F E AT U R E
Unmasking AI:
Informing Authenticity
Decisions by Labeling
AI-Generated Content
Olivia Burrus, Adobe
Amanda Curtis, Adobe and Oxford University
T
Laura Herman, Adobe and Oxford University
Insights The rapid proliferation of artificial understanding, as well as for ensuring
→ AI labeling can be a valuable intelligence technologies has ushered accountability and support for
tool in the fight against mis- and in a transformative era in content responsible AI development and
disinformation, helping viewers creation, and with it, shifting deployment.
understand the potential for perceptions of what makes content Many researchers have explored
biases or manipulations. authentic. The notion of authenticity AI-labeling solutions to support this
→ Creators wish to demonstrate (i.e., the accuracy, reliability, and type of transparency for viewers,
their creative process, while trustworthiness of content) becomes which involves providing metadata,
viewers seek information to particularly fraught in the context of tags, or annotations to content
inform authenticity perceptions. content made using generative AI produced by AI models [2,3,4]. Such
→ We need participation in industry- (GenAI) tools [1] such as DALL-E, studies have shown that surfacing
wide collaborations, adherence ChatGPT, or Adobe Firefly. Providing GenAI labels to consumers with text
to legislative measures, and transparency means presenting or graphics displayed on or next to
the integration of provenance viewers with clear and understandable content can be a valuable tool in the
information at both the creation information about how content was fight against mis- and disinformation,
and display stages of AI-generated created, and disclosing the role of AI in helping viewers understand the
content. that process, if any. Transparency is potential for biases or manipulations in
essential for fostering viewer trust and the content they are consuming. For
38 INTER ACTIONS J U LY – A U G U S T 2 0 2 4 @INTER ACTIONSMAG
example, Chloe Wittenberg et al. [4] metadata about images. In choosing to creative ownership. But what trust
propose a framework for labeling remain impact-agnostic, we wanted to signals do viewers need, what do
AI-generated media according to the understand whether we could creators want to share, and how can we
goal that specific labels are intending empower users to make their own trust design a truly effective labeling
to accomplish. The authors propose decisions without making any value system?
either process-based transparency (i.e., judgements for them (Figure 1).
O
communicating to users how and from Whether it is for news media, arts, OVERVIEW OF
where a particular piece of content or product marketing, effective labels OUR RESEARCH STUDIES
originated while “remaining agnostic should enable viewers to quickly Over the past two years, we have
I M A G E B Y S H U T T E R S T O C K A I G E N E R AT O R
about the potential consequences of recognize AI’s involvement, allowing conducted a number of interviews and
that content for viewers or society” [4]) them to evaluate source credibility, surveys exploring viewer and content
or harm-based transparency (i.e., verify content accuracy, gain creator perceptions of authenticity in
communicating to users that a piece of contextual knowledge, and ultimately AI-generated news, influencer
content is potentially misleading or make their own informed decisions marketing, and art. These studies were
deceptive). Building on this body of around trust and authenticity. AI part of our larger research initiative to
work, we conducted research on a labels should also empower content explore the future of digital trust, how
process-based transparency approach creators by recognizing innovation, to best mitigate harms, and how to
that presents viewers with key mitigating stigma, and reinforcing empower creators and viewers. Our
INTER ACTIONS. [Link] J U LY – A U G U S T 2 0 2 4 INTER ACTIONS 39
F E AT U R E
participants have been primarily gender identity, and political creators to use such a labeling system.
North American creators such as art affiliations). We felt it critical to gain In this piece, we report on high-
photographers, illustrators, photo and perspectives from both creators, who level insights synthesized across these
video journalists, and short-form video need to adopt this system, and viewers, studies. In our first study, we
influencers (using TikTok or Instagram who need to make sense of and use it. conducted ethnographic interviews on
Reels), as well as mostly North Future research should also explore digital trust with seven content
American viewers of these creative the needs of the implementers of AI creators and 10 viewers. We then dug
outputs. We purposefully sampled a labeling, including both platforms that deeper into creators’ perceptions of
wide range of viewers and creators enable the information to be surfaced authenticity and AI-labeling solutions
from the U.S. and Canada with diverse to viewers (social media sites, Web by conducting a second study with 10
lived experiences (including browsers, news media, etc.) as well as creators, observing their interaction
socioeconomic backgrounds, ethnicity, the creative tools that would allow with an AI-labeling prototype
embedded within a creative tool. In
this study, creators explored what such
labels could and should look like during
their creative processes. Our third
study was a concept test, in which we
interviewed 19 viewers and asked
them to interact with an example of AI
labeling on social media posts. Our
fourth study aimed to prioritize
possible trust signals—specific
metadata, tags, or annotations—based
on what creators and viewers would
most want to share or see, respectively.
This mixed-methods study included 66
creators (photojournalists and
influencers) and 126 viewers of this
Figure 1. Example of process-based transparency from the Content Authenticity Initiative. type of content; it also evaluated
participant knowledge of GenAI and
perceptions of its impact on content
authenticity. Finally, we conducted
two back-to-back unmoderated studies
with a total of 47 viewers of art and
news media to gather iterative
feedback on updated AI-labeling
prototypes that incorporated the
insights from the previous four studies.
Our findings have allowed us to
critically consider approaches to
generative AI labeling and implications
for creators and corporations. To
further our learnings, we recommend
additional studies that include
international perspectives,
particularly from non-Western and
nondemocratic countries.
CREATOR AND CONSUMER
PERCEPTIONS OF
A
AUTHENTICITY IN
THE CONTEXT OF AI
Across studies, we found that viewers’
perceptions of GenAI change as
contexts change, and thus the
information provided to them needs to
change accordingly. For example, in
one of our studies, viewers believed the
use of GenAI in news content to be a
clear signal that something is
untrustworthy, whereas the same
Figure 2. Example of AI labeling from TikTok (left) and YouTube (right). viewers found the use of GenAI to be
40 INTER ACTIONS J U LY – A U G U S T 2 0 2 4 @INTER ACTIONSMAG
Figure 3. Examples of a proposed content credentialing system in action, following the C2PA 2.0 UX Guidance, found at [Link]
specifications/specifications/1.4/ux/UX_Recommendations.html.
acceptable in the context of influencer- GenAI is, they did perceive that it communicate the involvement of AI in
marketing content. We also found that somehow reduced the authenticity of the creation process, fostering trust
edits perceived by viewers to be digital content. Participants’ confusion with their audiences. AI labeling was
examples of simple auto- around what exactly AI-generated also viewed as a way to showcase
enhancements, such as blemish content is adds a layer of complexity innovative approaches to content
removal or color corrections (whether around already tricky trust decisions, creation and position creators as
accurate or not), have less impact on where viewers continue to center the forward-thinking and adaptive to
perceptions of authenticity than those human creators, not the technologies, emerging technologies. On the other
perceived as complex GenAI in the content creation process. hand, creators who were not using AI
production, such as turning a frown Viewers wanted information about were also interested in sharing
into a smile, or adding, removing, or which elements had been enhanced or information about AI’s lack of
replacing an object. Interestingly, in generated with AI, what information involvement in their process, in order
comparison with their viewers, social that AI was trained on (particularly, to demonstrate the “handmade”
media influencers were less concerned whose work was used to train the AI nature of their work. So whether or not
about the impact of AI-powered model), and how to recenter the human creators are using AI, our research
auto-enhancements, but they were work that goes into content creation. indicated that most are not only
more concerned about changes made Based on these insights, we willing to adopt AI labeling but also
using GenAI. In sum, AI has a time recommend considering the content see value in doing so for themselves
and place in content creation. GenAI is creator’s intended audience, message, and society.
closely tied with mis- and goals, process, and platform when
MITIGATING POTENTIAL
T
disinformation for many of our designing a labeling system for
participants, and as such, they found it authenticity. We also suggest being CHALLENGES OF
to be more appropriately used in mindful of terminology, without AI-LABELING SOLUTIONS
“lower stakes” situations than news assuming that the average person There are a number of challenges to be
reporting. It is critical for viewers of knows what exactly generative AI considered when developing labeling
AI-labeling systems to have entails. solutions for generative AI:
information about the context of the When we shared our prototyped • First, it is critical for labeling
content in question. Further, how, labeling system with content creators, mechanisms to inspire consumer trust
when, and where GenAI is used in the many were open to this approach of in the system itself, as misleading labels
creation process is paramount, and any sharing an overview of their process can severely undermine viewers’
labeling solutions should carefully with their audience. Creators also saw confidence in the entire media
consider how to display this the value in clarifying the specific role ecosystem. This assurance is invaluable
information. AI played in the content generation in an environment where
It is important to note that many and editing processes. Specifically, misinformation and manipulation can
viewers did not know what generative content creators who were creating have far-reaching consequences.
AI meant, even if they reported using AI-generated content viewed AI • Second, AI labeling should provide
generative AI tools. Specifically, there labeling as a means to transparently transparency while also aligning with
were two areas of confusion. First, the
term itself was confusing to the
average person, where they understood
that tools such as DALL-E used AI, but
they had not encountered the term Effective labels should enable viewers
generative AI before. Second, it was not to quickly recognize AI’s involvement,
allowing them to evaluate source
clear to our participants if using the
label “generative AI” implied that an
entire piece of content was created
credibility, verify content accuracy, and
gain contextual knowledge.
using AI or if AI was used in part of the
process. Without fully knowing what
INTER ACTIONS. [Link] J U LY – A U G U S T 2 0 2 4 INTER ACTIONS 41
F E AT U R E
viewers’ mental models and information parts of that content) has been at both the creation and display stages
needs. However, substantial challenges generated by AI. Both the companies of AI-generated content. Through such
are presented by issues such as the developing GenAI software and the concerted efforts, the industry can
degree of algorithmic intervention and platforms that display AI-generated collectively advance toward a future
the temporal considerations of when in content have a responsibility to where transparency, user control, and
the creative process AI was employed. mitigate their users’ authenticity trust define the landscape of AI-
Viewer cognitive load further concerns, which are exhibited by generated content.
compounds this issue, requiring a creators and viewers of online content.
delicate balance between providing Companies should be proactively Endnotes
1. Epstein, Z. et al. Art and the science of
sufficient information and avoiding engaging in efforts to increase
generative AI. Science 380, 6650 (2023),
overwhelming users with technical transparency, accountability, and user 1110–1111.
details [5]. control over AI-generated content. 2. Epstein, Z., Mengying C.F., Arechar, A.A.,
• Third, a widespread focus on Creators would like to demonstrate the and Rand, A. What label should be applied
AI-generated content could mechanisms and processes by which to content produced by generative AI?
inadvertently foster viewer skepticism their content was made, while viewers PsyArxiv preprint, 2023. DOI: 10.31234/
toward media more broadly [3], as they seek information to inform their trust [Link]/v4mfz
3. Hoes, E., Aitken, B., Zhang, J., Gackowski,
become aware of the presence of AI in decisions. Therefore, creation software
T., and Wojcieszak, M. Prominent
the media ecosystem. This could result tools should embed provenance misinformation interventions reduce
in viewers discounting authentic information into the content they misperceptions but increase skepticism.
information. Such discounting could in produce, and content platforms should PsyArxiv preprint, 2023. DOI:10.31234/
turn create situations in which decision display this information in a readily [Link]/zmpdu
makers and influential figures dismiss accessible manner. 4. Wittenberg, C., Epstein, Z., Berkinsky,
A.J., and Rand, D.G. Labeling AI-
legitimate media as fake [4,6]. There is growing interest in labeling
generated content: Promises, perils, and
• Fourth, there is the risk of totalizing content generated by AI, as companies future directions. Topical Policy Brief,
and stigmatizing AI by labeling its including OpenAI, Adobe, YouTube MIT Schwarzman College of Computing,
content, perpetuating biases against (Google), Meta, and TikTok adopt this 2023; [Link]
GenAI and the creators that use it. As practice (see examples in Figure 2). content/uploads/2023/11/AI-Policy_
such, stringent labeling requirements This momentum is being reinforced by [Link]
may impede the adoption of AI in legislative measures such as the U.S. 5. Sweller, J. Cognitive load during problem
solving: Effects on learning. Cognitive
creative processes by causing content Congress’s AI Disclosure Act, the
Science 12, 2 (1988), 257–285.
creators to shy away from it, dismissing European AI Act, and similar 6. Ternovski, J., Kalla, J., and Aronow, P.M.
possible positive use cases for GenAI. regulations emerging from other The negative consequences of informing
• Fifth, mandatory labeling has the political bodies globally. These voters about deepfakes: Evidence from
potential to erode the sense of creative regulatory efforts are largely focused two survey experiments. Journal of
ownership among creators by forcing on governance and the ethical Online Trust and Safety 1, 2 (2022).
them to reveal the use of a tool that is application of AI technologies, DOI:10.54501/jots.v1i2.28
currently the subject of strong public stressing transparency and
Olivia Burrus was most recently leading UX
debates around copyright, originality, accountability in AI systems. research at Adobe for the Content Authenticity
and authorship. Effective AI labeling Recognizing the importance of Initiative, an effort focused on bringing more
should address this by serving to industry-wide collaboration, self- transparency to digital content by sharing how
reinforce the sense of creative regulation, and adherence to emerging content was made, who made it, and how it
ownership, assuring creators that their regulations, initiatives like the changed over time.
unique vision remains central to the cross-sector consortium the Coalition → oliviampburrus@[Link]
artistic process. for Content Provenance and Amanda Curtis previously led research for
These issues underscore the Authenticity (C2PA) provide an open Adobe’s Content Authentivity Initiative and is
importance of balancing the potential access standard for embedding currently a user experience researcher and
Ph.D. student at Oxford University’s Internet
risks and benefits of labeling AI- provenance information into the
Institute focusing predominantly on gaming,
generated content to foster a more publishing, creation, and consumption immersive, and emerging technologies.
I
nuanced and constructive approach in experiences of online content (Figure She researches video game experiences,
the evolving landscape of AI and 3). knowledge creation, and creativity.
media. This research has shown that there → [Link]@[Link]
is a clear need for a labeling system Laura Herman specializes in emerging
CALL FOR A DEEPLY that takes a deeply researched technologies’ impact on creative practices. She
RESEARCHED, INDUSTRY- approach, responding to cultural shifts is a research leader at Adobe and a doctoral
WIDE APPROACH TO AI in awareness, comprehension, and researcher at Oxford University’s Internet
LABELING trust of GenAI and its outputs. We Institute. She has held research roles at
Harvard, Princeton, and Intel and has worked
In light of the increasing sophistication need active participation in industry- with arts institutions such as the Serpentine
of GenAI technology, there is an wide collaborations, adherence to Galleries, the Tate, Studio Olafur Eliasson, and
urgent need for user-friendly solutions legislative measures, and the Ars Electronica.
that disclose which content (and which integration of provenance information → lherman@[Link]
DOI: 10.1145/3665321 COPYRIGHT HELD BY AUTHORS. PUBLICATION RIGHTS LICENSED TO ACM. $15.00
42 INTER ACTIONS J U LY – A U G U S T 2 0 2 4 @INTER ACTIONSMAG