Bus Inf Syst Eng 66(1):111–126 (2024)

[Link]

CATCHWORD

Generative AI
Stefan Feuerriegel • Jochen Hartmann • Christian Janiesch •

Patrick Zschech

Received: 29 April 2023 / Accepted: 7 August 2023 / Published online: 12 September

2023
© The Author(s) 2023

Keywords Generative AI · Artificial intelligence · The term generative AI refers to computational tech-
Decision support · Content creation · Information systems niques that are capable of generating seemingly new,
meaningful content such as text, images, or audio from
training data. The widespread diffusion of this technology
1 Introduction with examples such as Dall-E 2, GPT-4, and Copilot is
currently revolutionizing the way we work and communi-
Tom Freston is credited with saying ‘‘Innovation is taking cate with each other. Generative AI systems can not only
two things that exist and putting them together in a new be used for artistic purposes to create new text mimicking
way’’. For a long time in history, it has been the prevailing writers or new images mimicking illustrators, but they can
assumption that artistic, creative tasks such as writing and will assist humans as intelligent question-answering
poems, creating software, designing fashion, and compos- systems. Here, applications include information technology
ing songs could only be performed by humans. This (IT) help desks where generative AI supports transitional
assumption has changed drastically with recent advances in knowledge work tasks and mundane needs such as cooking
artificial intelligence (AI) that can generate new content in recipes and medical advice. Industry reports suggest that
ways that cannot be distinguished anymore from human generative AI could raise global gross domestic product
craftsmanship. (GDP) by 7% and replace 300 million jobs of knowledge
workers (Goldman Sachs 2023). Undoubtedly, this has
drastic implications not only for the Business & Informa-
Accepted after one revision by Susanne Strahringer. tion Systems Engineering (BISE) community, where we
will face revolutionary opportunities, but also challenges
S. Feuerriegel ( ) and risks that we need to tackle and manage to steer the
LMU Munich and Munich Center for Machine Learning,
technology and its use in a responsible and sustainable
Geschwister-Scholl-Platz 1, 80539 Munich, Germany
e-mail: feuerriegel@[Link] direction.
In this Catchword article, we provide a conceptualiza-
J. Hartmann tion of generative AI as an entity in socio-technical systems
Technical University of Munich, TUM School of Management, and provide examples of models, systems, and applica-
Arcisstr. 21, 80333 Munich, Germany
e-mail: [Link]@[Link] tions. Based on that, we introduce limitations of current
generative AI and provide an agenda for BISE research.
C. Janiesch Previous papers discuss generative AI around specific
TU Dortmund University, Otto-Hahn-Str. 12, 44319 Dortmund, methods such as language models (e.g., Teubner et al.
Germany
e-mail: [Link]@[Link] 2023; Dwivedi et al. 2023; Scho¨bel et al. 2023) or
specific applications such as marketing (e.g., Peres et
P. Zschech al. 2023), innovation management (Burger et al. 2023),
FAU Erlangen-Nu¨rnberg, Lange Gasse 20, 90403 scholarly research (e.g., Susarla et al. 2023; Davison et
Nu¨rnberg,
Germany
al. 2023), and education (e.g., Kasneci et al. 2023;
e-mail: [Link]@[Link] Gimpel et al.

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

112 S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126
(2024)
2023). Different from these works, we focus on generative Note that the modalities in Fig. 1 are neither complete
AI in the context of information systems, and, to this end, nor entirely distinctive and can be detailed further. In
we discuss several opportunities and challenges that are addition, many unique use cases such as, for example,
unique to the BISE community and make suggestions for modeling functional properties of proteins (Unsal et al.
impactful directions for BISE research. 2022) can be represented in another modality such as text.

2.2 A Model-, System-, and Application-Level View

2 Conceptualization of Generative AI

2.1 Mathematical Principles of Generative AI 2.2.1 Model-Level View

Generative AI is primarily based on generative modeling, A generative AI model is a type of machine learning
which has distinctive mathematical differences from dis- architecture that uses AI algorithms to create novel data
criminative modeling (Ng and Jordan 2001) often used in instances, drawing upon the patterns and relationships
data-driven decision support. In general, discriminative observed in the training data. A generative AI model is of
modeling tries to separate data points X into different critically central yet incomplete nature, as it requires fur-
classes Y by learning decision boundaries between them ther fine-tuning to specific tasks through systems and
(e.g., in classification tasks with Y 2 f0; 1g). In contrast applications.
to that, generative modeling aims to infer some actual Deep neural networks are particularly well suited for the
data distribution. Examples can be the joint probability purpose of data generation, especially as deep neural net-
distri- bution P(X, Y) of both the inputs and the outputs works can be designed using different architectures to
or P(Y), but where Y is typically from some high- model different data types (Janiesch et al. 2021; Kraus
dimensional space. By doing so, a generative model offers et al. 2020), for example, sequential data such as human
the ability to produce new synthetic samples (e.g., language or spatial data such as images. Table 1 presents
generate new obser- vation-target-pairs (X, Y) or new an overview of the underlying concepts and model archi-
observations X given a target value Y) (Bishop 2006). tectures that are common in the context of generative AI,
Building upon the above, a generative AI model refers such as diffusion probabilistic models for text-to-image
to generative modeling that is instantiated with a generation or the transformer architecture and (large) lan-
machine learning architecture (e.g., a deep neural guage models (LLMs) for text generation. GPT (short for
network) and, therefore, can create new data samples generative pre-trained transformer), for example, repre-
based on learned patterns.1 Further, a generative AI system sents a popular family of LLMs, used for text generation,
encompasses the entire infrastructure, including the model, for instance, in the conversational agent ChatGPT.
data processing, and user interface components. The Large generative AI models that can model output in
model serves as the core component of the system, which and across specific domains or specific data types in a
facilitates interaction and application within a broader comprehensive and versatile manner are oftentimes also
context. Lastly, generative AI applications refer to the called foundation models (Bommasani et al. 2021). Due to
practical use cases and imple- mentations of these their size, they exhibit two key properties: emergence,
systems, such as search engine opti- mization (SEO) meaning the behavior is oftentimes implicitly induced
content generation or code generation that solve real- rather than explicitly constructed (e.g., GPT models can
world problems and drive innovation across various create calendar entries in the .ical format even though such
domains. Figure 1 shows a systematization of models were not explicitly trained to do so), and homog-
generative AI across selected data modalities (e.g., text, enization, where a wide range of systems and applications
image, and audio) and the model-, system-, and application- can now be powered by a single, consolidated model (e.g.,
level perspectives, which we detail in the following Copilot can generate source code across a wide range of
1section. programming languages).
It should be noted, however, that advanced generative AI
models are often not based on a single modeling principle or Figure 1 presents an overview of generative AI models
learning mechanism, but combine different approaches. For example, along different, selected data modalities, which are pre-
language models from the GPT family first apply a generative trained on massive amounts of data. Note that we structure
pre-training stage to capture the distribution of language data
using a language modeling objective, while downstream systems the models in Fig. 1 by their output modality such as X-to-
typically then apply a discriminative fine-tuning stage to adapt text or X-to-image. For example, GPT-4 as the most recent
the model parameters to specific tasks (e.g., document generative AI model underlying OpenAI’s popular con-
classification, question answering). Similarly, ChatGPT combines versational agent ChatGPT (OpenAI 2023a) accepts both
techniques from generative modeling together with discriminatory
modeling and reinforcement learning (see Fig. 2). image and text inputs to generate text outputs. Similarly,

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126 (2024) 113

Fig. 1 A model-, system-, and

application-level view on
generative AI

Midjourney accepts both modalities to generate images. To 2.2.2 System-Level View

this end, generative AI models can also be grouped into
unimodal and multimodal models. Unimodal models take Any system consists of a number of elements that are
instructions from the same input type as their output (e.g., interconnected and interact with each other. For generative
text). On the other hand, multimodal models can take input AI systems, this comprises not only the aforementioned
from different sources and generate output in various generative AI model but also the underlying infrastructure,
forms. Multimodal models exist across a variety of data user-facing components, and their modality as well as the
modalities, for example for text, image, and audio. corresponding data processing (e.g., for prompts). An
Prominent examples include Stable Diffusion (Rombach example would be the integration of deep learning models,
et al. 2022) for text-to-image generation, MusicLM like Codex (Chen et al. 2021), into a more interactive and
(Agostinelli et al. 2023) for text-to-music generation, comprehensive system, like GitHub Copilot, which allows
Codex (Chen et al. 2021) and AlphaCode (Li et al. 2022) its users to code more efficiently. Similarly, Midjourney’s
for text-to-code generation, and as mentioned above GPT-4 image generation system builds on an undisclosed X-to-
for image-to-text as well as text-to-text generation image generation model that users can interact with to
(OpenAI 2023a). generate images using Discord bots. Thus, generative AI
The underlying training procedures vary greatly across systems embed the functionality of the underlying mathe-
different generative AI models (see Fig. 2). For example, matical model to provide an interface for user interaction.
generative adversarial networks (GANs) are This step augments the model-specific capabilities,
trained through two competing objectives enhancing its practicability and usability across real-world
(Goodfellow et al. use cases.
2014), where one is to create new synthetic samples Core concerns when embedding deep learning models in
while generative AI systems generally are scalability (e.g., dis-
the other tries to detect synthetic samples from the tributed computing resources), deployment (e.g., in various
actual training samples, so that the distribution of environments and for different devices), and usability (e.g.,
synthetic samples is eventually close to the a user-friendly interface and intent recognition). As pre-
distribution of the training samples. Differently, systems trained open-source alternatives to closed-source, propri-
such as ChatGPT- based conversational models use etary models continue to be released, making these models
reinforcement learning from human feedback (RLHF). available to their users (be it companies or individuals)
RLHF as used by ChatGPT proceeds in three steps to first becomes increasingly important. For both open-source and
create demonstration data for prompts, then to have users closed-source models, unexpected deterioration of model
rank the quality of different outputs for a prompt, and performance over time highlights the need for continuous
finally to learn a policy that generates desirable output model monitoring (Chen et al. 2023). Although powerful
via reinforcement learning so that the output would text-generating models existed before the release of the
score well during ranking (Ziegler et al. 2019).

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

114 S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126 (2024)

Table 1 Glossary of key concepts in generative AI

Concept Description

Diffusion probabilistic models Diffusion probability models are a class of latent variable models that are common for various tasks such
as image generation (Ho et al. 2020). Formally, diffusion probability models capture the image data by
modeling the way data points diffuse through a latent space, which is inspired by statistical physics.
Specifically, they typically use Markov chains trained with variational inference and then reverse the
diffusion process to generate a natural image. A notable variant is Stable Diffusion (Rombach et al.
2022). Diffusion probability models are also used in commercial systems such as DALL-E and
Generative adversarial network Midjourney.
A GAN is a class of neural network architecture with a custom, adversarial learning objective
(Goodfellow et al. 2014). A GAN consists of two neural networks that contest with each other in the form
of a zero-sum game, so that samples from a specific distribution can be generated. Formally, the first
network G is called the generator, which generates candidate samples. The second network D is called the
discriminator, which evaluates how likely the candidate samples come from a desired distribution.
Thanks to the adversarial learning objective, the generator learns to map from a latent space to a data
distribution of interest, while the discriminator distinguishes candidates produced by the generator from
(Large) language model the true data distribution (see Fig. 2).
A (large) language model (LLM) refers to neural networks for modeling and generating text data that
typically combine three characteristics. First, the language model uses a large-scale, sequential neural
network (e.g., transformer with an attention mechanism). Second, the neural network is pre-trained
through self-supervision in which auxiliary tasks are designed to learn a representation of natural
language without risk of overfitting (e.g., next-word prediction). Third, the pre-training makes use of
large-scale datasets of text (e.g., Wikipedia, or even multi-language datasets). Eventually, the language
model may be fine-tuned by practitioners with custom datasets for specific tasks (e.g., question
answering, natural language generation). Recently, language models have evolved into so-called LLMs,
which combine billions of parameters. Prominent examples of massive LLMs are BERT (Devlin et al.
2018) and GPT-3 (Brown et al. 2020) with ~ 340 million and ~ 175 billion parameters, respectively.
Reinforcement learning from
human feedback RLHF learns sequential tasks (e.g., chat dialogues) from human feedback. Different from traditional
reinforcement learning, RLHF directly trains a so-called reward model from human feedback and then
uses the model as a reward function to optimize the policy, which is optimized through data-efficient and
robust algorithms (Ziegler et al. 2019). RLHF is used in conversational systems such as ChatGPT
(OpenAI 2022) for generating chat messages, such that new answers accommodate the previous chat
dialogue and ensure that the answers are in alignment with predefined human preferences (e.g., length,
Prompt learning
style, appropriateness)
Prompt learning is a method for LLMs that uses the knowledge stored in language models for
downstream tasks (Liu et al. 2023). In general, prompt learning does not require any fine-tuning of the
language model, which makes it efficient and flexible. A prompt is a specific input to a language model
(e.g., ‘‘The movie was superb. Sentiment: ‘‘) and then the most probable output s 2 f‘‘positive’’;
‘‘negative’’g instead of the space is picked. Recent advances allow for more complex data-driven prompt
seq2seq
engineering, such as tuning prompts via reinforcement learning (Liu et al. 2023).
The term sequence-to-sequence (seq2seq) refers to machine learning approaches where an input sequence
is mapped onto an output sequence (Sutskever et al. 2014). An example is machine learning-based
translation between different languages. Such seq2seq approaches consist of two main components: An
encoder turns each element in a sequence (e.g., each word in a text) into a corresponding hidden vector
containing the element and its context. The decoder reverses the process, turning the vector into an output
element (e.g., a word from the new language) while considering the previous output to model
dependencies in language. The idea of seq2seq models has been extended to allow for multi-modal
Transformer mappings such as text-to-image or text-to-speech mappings.
A transformer is a deep learning architecture (Vaswani et al. 2017) that adopts the mechanism of self-
attention which differentially weights the importance of each part of the input data. Like recurrent
neural networks (RNNs), transformers are designed to process sequential input data, such as natural
language, with applications for tasks such as translation and text summarization. However, unlike
RNNs, transformers process the entire input all at once. The attention mechanism provides context for
any position in the input sequence. Eventually, the output of a transformer (or an RNN in general) is a
document embedding, which presents a lower-dimensional representation of text (or other input)
sequences where similar texts are located in closer proximity which typically benefits downstream tasks
Variational autoencoder as this allows to capture semantics and meaning (Siebers et al. 2022).
A variational autoencoder (VAE) is a type of neural network that is trained to learn a low-dimensional
representation of the input data by encoding it into a compressed latent variable space and then
reconstructing the original data from this compressed representation. VAEs differ from traditional
autoencoders by using a probabilistic approach to the encoding and decoding process, which enables
them to capture the underlying structure and variation in the data and generate new data samples from the
learned latent space (Kingma and Welling 2013). This makes them useful for tasks such as anomaly
detection and data compression but also image and text generation.

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126 (2024) 115

Table 1 continued
Concept Description

Zero-shot learning / few-shot Zero-shot learning and few-shot learning refer to different paradigms of how machine learning deals
learning with the problem of data scarcity. Zero-shot learning is when a machine is taught how to learn a task
from data without ever needing to access the data itself, while few-short learning refers to when there are
only a few specific examples. Zero-shot learning and few-shot learning are often desirable in practice as
they reduce the cost of setting up AI systems. LLMs are few-shot or zero-shot learners (Brown et al.
2020) as they just need a few samples to learn a task (e.g., predicting the sentiment of reviews), which
makes LLMs highly flexible as a general-purpose tool.

Fig. 2 Examples of different training procedures for generative AI models. a Generative adversarial network (GAN). b Reinforcement
learning from human feedback (RLHF) as used in conversational generative AI models

ChatGPT system in November 2022, ChatGPT’s ease of in many generative AI models are that they were trained on
use also for non-expert users was a core contributing factor historical data with specific cut-off date and thus do not
to its explosive worldwide adoption. store information beyond or that an information compres-
Moreover, on the system level, multiple components of sion takes place because of which generative AI models
a generative AI system can be integrated or connected to may not remember everything that they saw during training
other systems, external databases with domain-specific (Chiang 2023). Both limitations can be mitigated by aug-
knowledge, or platforms. For example, common limitations menting the model with functionality for real-time

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

116 S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126
(2024)
information retrieval, which can substantially enhance its (Berente et al. 2021). With generative AI, we are
accuracy and usefulness. Relatedly, in the context of text approaching a further point of refinement. In the past, the
generation, online language modeling addresses the prob- capability of AI was mostly understood to be analytic,
lem of outdated models by continuously training them on suitable for decision-making tasks. Now, AI gains the
up-to-date data.2 Thereby, such models can then be capability to perform generative tasks, suitable for content
knowledgeable of recent events that their static counter- creation. While the procedure of content creation to some
parts would not be aware of due to their training cut-off respect can still be considered analytic as it is inherently
dates. probabilistic, its results can be creative or even artistic as
generative AI combines elements in novel ways. Further,
2.2.3 Application-Level View IT artifacts were considered passive as they were used
directly by humans. With the advent of agentic IT artifacts
Generative AI applications are generative AI systems sit- (Baird and Maruping 2021) powered by LLMs (Park et al.
uated in organizations to deliver value by solving dedicated 2023), this human agency primacy assumption needs to be
business problems and addressing stakeholder needs. They revisited and impacts how we devise the relation between
can be regarded as human-task-technology systems or human and AI based on their potency. Eventually, this may
information systems that use generative AI technology to require AI capability models to structure, explain, guide,
augment human capacities to accomplish specific tasks. and constrain the different abilities of AI systems and their
This level of generative AI encompasses countless real- uses as AI applications.
world use cases: These range from SEO content generation Focusing on the interaction between humans and AI, so
(Reisenbichler et al. 2022), over synthetic movie genera- far, for analytic AI, the concept of delegation has been
tion (Metz 2023) and AI music generation (Garcia 2023), discussed to establish a hierarchy for decision-making
to natural language-based software development (Chen (Baird and Maruping 2021). With generative AI, a human
et al. 2021). uses prompts to engage with an AI system to create con-
Generative AI applications will give rise to novel tent, and the AI then interprets the human’s intentions and
technology-enabled modes of work. The more users will provides feedback to presuppose further prompts. At first
familiarize themselves with these novel applications, the glance, this seems to follow a delegation pattern as well.
more they will trust or mistrust them as well as use or Yet, the subsequent process does not, as the output of the
disuse them. Over time, applications will likely transition AI can be suggestive to the other and will inform their
from mundane tasks such as writing standard letters and further involvement directly or subconsciously. Thus, the
getting a dinner reservation to more sensitive tasks such as process of creation rather follows a co-creation pattern, that
soliciting medical or legal advice. They will involve more is, the practice of collaborating in different roles to align
consequential decisions, which may even involve moral and offer diverse insights to guide a design process
judgment (Kru¨gel et al. 2023). This ever-increasing (Ramaswamy and Ozcan 2018). Using the lens of agentic
scope and pervasiveness of generative AI applications give AI artifacts, initiation is not limited to humans.
rise to an imminent need not only to provide The abovementioned interactions also impact our cur-
prescriptions and principles for trustworthy and reliable rent understanding of hybrid intelligence as the integration
designs, but also for scrutinizing the effects on the user of humans and AI, leveraging the unique strengths of
to calibrate qualities such as trust appropriately. The both. Hybrid intelligence argues to address the
(continued) use and adoption of such applications by limitations of
end users and organiza- tions entails a number of each intelligence type by combining human intuition,
fundamental socio-technical considerations to descry cre-
innovation potential and affor- dances of generative AI ativity, and empathy with the computational power,
artifacts. accu- racy, and scalability of AI systems to achieve
enhanced decision-making and problem-solving
2.3 A Socio-Technical View on Generative AI
capabilities (Deller- mann et al. 2019). With generative AI
and the AI’s capa- bility to co-create, the understanding
As technology advances, the definition and extent of what
of what constitutes this collective intelligence begins to
constitutes AI are continuously refined, while the reference
shift. Hence, novel human-AI interaction models and
point of human intelligence stays comparatively constant
patterns may become necessary to explain and guide the
behavior of humans and AI systems to enable effective
2
See [Link] (accessed 25 and efficient use in AI applications on the one hand
Aug 2023) for a script that enables users to pull up-to-date data and, on the other hand, to ensure envelopment of AI
from the web for training online language models, for instance, from agency and reach (Asatiani et al. 2021).
Common Crawl and Wikipedia.
On a theoretical level, this shift in human-computer or
rather human-AI interaction fuels another important
13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126 (2024) 117

observation: The theory of mind is an established theoret- Moreover, the output of generative AI, especially that of
ical lens in psychology to describe the cognitive ability of LLMs, is typically not easily verifiable.
individuals to understand and predict the mental states, The correctness of generative AI models is highly
emotions, and intentions of others (Carlson et al. 2013; dependent on the quality of training data and the according
Baron-Cohen 1997; Gray et al. 2007). This skill is crucial learning process. Generative AI systems and
for social interactions, as it facilitates empathy and allows applications can implement correctness checks to inhibit
for effective communication. Moreover, conferring a mind certain out-
to an AI system can substantially drive usage intensity puts. Yet, due to the black-box nature of state-of-the-art
(Hartmann et al. 2023a). The development of a theory of AI
mind in humans is unconscious and evolves throughout an models (Rai 2020), the usage of such systems critically
individual’s life. The more natural AI systems become in hinges on users’ trust in reliable outputs. The closed source
terms of their interface and output, the more a theory of of commercial off-the-shelf generative AI systems aggra-
mind for human-computer interactions becomes necessary. vates this fact and prohibits further tuning and re-training
Research is already investigating how AI systems can of the models. One solution for addressing the downstream
become theory-of-mind-aware to better understand their implications of incorrect outputs is to use generative AI to
human counterpart (Rabinowitz et al. 2018; C¸ elikok et produce explanations or references, which can then be
al. 2019). However, current AI systems hardly offer any verified by users. However, such explanations are again
cues for interactions. Thus, humans are rather void of a probabilistic and thus subject to errors; nevertheless, they
theory to explain their understanding of intelligent may help users in their judgment and decision-making
behavior by AI systems, which becomes even more when to accept outputs of generative AI and when not.
important in a co- creation environment that does not Bias and fairness. Societal biases permeate everyday
follow a task delegation pattern. A theory of the artificial human-generated content (Eskreis-Winkler and Fishbach
mind that explains how individuals perceive and assume 2022). The unbiasedness of vanilla generative AI is very
the states and rationale of AI systems to better collaborate much dependent on the quality of training data and the
with them may alleviate some of these concerns. alignment process. Training deep learning models on
biased data can amplify human biases, replicate toxic
language, or perpetuate stereotypes of gender, sexual ori-
3 Limitations of Current Generative AI entation, political leaning, or religion (e.g., Caliskan et al.
2017; Hartmann et al. 2023b). Recent studies expose the
In the following, we discuss four salient boundaries of harmful biases embedded in multimodal generative AI
generative AI that, we argue, are important limitations in models such as CLIP (contrastive language-image pre-
real-world applications. The following limitations are of training; Wolfe et al. 2022) and the CLIP-filtered LAION
technical nature in that they refer to how current generative dataset (Birhane et al. 2021), which are core components
AI models make inferences, and, hence, the limitations of generative AI models (e.g., Dall-E 2 or Stable Diffu-
arise at the model level. Because of this, it is likely that sion). Human biases can also creep into the models in other
limitations will persist in the long run, with system- and stages of the model engineering process. For instruction-
application-level implications. based language models, the RLHF process is an additional
Incorrect outputs. Generative AI models may produce source of bias (OpenAI 2023b). Careful coding guidelines
output with errors. This is owed to the underlying nature of and quality checks can help address these risks.
machine learning models relying on probabilistic Addressing bias and thus fairness in AI receives
algo- rithms for making inferences. For example, increasing attention in the academic literature (Dolata et al.
generative AI 2022; Schramowski et al. 2022; Ferrara 2023; De-Arteaga
models generate the most probable response to a et al. 2022; Feuerriegel et al. 2020; von Zahn et al. 2022),
prompt, but remains an open and ongoing research question. For
not necessarily the correct response. As such, example, the developers of Stable Diffusion flag ‘‘probing
challenges arise as, by now, outputs are and understanding the limitations and biases of generative
indistinguishable from authentic content and may models’’ as an important research area (Rombach et al.
present misinformation or deceive users (Spitale et al. 2022). Some scholars even attest to models certain moral
2023). In LLMs, this problem in emergent behavior is self-correcting capabilities (Ganguli et al. 2023), which
called hallucination (Ji et al. 2023), which refers to may attenuate concerns of embedded biases and result in
mistakes in the generated text that are semantically or more fairness. In addition, on the system and application
syntactically plausible but are actually nonsensical or level, mitigation mechanisms can be implemented to
incorrect. In other words, the generative AI model address biases embedded in the deep learning models and
produces content that is not based on any facts or create more diverse outputs (e.g., updating the prompts
evidence, but rather on its own assumptions or biases. ‘‘under the hood’’ as done by Dall-E 2 to increase the
13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

118 S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126
(2024)
demographic diversity of the outputs). Yet, more research other hand, offer numerous research opportunities, espe-
is needed to get closer to the notion of fair AI. cially for BISE researchers due to their interdisciplinary
Copyright violation. Generative AI models, systems, background. We organize our considerations according to
and applications may cause a violation of copyright laws the individual departments of the BISE journal (see
because they can produce outputs that resemble or Table 2 for an overview of exemplary research questions).
even copy existing works without permission or
compensation to 4.1 Business Process Management
the original creators (Smits and Borghuis 2022). Here,
two Generative AI will have a strong impact on the field of
potential infringement risks are common. On the one Business Process Management (BPM) as it can assist in
hand, generative AI may make illegal copies of a automating routine tasks, improving customer and
work, thus violating the reproduction right of creators. employee satisfaction, and revealing process innovation
Among others, this may happen when a generative AI opportunities (Beverungen et al. 2021), especially in cre-
was trained on original content that is protected by ative processes (Haase and Hanel 2023). Concrete impli-
copyright but where the generative AI produces cations and research directions can be connected to various
copies. Hence, a typical implication is that the training phases of the BPM lifecycle model (Vidgof et al. 2023).
data for building generative AI systems must be free of For example, in the context of process discovery, genera-
copyrights. Crucially, copyright violation may tive AI models could be used to generate process
nevertheless still happen even when the generative AI descriptions, which can help businesses identify and
has never seen a copyrighted work before, such as, for understand the different stages of a process (Kecht et al.
example, when it simply produces a trade- marked 2023). From the perspective of business process improve-
logo similar to that of Adidas but without ever ment, generative process models could be used for idea
having seen that logo before. On the other hand, generation and to support innovative process (re-)design
generative AI may prepare derivative works, thus violating initiatives (van Dun et al. 2023). In this regard, there is
the transformation right of creators. To this end, legal great potential for generative AI to contribute to both
questions arise around the balance of originality and cre- exploitative as well as explorative BPM design strategies
ativity in generative AI systems. Along these lines, legal (Grisold et al. 2022). In addition, natural language pro-
questions also arise around who holds the intellectual cessing tasks related to BPM such as process extraction
property for works (including patents) produced by a from text could benefit from generative AI without further
generative AI. fine-tuning using prompt engineering (Busch et al. 2023).
Environmental concerns. Lastly, there are substantial Likewise, other phases can benefit owing to generative
environmental concerns from developing and using gen- AI’s ability to learn complex and non-linear relationships
erative AI systems due to the fact that such systems are in dynamic business processes that can be used for
typically built around large-scale neural networks, and, implementation as well as in simulation and predictive
therefore, their development and operation consume large process monitoring among other things.
amounts of electricity with immense negative carbon In the short term, robotic process automation (van der
footprint (Schwartz et al. 2020). For example, the carbon Aalst et al. 2018; Herm et al. 2021) will benefit as formerly
emission for training a generative AI model such as GPT-3 handcrafted processing rules can not only be replaced, but
was estimated to have produced the equivalent of entirely new types of automation can be enabled by ret-
552 t CO2 and thus amounts to the annual CO2 emissions rofitting and thus intelligentizing legacy software. In the
of several dozens of households (Khan 2021). Owing to long run, we also see a large potential to support the phase
this, there are ongoing efforts in AI research to make the of business process execution in traditional BPM. Specifi-
development and deployment of AI algorithms more car- cally, we anticipate the development of a new generation of
bon-friendly, through more efficient training algorithms, process guidance systems. While traditional system designs
through compressing the size of neural network architec- are based on static and manually-crafted knowledge bases
tures, and through optimized hardware (Schwartz et al. (Morana et al. 2019), more dynamic and adaptive systems
42020).
Implications and Future Directions for the BISE are feasible on the basis of large enterprise-wide trained
Community language models. Such systems could improve knowledge
retrieval tasks from a wide variety of heterogeneous sour-
ces, including manuals, handbooks, e-mails, wikis, job
In this section, we draw a number of implications and
descriptions, etc. This opens up new avenues of research
future research directions which, on the one hand, are of
into how unstructured and distributed organizational
direct relevance to the BISE community as an application-
oriented, socio-technical research discipline and, on the

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126 (2024) 119

Table 2 Examples of research questions for future BISE research on generative

AI
BISE department Research questions (examples)
Business process management How can generative AI assist in automating routine tasks?
How can generative AI reveal process innovation opportunities and support process
(re-)design initiatives?
Decision analytics and data How can generative AI models be effectively fine-tuned for domain-specific
science applications? How can the reliability of generative AI systems be improved?
Digital business management and How can generative AI support managerial tasks such as resource allocation?
digital leadership How will the digital work of employees change with smart assistants powered by generative
Economics of information systems AI? What are the welfare implications of generative AI?
Which jobs and tasks are affected most by generative AI?
Enterprise modeling and enterprise How can generative AI be used to support the construction and maintenance of enterprise
engineering models? How can generative AI support in enterprise applications (e.g., CRM, BI, etc.)?
Human computer interaction and social How should generative AI systems be designed to foster trust?
computing What countermeasures are effective to prevent users from falling for AI-generated
disinformation? To what extent can generative AI replace or augment crowdsourcing tasks?
How can generative AI assist in education?
Information systems engineering and What are effective design principles for developing generative AI systems?
technology How can generative AI support design science projects to foster creativity in the development
of new IT artifacts?

knowledge can be incorporated into intelligent process better ways how outputs can be verified (e.g., by offering
guidance systems. additional explanations or references).
Finally, questions arise about how generative AI can
4.2 Decision Analytics and Data Science natively support decision analytics and data science pro-
jects by closing the gap between modeling experts and
Despite the huge progress in recent years, several analytical domain users (Zschech et al. 2020). For instance, it is
and technical questions around the development of gener- commonly known that many AI models used in business
ative AI have yet to be solved. One open question relates to analytics are difficult to understand by non-experts (cf.
how generative AI can be effectively customized for Senoner et al. 2022). As a remedy, generative AI could be
domain-specific applications and thus improve perfor- used to generate descriptions that explain the logic of
mance through higher degrees of contextualization. For business analytics models and thus make the decision logic
example, novel and scalable techniques are needed to more intelligible. One promising direction could be, for
customize conversational agents based on generative AI for example, to use generative AI for translating post hoc
applications in medicine or finance. This will be crucial in explanations derived from approaches like SHAP or LIME
practice to solve specific BISE-related tasks where cus- into more intuitive textual descriptions or generate user-
tomization may bring additional performance gains. Novel friendly descriptions of models that are intrinsically inter-
techniques for customization must be designed in a way pretable (Slack et al. 2023; Zilker et al. 2023).
that ensures the safety of proprietary data and prevents the
data from being disclosed. Moreover, new frameworks are 4.3 Digital Business Management and Digital
needed for prompt engineering that are designed from a Leadership
user-centered lens and thus promote interpretability and
usability. Generative AI has great potential to contribute to different
Another important research direction is to improve the types of value creation mechanisms, including knowledge
reliability of generative AI systems. For example, algo- creation, task augmentation, and autonomous agency.
rithmic solutions are needed on how generative AI can However, this also requires the necessary organizational
detect and mitigate hallucination. In addition to algorithmic capabilities and conditions, where further research is nee-
solutions, more effort is also needed to develop user-cen- ded to examine these ingredients more closely for the
tered solutions, that is, how users can reduce the risk of context of generative AI to steer the technological possi-
falling for incorrect outcomes, for example, by developing bilities in a successful direction (Shollo et al. 2022).

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

120 S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126
(2024)
That is, generative AI will lead to the development of example, AI-based translation between different languages
new business ideas, unseen product and service innova- is responsible for significant economic gains (Brynjolfsson
tions, and ultimately to the emergence of completely new et al. 2019). The BISE community can contribute by pro-
business models. At the same time, it will also have a viding quantification through rigorous causal evidence.
strong impact on intra-organizational aspects, such as work Given the velocity of AI research, it may be necessary to
patterns, organizational structures, leadership models, and take a more abstract problem view instead of a concrete
management practices. In this regard, we see that AI-based tool view. For example, BISE research could run field
assistant systems previously centered around desktop experiments to compare programmers with and without AI
automation taking over more and more routine tasks such support and thereby assess whether generative AI systems
as event management, resource allocation, and social for coding can improve the speed and quality of code
media account management to free up even more human development. Similarly, researchers could test whether
capacity (Maedche et al. 2019). Further, in algorithmic generative AI will make artists more creative as they can
management (Benlian et al. 2022; Cameron et al. 2023), it more easily create new content. A similar pattern was
should be examined how existing theories and frameworks previously observed for AlphaGo, which has led humans to
need to be contextualized or fundamentally extended in become better players in the board game Go (Shin et al.
light of the increasingly powerful capabilities of generative 2023).
AI. Generative AI is likely to transform the industry as a
However, there are not only implications at the man- whole. This may hold true in the case of platforms that
agement level. The future of work is very likely to change make user-generated content available (e.g.,
at all levels of an organization (Feuerriegel et al. 2022). shutterstock.- com, [Link], [Link]),
Due to the multi-modality of generative AI models, it is which may be
conceivable that employees will work increasingly via replaced by generative AI systems. Here, further
smart, speech-based interfaces, whereby the formulation of research
prompts and the evaluation of their results could become a questions arise as to whether the use of generative AI
key activity. Against this background, it is worth investi- can lead to a competitive advantage and how generative
gating which new competencies are required to handle this AI changes competition. For example, what are the
emerging technology (cf. Debortoli et al. 2014) and which economic implications if generative AI is developed as
entirely new job profiles, such as prompt engineers, may open-source vs. closed-source systems? In this regard, a
evolve in the near future (Strobelt et al. 2023). salient success factor for the development of
Generative AI is also expected to fundamentally reform conversational agents based on generative AI (e.g.,
the way organizations manage, maintain, and share ChatGPT) are data from user interactions through
knowledge. Referring to the sketched vision of a new dialogues and feedback on whether the dialog was helpful.
process guidance system in Sect. 4.1, we anticipate a Hence, the value of such interaction data is poorly
number of new opportunities for digital knowledge man- understood and what it means if such data are only
agement, among others automated knowledge discovery available to a few Big Tech companies.
based on large amounts of unstructured distributed data The digital transformation from generative AI also poses
(e.g., identification of new product combinations), challenges and opportunities for economic policy. It may
improved knowledge sharing by automating the process of affect future work patterns and, indirectly, worker capa-
creating, summarizing, and disseminating content (e.g., bility via restructured learning mechanisms. It may also
automated creation of wikis and FAQs in different lan- affect content sharing and distribution and, hence, have
guages), and personalized knowledge delivery to individual non-trivial implications on the exploitation and protection
employees based on their specific needs and preferences of intellectual properties. On top of that, a growing con-
(e.g., recommendations for specific training material). centration of power over AI innovation in the hands of a
few companies may result in a monopoly of AI capabilities
and hamper future innovation, fair competition, scientific
4.4 Economics of Information Systems
progress, and thus welfare and human development at
large. All of these future impacts are important to under-
Generative AI will have significant economic implications
stand and provide meaningful directions for shaping eco-
across various industries and markets. Generative AI can 4.5 Enterprise Modeling and Enterprise Engineering
nomic policy.
increase efficiency and productivity by automating many
tasks that were previously performed by humans, such as Enterprise models are important artifacts for capturing
content creation, customer service, code generation, etc. insights into the core components and structures of an
This can reduce costs and open up new opportunities for organization, including business processes, resources,
growth and innovation (Eloundou et al. 2023). For information flows, and IT systems (Vernadat 2020). A

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126 (2024) 121

major drawback of traditional enterprise models is that they if necessary, need to be extended. Nevertheless, this situ-
are static and may not provide the level of abstraction that ation also offers the potential to explore and design new
is required by the end user. Likewise, their construction approaches for more effective API management (e.g.,
and maintenance are time-consuming and expensive and including novel app store solutions, privacy and security
require manual effort and human expertise (Silva et al. mechanisms, service level definitions, pricing, and licens-
2021). With generative AI, we see a large potential that ing models) so that generative AI solutions can be
many of these limitations can be addressed by generative smoothly integrated into existing enterprise IT infrastruc-
AI as assistive technology (Sandkuhl et al. 2018), for tures without risking any unauthorized use and confiden-
example by automatically creating and updating enterprise tiality breaches.
models at different levels of abstraction or generating
multi-modal representations. 4.6 Human Computer Interaction and Social
First empirical results suggest that generative AI is able Computing
to generate useful conceptual models based on textual
problem descriptions. Fill et al. (2023) show that ER, Salient behavioral questions related to the interactions
BPMN, UML, and Heraklit models can not only be gen- between humans and generative AI systems are still
erated with very high to perfect accuracy from textual unanswered. Examples are related to the perception,
descriptions, but they also explored the interpretation of acceptance, adoption, and trust of systems using generative
existing models and received good results. In the near AI. A study found that news was believed less if generated
future, we expect more research that deals with the by generative AI instead of humans (Longoni et al. 2022)
development, evaluation, and application of more and another found that there is a replicant effect (Jakesch
advanced approaches. Specifically, we expect that learned et al. 2019). Such behavior is likely to be context-specific
representations of enterprise models can be transformed and will vary by other antecedents highlighting the need
into more application-specific formats and can either be for a principled theoretical foundation to build
enriched with further details or reduced to the essential successful generative AI systems. The BISE
content. community is well positioned to develop rigorous design
Against this background, the concept of ‘‘digital twins’’, recommendations.
virtual representations of enterprise assets, may experience Further, generative AI is a key enabler for developing
new accentuation and extensions (Dietz and Pernul high-quality interfaces for information systems based on
2020). Especially, in the public sector, where most natural language that promote usability and
organizational accessibility. For example, such interfaces will not only
assets are non-tangible in the form of defined make interac-
services, tions more intuitive but will also facilitate people
specified procedures, legal texts, manuals, and with
organiza- tional charts, generative AI can play a crucial disabilities. Generative AI is likely to increase the
role in dig- itally mirroring and managing such assets ‘‘degree of intelligence’’ of user assistance systems.
along their lifecycles. Similar benefits could be explored However, the design of effective interactions must also
with physical assets in Industry 4.0 environments (Lasi et be considered when increasing the degree of intelligence
al. 2014). (Maedche et al. 2016). Similarly, generative AI will
In enterprise engineering, the role of generative AI undoubtedly have an impact on (computer-mediated)
systems in existing as well as newly emerging IT land- communication and col- laboration, such as within
scapes to support the business goals and strategies of an companies. For example, gen- erative AI can create
organization gives rise to numerous opportunities (e.g., in optimized content for social media, emails, and reports.
office solutions, customer relationship management and It can also help to improve the onboarding of new
business analytics applications, knowledge management employees by creating personalized and interactive training
systems, etc.). Generative AI systems have the potential to materials. It can also enhance collabo- ration within teams
evolve into core enterprise applications that can either be by providing creative and intelligence conservation agents
hosted on-premise or rented in the cloud. Unsanctioned use that suggest, summarize, and synthe- size information
bears the risk that third-party applications will be used for based on the context of the team (e.g., automated
job-related tasks without explicit approval or even knowl- meeting notes).
edge of the organization. This phenomenon is commonly Several applications and research opportunities are
known as shadow IT and theories and frameworks have related to the use of generative AI in marketing and,
been proposed to explain this phenomenon, as well as especially, e-commerce. It is expected that generative AI
recommending actions and policies to mitigate associated can automate the creation of personalized marketing con-
risks (cf. Haag and Eckhardt 2017; Klotz et al. 2022). In tent, for instance, different sales slogans for introverts vs.
the light of generative AI, however, such approaches have extroverts (Matz et al. 2017) or other personality traits as
to be revisited for their applicability and effectiveness and, personalized marketing content is more effective than a
one-content-fits-all approach (Matz et al. 2023).

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

13
122 S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126
(2024)
Generative AI may automate various tasks in marketing vein, the widespread availability of generative AI systems
and media where content generation is needed (e.g., may also propel research around virtual assistants. Previ-
writing news stories, summarizing web pages for mobile ously, research made use of ‘‘Wizard-of-Oz’’ experiments
devices, creating thumbnail images for news stories, (Diederich et al. 2020), while future research may build
translating written news to audio for blind people and upon generative AI systems instead.
Braille-sup- ported formats for deaf people) that may Crucially, automated content generation using genera-
be studied in future research. Moreover, generative AI tive AI is a new phenomenon, but automation in general
may be used in recommender systems to boost the and how people are affected by automated systems
effectiveness of infor- mation dissemination through has been studied by scholars for decades. Thus, existing
personalization as content can be tailored better to the theo-
abilities of the recipient. ries on the interplay of humans with automated
The education sector is another example that will need systems may be contextualized to generative AI systems.
to reinvent in some parts following the availability of 4.7 Information Systems Engineering and Technology
conversational agents (Kasneci et al. 2023; Gimpel et
al. 2023). At first glance, generative AI seems to constitute Generative AI offers many engineering- and technology-
an oriented research opportunities for the Information Systems
unauthorized aid that jeopardizes student grading so community as a design-oriented discipline. This includes
far developing and evaluating design principles for generative
relying on written examinations and term papers. AI systems and applications to extend the limiting
However, over time, examinations will adapt, and boundaries of this technology (cf. Section 3). As such,
generative AI will enable the development of design principles can focus on how generative AI systems
comprehensive digital teaching assistants as well as the can be made explainable to enable interpretability, under-
creation of supplemental teaching material such as standing, and trust; how they can be designed reliable to
teaching cases and recap questions. Fur- ther, the avoid discrimination effects or privacy issues; and how
educator’s community will need to develop novel they can be built more energy efficient to promote envi-
guidelines and governance frameworks that educate ronmental sustainability (cf. Schoormann et al. 2023b).
learners to rely appropriately on generative AI systems, While a lot of research is already being conducted in
how to verify model outputs, and to engineer prompts technology-oriented disciplines such as computer science,
rather than the output itself. the BISE community can add its strength by looking at
In addition, generative AI, specifically LLMs, can not design aspects through a socio-technical lens, involving
only be used to spot harmful content on social media (e.g., individuals, teams, organizations, and societal groups in
Maarouf et al. 2023), but it can also create realistic design activities, and thereby driving the field forward with
disin- formation (e.g., fake news, propaganda) that is new insights from a human–machine perspective (Maedche
hard to et al. 2019).
detect by humans (Kreps et al. 2022; Jakesch et al. 2023). Further, we see great potential that generative AI can be
Notwithstanding, AI-generated disinformation has previ- leveraged to improve current practices in design science
ously evolved as so-called deepfakes (Mirsky and Lee research projects when constructing novel IT artifacts
2021), but recent advances in generative AI reduce the cost (see Hevner et al. 2019). Here, one of the biggest
of creating such disinformation and allow for unprece- potentials
dented personalization. For example, generative AI can could lie in the support of knowledge retrieval
automatically adapt the tone and narrative of misinforma- tasks.
tion to specific audiences that identify as extroverts or Currently, design knowledge in the form of
introverts, left- or right-wing partisans, or people with design requirements, design principles, and design
particular religious beliefs. features is often only available in encapsulated written
Lastly, generative AI can facilitate—or even replace— papers or implicitly embedded in instantiated artifacts.
traditional crowdsourcing where annotations or other Generative AI has the potential to extract such design
knowledge tasks are handled by a larger pool of crowd knowledge that is spread over a broad body of
workers, for example in social media content annotation interdisciplinary research and make it available in a
(Gilardi et al. 2023) or market research on willingness-to- collective form for scholars and practitioners. This could
pay for services and products (Brand et al. 2023). In gen- also overcome the limitation that design knowledge is
eral, we expect that generative AI will automate many currently rarely reused, which ham- pers the fundamental
other tasks being a zero-shot / few-shot learner. However, idea of knowledge accumulation in design science
this may also unfold negative implications: Users may research (Schoormann et al. 2021).
contribute less to question-answering forums such as Besides engineering actual systems and applications, the
[Link], which thus may reduce human-based BISE community should also investigate how generative
13
knowledge creation impairing the future performance of AI can be used to support creativity-based tasks when
AI-based question-answering systems that rely upon
human question-answering content for training. In a similar
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126 (2024) 123

initiating new design projects. In this regard, a promising indicated otherwise in a credit line to the material. If material is
direction could be to incorporate generative AI in design not included in the article’s Creative Commons licence and your
thinking and similar methodologies to combine human intended use is not permitted by statutory regulation or exceeds the
permitted use, you will need to obtain permission directly from the
creativity with computational creativity (Hawlitschek
copyright holder. To view a copy of this licence, visit
2023). This may support different phases and steps of [Link] org/licenses/by/4.0/.
innovation projects, such as idea generation, user needs
elicitation, prototyping, design evaluation, and design
automation, in which different types of generative AI References
models and systems could be used and combined with each
other to form applications for creative industries (e.g., Agostinelli A, Denk TI, Borsos Z, Engel J, Verzetti M, Caillon A,
generated user stories with textual descriptions, visual Huang Q, Jansen A, Roberts A, Tagliasacchi M, et al (2023)
MusicLM: generating music from text. arXiv:2301.11325
mock-ups for user interfaces, and quick software proto- Asatiani A, Malo P, Nagbøl PR, Penttinen E, Rinta-Kahila T,
types for proofs-of-concept). If generative AI is used to co- Salovaara A (2021) Sociotechnical envelopment of
create innovative outcomes, it may also enable better artificial intelligence: an approach to organizational
reflection of the different design activities to ensure the deployment of inscrutable artificial intelligence systems. J
Assoc Inf Syst 22(2):8
necessary learning (Schoormann et al. 2023a). Baird A, Maruping LM (2021) The next generation of research on
IS use: a theoretical framework of delegation to and from
agentic IS artifacts. MIS Q 45(1):315–341
Baron-Cohen S (1997) Mindblindness: an essay on autism and
5 Conclusion theory of mind. MIT Press, Cambridge
Benlian A, Wiener M, Cram WA, Krasnova H, Maedche A,
Generative AI is a branch of AI that can create new content Mo¨hlmann M, Recker J, Remus U (2022) Algorithmic manage-
such as texts, images, or audio that increasingly often ment. Bus Inf Syst Eng 64(6):825–839. [Link]
s12599-022-00764-w
cannot be distinguished anymore from human craftsman- Berente N, Gu B, Recker J, Santhanam R (2021) Special issue
ship. For this reason, generative AI has the potential to editor’s comments: managing artificial intelligence.
transform domains and industries that rely on creativity, MIS Q 45(3):1433–1450
innovation, and knowledge processing. In particular, it Beverungen D, Buijs JCAM, Becker J, Di Ciccio C, van der Aalst
WMP, Bartelheimer C, vom Brocke J, Comuzzi M, Kraume K,
enables new applications that were previously impossible Leopold H, Matzner M, Mendling J, Ogonek N, Post T, Resinas
or impractical for automation, such as realistic virtual
M, Revoredo K, del R´ıo-Ortega A, La Rosa M, Santoro
assistants, personalized education and service, and digital FM, Solti A, Song M, Stein A, Stierle M, Wolf V (2021)
art. As such, generative AI has substantial implications for Seven paradoxes of business process management in a hyper-
BISE practitioners and scholars as an interdisciplinary connected world. Bus Inf Syst Eng 63(2):145–156.
research community. In our Catchword article, we offered [Link] s12599-020-00646-z
Birhane A, Prabhu VU, Kahembwe E (2021) Multimodal datasets:
a conceptualization of the principles of generative AI along
misogyny, pornography, and malignant stereotypes. arXiv:2110.
a model-, system-, and application-level view as well as a 01963
social-technical view and described limitations of current Bishop C (2006) Pattern recognition and machine learning. Springer,
generative AI. Ultimately, we provided an impactful New York
Bommasani R, Hudson DA, Adeli E, Altman R, Arora S, von Arx S,
research agenda for the BISE community and thereby
Bernstein MS, Bohg J, Bosselut A, Brunskill E, Brynjolfsson
highlight the manifold affordances that generative AI E, Buch S, Card D, Castellon R, Chatterji NS, Chen AS, Creel
offers through the lens of the BISE discipline. KA, Davis J, Demszky D, Donahue C, Doumbouya M,
Durmus E, Ermon S, Etchemendy J, Ethayarajh K, Fei-Fei L,
Acknowledgements During the preparation of this Catchword, Finn C, Gale T, Gillespie LE, Goel K, Goodman ND,
we contacted all current department editors at BISE to actively seek Grossman S, Guha N, Hashimoto T, Henderson P, Hewitt J,
their feedback on our suggested directions. We gratefully Ho DE, Hong J, Hsu K, Huang J, Icard TF, Jain S, Jurafsky D,
acknowledge their support. Kalluri P, Karamcheti S, Keeling G, Khani F, Khattab O, Koh
PW, Krass MS, Krishna R, Kuditipudi R, Kumar A, Ladhak F,
Lee M, Lee T, Leskovec J, Levent I, Li XL, Li X, Ma T,
Funding Open Access funding enabled and organized by Malik A, Manning CD, Mirchandani SP, Mitchell E,
Projekt DEAL. Munyikwa Z, Nair S, Narayan A, Narayanan D, Newman B,
Nie A, Niebles JC, Nilforoshan H, Nyarko JF, Ogut G, Orr L,
Open Access This article is licensed under a Creative Papadimitriou I, Park JS, Piech C, Portelance E, Potts C,
Commons Attribution 4.0 International License, which permits Raghunathan A, Reich R, Ren H, Rong F, Roohani YH, Ruiz
use, sharing, adaptation, distribution and reproduction in any C, Ryan J, R’e C, Sadigh D, Sagawa S, Santhanam K,
medium or format, as long as you give appropriate credit to the Shih A, Srinivasan KP, Tamkin A, Taori R, Thomas AW,
original author(s) and the source, provide a link to the Creative
Trame`r F, Wang RE, Wang W, Wu B, Wu J, Wu Y, Xie SM,
Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Yasunaga M, You J, Zaharia MA, Zhang M, Zhang T, Zhang X,
Creative Commons licence, unless Zhang Y, Zheng L, Zhou K, Liang P (2021) On the

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

124 S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126
(2024)
opportunities and risks of foundation models. Dwivedi YK, Kshetri N, Hughes L, Slade EL, Jeyaraj A, Kar AK,
arXiv:2108. 07258[Link] Baabdullah AM, Koohang A, Raghavan V, Ahuja M et al (2023)
Brand J, Israeli A, Ngwe D (2023) Using GPT for market research. ‘‘So what if ChatGPT wrote it?’’ Multidisciplinary perspectives
SSRN 4395751 on opportunities, challenges and implications of
Brown T, Mann B, Ryder N, Subbiah M, Kaplan JD, Dhariwal P, generative conversational AI for research, practice and
Neelakantan A, Shyam P, Sastry G, Askell A et al (2020) policy. Int J Inf Manag 71(102):642
Language models are few-shot learners. Adv Neural Inf Eloundou T, Manning S, Mishkin P, Rock D (2023) GPTs are GPTs:
Process Syst 33:1877–1901 an early look at the labor market impact potential of
Brynjolfsson E, Hui X, Liu M (2019) Does machine translation affect large language models. arxiv:2303.10130, accessed 03 April
international trade? Evidence from a large digital platform. 2023
Manag Sci 65(12):5449–5460 Eskreis-Winkler L, Fishbach A (2022) Surprised elaboration: when
Burger B, Kanbach DK, Kraus S, Breier M, Corvello V (2023) On white men get longer sentences. J Personal Soc
the use of AI-based tools like ChatGPT to support Psychol 123:941–956
management Ferrara E (2023) Should ChatGPT be biased? Challenges and risks
research. Europ J Innov Manag 26(7):233–241. [Link] of
10.1108/EJIM-02-2023-0156 bias in large language models. arXiv:2304.03738
Busch K, Rochlitzer1 A, Sola D, Leopold H (2023) Just tell Feuerriegel S, Dolata M, Schwabe G (2020) Fair AI: challenges
me: Prompt engineering in business process management. and opportunities. Bus Inf Syst Eng 62:379–384
arXiv: Feuerriegel S, Shrestha YR, von Krogh G, Zhang C (2022) Bringing
2304.07183 artificial intelligence to business management. Nat Machine
Caliskan A, Bryson JJ, Narayanan A (2017) Semantics Intell 4(7):611–613
derived automatically from language corpora contain human- Fill HG, Fettke P, Ko¨pke J (2023) Conceptual modeling and large
like biases. Sci 356(6334):183–186 language models: impressions from first experiments with
Cameron L, Lamers L, Leicht-Deobald U, Lutz C, Meijerink J, ChatGPT. EMISAJ 18(3):1–15. [Link]
Mo¨hlmann M (2023) Algorithmic management: its implications 18.3
for information systems research. Commun AIS 52(1):518–537. Ganguli D, Askell A, Schiefer N, Liao T, Lukosˇiu¯te_ K, Chen
[Link] A, Goldie A, Mirhoseini A, Olsson C, Hernandez D, et al
Carlson SM, Koenig MA, Harms MB (2013) Theory of mind. (2023)
WIREs Cogn Sci 4:391–402 The capacity for moral self-correction in large language models.
C¸ elikok MM, Peltola T, Daee P, Kaski S (2019) Interactive AI with arXiv:2302.07459
a Garcia T (2023) David Guetta replicated Eminem’s voice in a
theory of mind. In: ACM CHI 2019 workshop: computational song using artificial intelligence.
modeling in human-computer interaction, vol 80, pp 4215–4224 [Link] news/david-guetta-eminem-
Chen L, Zaharia M, Zou J (2023) How is chatgpt’s behavior artificial-intelligence-1235516924/, accessed 25 Aug 2023
changing Gilardi F, Alizadeh M, Kubli M (2023) ChatGPT outperforms
over time? arXiv:2307.09009 crowd- workers for text-annotation tasks. arXiv:2303.15056
Chen M, Tworek J, Jun H, Yuan Q, Pinto HPdO, Kaplan J, Edwards Gimpel H, Hall K, Decker S, Eymann T, La¨mmermann L, Ma¨dche A,
H, Burda Y, Joseph N, Brockman G, et al (2021) Evaluating Ro¨glinger M, Ruiner C, Schoch M, Schoop M, et al (2023)
large language models trained on code. arXiv:2107.03374 Unlocking the power of generative ai models and systems such as
Chiang T (2023) ChatGPT is a blurry JPEG of the web. [Link] GPT-4 and ChatGPT for higher education. [Link]
[Link]/tech/annals-of-technology/chatgpt-is-a-blurry- [Link]/fileadmin/einrichtungen/digital/Generative_AI_
jpeg-of-the-web, accessed 25 Aug 2023 and_ChatGPT_in_Higher_Education.pdf, accessed 25 Aug
Davison RM, Laumer S, Tarafdar M, Wong LHM (2023) ISJ 2023
editorial: pickled eggs: generative AI as research assistant or co- Goldman Sachs (2023) Generative AI could raise global GDP by
author? Inf Syst J Early View. [Link] 7%. [Link]
De-Arteaga M, Feuerriegel S, Saar-Tsechansky M (2022) Algorith- [Link]
mic fairness in business analytics: directions for research Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D,
and Ozair S, Courville A, Bengio Y (2014) Generative adversarial
practice. Prod Oper Manag 31(10):3749–3770 nets. Adv Neural Inf Process Syst 27:2672–2680
Debortoli S, Mu¨ller O, vom Brocke J (2014) Comparing business Gray HM, Gray K, Wegner DM (2007) Dimensions of mind
intelligence and big data skills. Bus Inf Syst Eng 6(5):289– perception. Sci 315(5812):619–619
300. [Link] Grisold T, Groß S, Stelzl K, vom Brocke J, Mendling J, Ro¨glinger
Dellermann D, Ebel P, So¨llner M, Leimeister JM (2019) Hybrid M, Rosemann M (2022) The five diamond method for
intelligence. Bus Inf Syst Eng 61(5):637–643. [Link] explorative business process management. Bus Inf Syst Eng
1007/s12599-019-00595-2 64(2):149–166. [Link]
Devlin J, Chang MW, Lee K, Toutanova K (2018) BERT: Pre- Haag S, Eckhardt A (2017) Shadow IT. Bus Inf Syst Eng
training of deep bidirectional transformers for language 59(6):469–473. [Link]
under- standing. arXiv:1810.04805 Haase J, Hanel PHP (2023) Artificial muses: generative artificial
Diederich S, Brendel AB, Kolbe LM (2020) Designing anthropo- intelligence chatbots have risen to human-level creativity. arXiv:
morphic enterprise conversational agents. Bus Inf Syst Eng 2303.12003
62(3):193–209 Hartmann J, Bergner A, Hildebrand C (2023a) MindMiner: uncov-
Dietz M, Pernul G (2020) Digital twin: empowering enterprises ering linguistic markers of mind perception as a new lens to
towards a system-of-systems approach. Bus Inf Syst Eng understand consumer-smart object relationships. J Consum
62(2):179–184. [Link] Psy- chol. [Link]
Dolata M, Feuerriegel S, Schwabe G (2022) A sociotechnical view Hartmann J, Schwenzow J, Witte M (2023b) The political ideology
of algorithmic fairness. Inf Syst J 32(4):754–818 of
van Dun C, Moder L, Kratsch W, Ro¨glinger M (2023) ProcessGAN: conversational AI: converging evidence on ChatGPT’s pro-
supporting the creation of business process improvement environmental, left-libertarian orientation. arXiv:2301.01768
13 ideas
through generative machine learning. Decis Support Syst
Hawlitschek F (2023) Interview with Samuel Tschepe on ‘‘Quo vadis
design thinking?’’. Bus Inf Syst Eng 65(2):223–228. [Link]
165(113):880. [Link] org/10.1007/s12599-023-00792-0

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126 (2024) 125

Herm LV, Janiesch C, Reijers HA, Seubert F (2021) From symbolic Maedche A, Legner C, Benlian A, Berger B, Gimpel H, Hess T, Hinz
RPA to intelligent RPA: challenges for developing and O, Morana S, So¨llner M (2019) AI-based digital assistants:
operating intelligent software robots. In: International opportunities, threats, and research perspectives. Bus Inf
conference on business process management, pp 289–305 Syst Eng 61(4):535–544. [Link]
Hevner A, vom Brocke J, Maedche A (2019) Roles of 00600-
digital innovation in design science research. Bus Inf 8
Syst Eng 61(1):3–8. [Link] Matz S, Teeny J, Vaid SS, Harari GM, Cerf M (2023) The potential
z of generative AI for personalized persuasion at scale.
Ho J, Jain A, Abbeel P (2020) Denoising diffusion probabilistic PsyArXiv
models. Adv Neural Inf Process Syst 33:6840–6851 Matz SC, Kosinski M, Nave G, Stillwell DJ (2017) Psychological
Jakesch M, French M, Ma X, Hancock JT, Naaman M (2019) AI- targeting as an effective approach to digital mass persuasion.
mediated communication: how the perception that profile Proc Natl Acad Sci 114(48):12,714-12,719
text Metz C (2023) Instant videos could represent the next leap in
was written by AI affects trustworthiness. In: Conference on A.I. technology.
human factors in computing systems (CHI) [Link] runway-ai-
Jakesch M, Hancock JT, Naaman M (2023) Human heuristics for [Link], accessed 25 Aug 2023
AI- generated language are flawed. Proc Natl Acad Mirsky Y, Lee W (2021) The creation and detection of deepfakes:
Sci 120(11):e2208839 a survey. ACM Comput Survey 54(1):1–41
Janiesch C, Zschech P, Heinrich K (2021) Machine learning and deep Morana S, Maedche A, Schacht S (2019) Designing process
learning. Electron Market 31(3):685–695. [Link] guidance systems. J Assoc Inf Syst pp 499–535,
1007/s12525-021-00475-2 [Link]
Ji Z, Lee N, Frieske R, Yu T, Su D, Xu Y, Ishii E, Bang YJ, Madotto 1jais.00542
A, Fung P (2023) Survey of hallucination in natural language Ng A, Jordan M (2001) On discriminative vs. generative classifiers:
generation. ACM Comput Surv 55(12):1–38 a comparison of logistic regression and naive Bayes. In:
Kasneci E, Seßler K, Ku¨chemann S, Bannert M, Dementieva Advances in Neural Information Processing Systems, vol 14,
D, Fischer F, Gasser U, Groh G, Gu¨nnemann S, Hu¨llermeier E pp 841–848, [Link]
et al [Link], accessed
(2023) ChatGPT for good? On opportunities and challenges of 25 Aug 2023
large language models for education. Learn Individ OpenAI (2022) Introducing ChatGPT. [Link]
Differ 103(102):274 chatgpt, accessed 25 Aug 2023
Kecht C, Egger A, Kratsch W, Ro¨glinger M (2023) Quantifying OpenAI (2023a) GPT-4 technical report. arXiv:2303.08774
chatbots’ ability to learn business processes. Inf Syst OpenAI (2023b) How should AI systems behave, and who
113(102):176. [Link] should decide? [Link]
Khan J (2021) AI’s carbon footprint is big, but easy to reduce, behave, accessed 25 Aug 2023
Google Park JS, O’Brien JC, Cai CJ, Morris MR, Liang P, Bernstein MS
researchers say. Fortune (2023) Generative agents: interactive simulacra of human
Kingma DP, Welling M (2013) Auto-encoding variational Bayes. behavior. arXiv:2304.03442
[Link] Peres R, Schreier M, Schweidel D, Sorescu A (2023) On ChatGPT
Klotz S, Westner M, Strahringer S (2022) Critical success factors of and beyond: how generative artificial intelligence may
business-managed IT: it takes two to tango. Inf Syst Manag affect
39(3):220–240 research, teaching, and practice. Int J Res Market 40:269–275
Kraus M, Feuerriegel S, Oztekin A (2020) Deep learning in Rabinowitz NC, Perbet F, Song HF, Zhang C, Eslami SMA,
business analytics and operations research: models, Botvinick MM (2018) Machine theory of mind. In:
applications and International conference on machine learning, PMLR, vol 80,
managerial implications. Europ J Oper Res 281(3):628–641. pp 4215–4224,
[Link] [Link]
Kreps S, McCain RM, Brundage M (2022) All the news that’s fit to accessed 25 Aug 2023
fabricate: AI-generated text as a tool of media Rai A (2020) Explainable AI: from black box to glass box. J
misinformation. Acad Market Sci 48:137–141
J Exp Polit Sci 9(1):104–117 Ramaswamy V, Ozcan K (2018) What is co-creation? An interac-
Kru¨gel S, Ostermaier A, Uhl M (2023) ChatGPT’s inconsistent moral tional creation framework and its implications for value creation. J
advice influences users’ judgment. Sci Report 13(1):4569 Bus Res 84:196–205
Lasi H, Fettke P, Kemper HG, Feld T, Hoffmann M (2014) Industry Reisenbichler M, Reutterer T, Schweidel DA, Dan D (2022)
4.0. Bus Inf Syst Eng 6(4):239–242. [Link] Frontiers: supporting content marketing with natural language
s12599-014-0334-4 generation. Market Sci 41(3):441–452
Li Y, Choi D, Chung J, Kushman N, Schrittwieser J, Leblond R, Rombach R, Blattmann A, Lorenz D, Esser P, Ommer B (2022) High-
Eccles T, Keeling J, Gimeno F, Dal Lago A et al (2022) resolution image synthesis with latent diffusion models. In:
Competition-level code generation with alphacode. IEEE/CVF conference on computer vision and pattern
Science 378(6624):1092–1097 recogni- tion, pp 10684–10695
Liu P, Yuan W, Fu J, Jiang Z, Hayashi H, Neubig G (2023) Pre-train, Sandkuhl K, Fill H, Hoppenbrouwers S, Krogstie J, Matthes F,
prompt, and predict: a systematic survey of prompting Opdahl AL, Schwabe G, Uludag O¨ , Winter R (2018) From
methods in natural language processing. ACM Comput Surv
expert discipline to common practice: a vision and
55(9):1–35 Longoni C, Fradkin A, Cian L, Pennycook G (2022)
research agenda for extending the reach of enterprise
News from generative artificial intelligence is believed less.
modeling. Bus Inf Syst Eng 60(1):69–80.
In: ACM conference on fairness, accountability, and
[Link] y
transparency
Schoormann T, Mo¨ller F, Hansen MRP (2021) How do researchers
(FAccT), pp 97–106
(re-)use design principles: An inductive analysis of
Maarouf A, Ba¨r D, Geissler D, Feuerriegel S (2023) HQP: a cumulative research. In: The Next Wave of Sociotechnical
human- annotated dataset for detecting online propaganda.
arXiv:2304. 14931
13
Design, Springer, Cham, Lecture Notes in Computer Science,
pp 188–194, https:// [Link]/10.1007/978-3-030-82405-1_20
Maedche A, Morana S, Schacht S, Werth D, Krumeich J (2016)
Advanced user assistance systems. Bus Inf Syst Eng 58:367–
370 Content courtesy of Springer Nature, terms of use apply. Rights reserved.
126 S. Feuerriegel et al.: Generative AI, Bus Inf Syst Eng 66(1):111–126
(2024)
Schoormann T, Stadtla¨nder M, Knackstedt R (2023) Act and Susarla A, Thatcher RGJB, Sarker S (2023) Editorial: the janus effect
reflect: integrating reflection into design thinking. J Manag of generative AI: charting the path for responsible conduct
Inf Syst 40(1):7–37. of scholarly activities in information systems. Inf Syst
[Link] Res 34(2):399–408. [Link]
Schoormann T, Strobel G, Mo¨ller F, Petrik D, Zschech P (2023) Sutskever I, Vinyals O, Le QV (2014) Sequence to sequence learning
Artificial intelligence for sustainability: a systematic review with neural networks. Adv Neural Inf Process Syst
of information systems literature. Commun AIS 52(1):8 27:3104–3112
Schramowski P, Turan C, Andersen N, Rothkopf CA, Kersting K Teubner T, Flath CM, Weinhardt C, van der Aalst W, Hinz O (2023)
(2022) Large pre-trained language models contain human- Welcome to the era of ChatGPT. Bus Inf Syst Eng 65(2):95–
like biases of what is right and wrong to do. Nat Machine 101.
Intell 4(3):258–268 [Link]
Schwartz R, Dodge J, Smith NA, Etzioni O (2020) Green AI. Unsal S, Atas H, Albayrak M, Turhan K, Acar AC, Dog˘an T
Commun ACM 63(12):54–63 (2022) Learning functional properties of proteins with language
Scho¨bel S, Schmitt A, Benner D, Saqr M, Janson A, Leimeister JM models.
(2023) Charting the evolution and future of conversational Nat Machine Intell 4(3):227–245
agents: a research agenda along five waves and new frontiers. van der Aalst WMP, Bichler M, Heinzl A (2018) Robotic process
Inf Syst Front. [Link] automation. Bus Inf Syst Eng 60(4):269–272. [Link]
Senoner J, Netland T, Feuerriegel S (2022) Using explainable 1007/s12599-018-0542-4
artificial intelligence to improve process quality: evidence from Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN,
semiconductor manufacturing. Manag Sci 68(8):5704–5723 Kaiser Ł, Polosukhin I (2017) Attention is all you need. Adv
Shin M, Kim J, van Opheusden B, Griffiths TL (2023) Superhuman Neural Inf Process Syst 30:6000–6010
artificial intelligence can improve human decision-making by Vernadat F (2020) Enterprise modelling: research review and
increasing novelty. Proc Natl Acad Sci 120(12):e2214840,120 outlook.
Shollo A, Hopf K, Thiess T, Mu¨ller O (2022) Shifting ML value Comput Indust 122(103):265. [Link]
creation mechanisms: a process model of ML value 2020.103265
creation. Vidgof M, Bachhofner S, Mendling J (2023) Large language models
J Strateg Inf Syst 31(3):101,734. [Link] for business process management: opportunities and
2022.101734 challenges.
Siebers P, Janiesch C, Zschech P (2022) A survey of text represen- In: Business process management forum. Lecture Notes in
tation methods and their genealogy. IEEE Access 10:96,492- Computer Science, Springer, Cham, pp 107-123
96,513. [Link] von Zahn M, Feuerriegel S, Kuehl N (2022) The cost of fairness
Silva N, Sousa P, Mira da Silva M (2021) Maintenance of enterprise in AI: evidence from e-commerce. Bus Inf Syst Eng 64:335–
architecture models. Bus Inf Syst Eng 63(2):157–180. 348
[Link] org/10.1007/s12599-020-00636-1 Wolfe R, Banaji MR, Caliskan A (2022) Evidence for hypodescent
Slack D, Krishna S, Lakkaraju H, Singh S (2023) Explaining in
machine learning models with interactive natural language visual semantic AI. In: ACM conference on fairness, account-
conversations using TalkToModel. Nat Machine Intell 5:873– ability, and transparency, pp 1293–1304
883 Ziegler DM, Stiennon N, Wu J, Brown TB, Radford A, Amodei D,
Smits J, Borghuis T (2022) Generative AI and intellectual Christiano P, Irving G (2019) Fine-tuning language models from
property rights. Law and artificial intelligence: regulating AI human preferences. arXiv:1909.08593
and applying ai in legal practice. Springer, Heidelberg, pp 323– Zilker S, Weinzierl S, Zschech P, Kraus M, Matzner M (2023) Best
344 of
Spitale G, Biller-Andorno N, Germani F (2023) AI model GPT-3 both worlds: combining predictive power with interpretable and
(dis) explainable results for patient pathway prediction. In:
informs us better than humans. Sci Adv 9:eadh1850 Proceed- ings of the 31st European Conference on
Strobelt H, Webson A, Sanh V, Hoover B, Beyer J, Pfister H, Rush Information Systems (ECIS), Kristiansand, Norway
AM (2023) Interactive and visual prompt engineering for ad- Zschech P, Horn R, Ho¨schele D, Janiesch C, Heinrich K (2020)
hoc task adaptation with large language models. IEEE Intelligent user assistance for automated data mining
Transact Visual Comput Graphics 29(1):1146–1156. method selection. Bus Inf Syst Eng 62(3):227–247.
[Link] 1109/TVCG.2022.3209479 [Link] 1007/s12599-020-00642-3

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature
in writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

onlineservice@[Link]