Kingdom of Saudi Arabia

Ministry of Education
Aljouf University
College of Arts and Sciences

A comparative analysis of Al and human

translations of cultural texts from Arabic
to English

Prepare By Student:
Abdulrahman Salim Al-Suhaimi
ID: 401110654
Instructor:
Khaled Mutlaq Al-Harbi
Introduction
For years now, translation has been the means through which different peoples and cultures have
communicated with each other. It has traditionally been the practice of replacing the words of a
source text with equivalent words in a target text, often needing a high degree of cultural context,
grammatical and linguistic skills. Things have taken a radical shift through the years with the
introduction of AI based systems now commonly referred to as machine translation (MT). With
the likes of Google Translate, DeepL and OpenAI’s ChatGPT systems, translation does not take
rocket science anymore. This wider system of translation has brought with it aspects that never
existed before such as speed, affordability and easy access at an unprecedented scale.

While these new AI based translation systems are on the rise, it begs the question of whether they
exceed humans in translation work. While the AI systems can complete scans on thousands of
texts in a matter of seconds, there are questions as to whether AI can maintain things such as
idioms, tone and cultural context. Human translators in contrast remain more intuitive, creative
and culturally aware ,traits that machines so far have been unable to replicate.

The present research sets out to compare, in a systematic manner, the advantages and
shortcomings of AI translators with the human ones. We endeavor to assess their capabilities for
diverse domains such as the translation of technical and medical texts, works of literature and
poetry among others. In addition, this work analyzes the use of AI translation including the social
and political aspects of the technology such as bias, addiction to AI and alterations to the
professional translator role in the future.

This research will answer the question: Is there a future for both human and AI translation and
how exactly will both work together? Additionally, the research will take a hybrid approach that
combines qualitative and quantitative research methods for a complete assessment of human and
AI translation capabilities to their existence. The results will be useful not only for businesses
and scholars but also for ordinary people who are able to communicate in several languages.

1
Research Objectives

The phenomenon of the globalization of the economy, international relations, and the emergence
of transnational networks raises new perspectives and research questions about the role of
artificial intelligence (AI) in the English language. The aim of this research is to evaluate the role
of AI in the translation field and in this aspect – how AI can complement human translation in
the future.

Quality, Accuracy, and Cultural Honesty

Overriding aim is to evaluate the quality and accuracy as well as a cultural context of human
translations in comparison with AI translations. Although AI translation services such as one
offered by Google Translate and DeepL have employed neural machine translation (NMT) and
provide accurate results in multiple scenarios, limitations definitely persist in rendering
figurative language, context, and cultural subtleties (Toral et al., 2018). Language users are able
to contextualize the language meaning into that appropriate for a given cultural context,
something which is particularly needed in the translation of literary works and other culture
oriented works. Applying both methods to the output evaluation of different content types will
aid in pointing out the advantages and disadvantages of both approaches in providing accurate
and cultural translations.

Speed and Cost-Efficiency Another important aspect is the speed and cost-effectiveness of AI
translation as compared to human translators. AI systems reduce the time and expenses incurred
in the business and hence are likely to be lauded by companies wishing to have economies of
scale (Lilt, 2020). Human translation can be accepted however slow and costly: the translation of
complex or sensitive texts like legal or medical documents, tends to be of higher quality (Gaspari
et al., 2015). By establishing the input time and money costs associated with each of the
methods, this study will contribute in overcoming the synergies defining the reliance of AI
translation where it would cost less and where quality would be available. User Perceptions
across Domains Every user perception becomes an underlying attribute that determines that
success of translation systems. However, users of various fields such as, legal, medical, business,
and creative will expect a certain degree of precision and creativity. It has been evidenced in

2
several studies that in such sensitive fields like medicine, users sometimes think that human
translators are more contextualized, more reasonable, and more prepared (Castilho et al., 2017).
This study will seek to assess the user reviews on both types of translation to find out how trust,
preference and satisfaction are portrayed in these areas.

Limitations, ethical issues, and their impact on the problem

In spite of tremendous achievements in development, AIs improvement has its own setbacks. For
one thing, They reproduce the same biases contained in the training data, Provide poor output
with low-resource languages, and fail to deliver on context-adaptive translations (Bender et al.,
2021). Other ethical correlates include the aspects of data protection, the risk of adult translator
displacement, and excessive dependence on AI tools. Evaluating these limitations will shed light
on the burden challenges associated with the growing presence of AI in translation technology
and practice.

Translation scenarios with an Addition of AI

Lastly, this research will discuss the possibility of integrating AI and human participation in
translation. Hybrid approaches that incorporate AI are becoming increasingly common: AI drafts
an initial version that is revised by professionals (Federico et al., 2020). This study will identify
situations where hybrid designers correspond accurately and propose hybridization designs for
translation workflows to identify how to get the best out of the two approaches.

Thus defining the objectives of the study, this research will makes it possible to have a deep
understanding of the effectiveness of the human AI enabled translator and his role in today’s
multilingual globalized world.

Research Questions

1. How do humans and AI explain languages in their view what are the pros and cons of
both human and artificial intelligence?

With each human and AI translation, English also produces respect challenges and
advantages that are associated with the translation processes. Translators overtime have been
exceptional in translating not only words but also phrases, cultures, and people that are
context-rich and require creative and cultural coherence (Toral et al., 2018). These are the

3
 advantages that are necessary in fields, such as literature, where there is an emphasis on tone
and other stylistic features. Nevertheless, it is important to mention that human translation
can be regarded as a lengthy and expensive looking which may be a drawback for outreach
on a larger scale or tasks where time is of the essence (Gaspari et al., 2015). Impressive
accuracy and cost effectiveness, mainly for simple or repetitive text tasks is however the
strength of AI translation systems which through neural machine translation (NMT) have
great speed (Vashee, 2020). Although, AI is having this competence, it is such that it cannot
deal with ambiguity as it often makes mistakes with intricate syntax or when dealing with
highly localized semantics (Läubli et al., 2020).

2. How does AI compare to actual language users in translating idioms, cultural elements,
and contextual features?

The automatic translation of idioms and other parts of contexts has improved considerably
with the development of neural language models, for example, Google Translate or DeepL
(Bahdanau et al., 2015). Nevertheless, they do not reach the performance of a professional
translator, who is fully capable of deciphering the idioms or context-specific language, more
precisely and accurately. As an example, AI has a tendency to translate idioms and phrases
literally, which often results in misconstrued idioms (Toral et al, 2018). One more area which
is a problem is the context; AI can easily confuse words such as homonyms or even fail to
provide the proper tone even if the target audience is other people (Bender et al, 2021). In
contrast, the translation performed by human interpreters is more oriented, because, while
working on the task, they constantly rely on a lot of socio-cultural knowledge as well as their
own personal experience.

3. In what circumstances do users gravitate towards human translations and in which

circumstance do they prefer AI translations?

User preferences for human versus AI translation depend mostly on the situation and the
outcomes associated with it. In high risk situations like legal and medical translations, users
tend to opt for human translators because they are perceived as more accurate (Castilho et al,
2017). Likewise, the creative sectors, for instance in literature or marketing, also find
translation by people useful as it carries emotional weight. In contrast, it is AI translation that
is preferred for low-risk, or simply mundane procedures, whereby translation involves a
4
 simple set of instructions or translation of documents that are not crucial (Gaspari et al,
2015). AI translation is quick and economical and is therefore excellent for organisations that
deal with great amounts of material.

4. How does AI translation technology work and where are its limitations?

Even though AI translation systems are progressing quite fast, they still face certain
technological limitations. Koehn & Knowles (2017) point out that AI has problems with low
resource languages, as there is not enough of them for training. Also, we have seen that
machine language translation systems are not free from the biases of the training dataset,
leading to problematical ethical outputs (Bender et al., 2021). The inability of AI to
comprehend and interpret context in its entirety remains a significant obstacle, because
translations tend to be more rigid and less adaptable than the work of humans (Läubli et al.,
2020). As such, these limitations also illustrate the vast chasm that exists between the current
capabilities of AI and the standard of translation excellence in the professional sphere.

5. Are there ethical issues concerning AI in the context of translation and other inter-
linguistic communication technologies?

This issue is quite pertinent in case of the increasing use of AI in translation. A major concern
is the loss of human translators especially in the contexts where AI use is widespread
(Gaspari et al., 2015). In addition, AI even may incorporate some inappropriate language or
biased opinions as an artificial intelligence system can trained on biased data (Bender et al.,
2021). Data confidentiality is another such issue since the translated text can be stored or
abused by these translation service platforms (Toral et al., 2018). Such ethical issues need to
be addressed by being open about AI usage in translation and finding a middle ground on
how these translation technologies should be used effectively.

Literature Review

Scope

The emergence and popularization of specific systems of machine translation (Google Translate,
DeepL, models from OpenAI, etc) has changed the dimension completely. This study aims to
analyze the time course perspective of machine translation as one which has developed from

5
logic-based models to the current reliance on neural networks. It also studies case studies that
specifically pit both AI and human translations side by side, neural networks’ and deep learning’s
emergence as key factors in modern MT’s development approaches and AI’s usage for translation
purposes, ethics of use including biases, privacy issues, etc.

Historical Evolution of Machine Translation

Machine translation has experienced quite a lot of changes from the time it was invented. The
first systems were based on the so-called rules and the language was taken care of through the
use of dictionaries prepared capable of that (Hutchins, 2005). While such systems were
revolutionary, they had a number of problems with operationalizing the different dimensions of
natural language thus most times resulting in clumsy or incorrect translations. Then came the
1980s and 1990s and a new form of translation system known as statistical machine translation
emerged through the use of having large bilingual corpora and using complex mathematical
predictions in search of word placements and sentence structure (Koehn, 2009). These systems
managed to greatly improve upon the older models but as the case has always been with
advancing technology steps STL systems such as Google Translate still had problems with
handling context bugs.

In essence, the improved functionality of translation systems, in particular neural machine

translation (NMT) which emerged in the 2010’s and was developed through deep learning
models that displaced the use of phrase based translation, has altered the translation system. The
deep learning models encrypted phrases into vectors enabling the translation to have context
(Bahdanau et al., 2015). As such, Google translate using NMT was able to understand words in a
context of a sentence enabling fluent and precise translations, DeepL on the other hand was able
to use transformer architectures encoders to enable high output across most European languages
(Toral et al., 2018) . Because of this transition, NMT systems are increasingly outshining the
human translators on the basic translation of text documents.

Case Studies Comparing Human and AI Translations

According to research, AI is gaining an upper hand in terms of translation efficiency boost.

However, as AI develops, human intervention may still be required; AI translations of literary
texts , despite appearing coherent, architectural or grammatical ,formatted most common

6
language some linguistic meaning, style and authentisity in a way legal context (Toral & Way,
2018). Läubli, Thordal and Way decided to check the documents in a juridical context in order to
see what standards were maintained while translating from A to B language and how much they
managed to render in the volume. The study revealed the importance of the core message as the
key to indicate success in regards to getting the needed information from the AI systems.
Unfortunately, it is clearly stated that the AI systems in those days were accurate but too slow,
and cost efficient but too incompetent (2020).

Vashee (2020) pointed out that AI systems have taken a front seat when it comes to responding to
customer inquiries and writing product manuals since they can process and understand a lot of
text information reasonably well. Further investigations by Castilho et al. (2017) stress out,
however, that the input need not consist of translations alone as it may also require professional
aspects that are often overlooked, such as editing in translation which is more essential in
marketing. It is demonstrated in such studies that artificial intelligence can well do translation in
terms of speed and volume, but culture specific and socially nuanced translations remain a
domain for human translators.

Cognition of NMT: Neural Networks and Deep Learning

Also, success of modern MT systems has a lot to do with increased utilization of neural networks
and deep learning. Encoder-decoder architecture forms the basis of neural machine translation
and works by first encoding input sentences through continuous embedding vocabulary
representation before decoding them into the target language (Cho et al., 2014). The emergence
of attention mechanisms, especially in transformer models, was transformative for NMT as it
permitted machines to concentrate on certain portions of the input sentence when translating
(Vaswani et al., 2017).

The use of deep learning has solved the major issues users had with dealing with idioms, dealing
with outliers, and also dealing with long-range dependencies (Bahdanau et al., 2015). For
example, Google Translate and DeepL, trained on large-scale parallel corpora, reported high
BLEU scores, which is a yardstick for gauging the quality of translations (Toral et al., 2018).
Additionally, the emergence and rise of the pre-trained language model such as OpenAI’s GPT
has also improved the quality of translation due to the models’ better context and more robust
style uniformity (Brown et al., 2020).
7
Still, there are some unresolved issues. With respect to the training of neural networks, it has
very expensive computational needs and requires a large dataset, thus making them impossible to
be applied to low resource languages (Koehn & Knowles 2017). In addition to, these systems
also have difficulties in tackling domain-specific terminologies or context-specific translations
and usually require post-editing by a professional translator.

Ethical Implications Regarding AI in Translation

Translation is becoming heavily reliant on AI and in this process, a few concerns arise which
include bias, privacy concerns and replacement of humans with other tools. Such bias in AI
translation is usually sourced from a lack of balance in the training set which translates to the AI
being trained on stereotypes or biased cultures. In this way, for instance, Prates et al. (2019)
showed that gender bias appeared in Google Translate where the system managed to e.g.
translate “owned a business” into a male pronoun but “ raised children” into a female pronoun.
Biased assumptions of this kind contribute to the maintenance of oppressive stereotypes and
discredit translations in areas of professionalism.

Privacy bias is other concern but a greater one. Most MT providers conceptualise the need to
input sensitive text for translation and such highlights the need for data protection and ownership
as a key area. Cases noted by Toral et al. (2018) included where the translation provider
accumulated such data from clients without their consent which poses danger in subsects such as
health care and legal sectors.

Furthermore, the growing use of AI translation systems may knock out human translators,
especially from developing countries which price their work very low and get employed in bulk
for various translation jobs (Gaspari et al., 2015). Although AI performs cost-effective measures,
does generate revenue for companies, the capability of general AI to tackle tasks of an intricate
and creative nature is unfortunately quite limited hence retaining the necessity for human
involvement in such endeavors. Preserving the translation profession in practice as a viable way
of earning income while conversely leveraging the economic benefits offered by AI is one of the
major issues that need to be addressed by industry and governmental policies.

Methodology

8
The methodology for this research is intended to allow for the overall assessment of human
translation as well as AI-assisted translation systems. It fuses qualitative and quantitative analysis
with experimental designs and ethical approaches so as to answer the research questions in a
credible manner.

Qualitative Analysis

The qualitative dimension of this study in translation studies is concerned with the contextual
and cultural elements of translating a given text. First, case studies will be undertaken in
translation of literature, medicine, and technical translation. Each of these fields poses its own
unique challenges: literature calls for creativity and sensitivity to cultural aspects, while technical
documents need attention to detail and clarity (Läubli et al., 2020). In this regard, the research
will also analyze outputs of translation created by AI systems as well as those done by human
translators in terms of their contexts.

Second, reports from case studies as well as surveys will aim at understanding the professional
practices and quality expectations of both translators and end-users. Other scholars argue that
nuances are often lost by AI systems, but not by actual translators who know how to use
shorthand phrases and sensitive cultural references (Toral & Way, 2018). User feedback will
facilitate the decision of which translation method, AI or human, is more appropriate with regard
to context and domain.

Finally, all factors of the culture and idioms will also be taken into account, with special regard
on the use of idioms and culturally specific content in human and AI translations. Such aspects
are usually absent or exist in poorly formed states in the output of AI systems because of their
rigid patterns of decodes, while their human equivalents, as a rule, succeed to these tasks (Prates
et al., 2019).

Quantitative Analysis

Quantitative analysis will target measurable aspects of the translation performance. BLEU
(Bilingual Evaluation Understudy) and METEOR (Metric for Evaluation of Translation with
Explicit ORdering), the two best metrics for estimating translation performance of AI systems,
will be used to assess the outcome. These procedures which are mainly applied in the computer-

9
aided translation, provide unbiased intervention by assessing the extent to which AI translations
reach the expected level of reference translations (Koehn & Knowles, 2017).

Besides, user satisfaction surveys will be interpretation in order to obtain dedicators about
translation quality. The respondents will examine the translated versions on the aspects of
grammar, comprehensibility, and suitability for the respective culture. This information may
assist in drawing certain trends of respondent’s hopes to AI or human translation in some
domains.

The evaluation design will focus on the time and expense of translations carried out by AI versus
those done by a human. Owing to their speed and cost efficiency, AI can be utilized in the
performance of high volume but low stakes activities. However, a human translator takes longer
and incurs higher costs but in most cases such complex texts exhibit a higher quality (Vashee,
2020) for the texts.

Experimental Procedure

An experimental design will be adopted whereby, identical texts will be translated using a
number of AI including ChatGPT, Google Translate, and DeepL as well as through human
translators. A group of linguists will assess the translations and divide them into three broad
categories: grammatical correctness, cultural relevance, and stylistic fluency. This type of case
study will help to highlight which areas are best suited for machine translations and which are
best performed by humans (Castilho et al., 2017).

Social Justice Language

A more profound consideration of the ethical issues of using AI for the purposes of translation
will be engaged. In the case of the industry of localization and international communication, this
also involves considering the impact of a high factor of AI dependence. For instance, AI systems
can head the translation industry and capsize the market leaving human agents out of the work
ahighlu like in most systems. This therefore leads to the issue of employability for workers and
the pertinence of their skills (Gaspari et al., 2015).

Furthermore, the research will assess the prejudice which exists in AI systems and its risks in
reinforcing stereotypes. Some studies have shown that biases contained in training datasets can

10
result in bias in translation, mainly in reference to gender and cultural roles (Prates et al., 2019).
The analysis of these biases will also help in expanding the understanding the possible negative
implications of AI translation on the society.

Analysis and Interpretation

Comparison Based on Quantitative and Qualitative Metrics

As for the analysis, the human and AI translations are assessed using both quantitative and
qualitative measures. BLEU and METEOR score – are among the quantitative measures and this
ones seek to establish how successful an AI translation is relative to some written in a text
translation. For instance, it can be expected that DeepL or Google Translate will achieve high
BLEU scores for structured texts like technical manuals, but lower scores when dealing with
idiomatic texts or texts with cultural references, Koehn & Knowles, (2017). However, for
Natural Language Processing, where translation is primarily done by AI, human translations
serve as the gold standard primarily because AI doesn’t perform well as of now.

The qualitative measures have been used to investigate questions like cultural sensitivity,
grammatical correctness, or overall readibility. These investigations are carried out by language
professionals. For example, both human and machine translators have their competitive
advantages: human capabilities are better for literary works which require distinct and
sophisticated nuances, while specialized AI performs well with simple repetitive texts. In this
way, the study maps each perspective and the capabilities associated with them and where the
gaps are located.

Evaluation of Error Patterns

Error patterns are among those issues that are discussed in relation to machine translation in
general and AI translations in particular as an area of weakness in AI grade machine translation
when compared to a human translator's work. For instance, any literal translation of idioms, or
context-based phrases, AI tends to omit the original intent, which often results annoying
translations (Toral et al., 2018). Phrases that involve an element of culture, for example, AI
translates them but loses their intended meaning. Also, AI would find it difficult to easily resolve
transfer of meaning of domain specific terms related to legal or medical writings where adequate
training data is not available.

11
Whereas human translators usually deal with fewer systemic errors, they may also suffer from
non-systemic uniformity error due to specific interpretation as well as lack of stamina for
continued tasks over prolonged period of time. If these models of behavior are analyzed, it is
possible to implement some measures aimed at decreasing the risks, such as creating an outline
using a machine translation, which is then corrected by a human.

Analysis of Survey Results of the End Users

Results from the user surveys present useful information on the user’s preferences and level of
content satisfaction with the human and AI generated translations. However, it appears that users’
preferences lean more towards human translations when content is sensitive or when high stakes
are involved (e.g., legal and creative domains) while low and routine tasks AI output is
acceptable (Castilho et al ,2017). The results of this survey will be structured in a way that will
cut across domains and provide a deeper understanding of the expectations and trust users place
on the translation systems. These insights will then assist in outlining the situations that could
tolerate the use of AI English language translation or human translations as well as the
possibilities of a combination.

Discussion

Information on the Comparison Between the Quality of Human Translations with AI

Translations

Presumably, the result of the investigation would show that even though AI is good in speed,
scale and cost, human translators are much better in understanding theme and creative and
culture coded materials. For instance, AI could fared better in translating technical manuals but it
would probably under perform in translation of poems and novels in which a lot of idiomatical
expressions are present.

Identifying Economically Reachable Domains for both AI and Humans

The investigation will detail situations when AI translations have practical application because
they are effective, such as in technical documents, customer care, or support for non-critical

12
content in several languages. On the other hand, it will accentuate areas where human labor is
not replaceable such as legal, medical, or even literary translations because they demand
accuracy, ethical and cultural concerns, respectively.

Integration of AI and Human Workflows Recommendations

The research will offer practical suggestions for companies and institutions to maximize the
advantages brought by both AI and human translators. Suggestions may include:

 Employing AI for first draft documents to reduce both time and resources.
 Using professional authors for final touches in high profile or imaginative projects.
 Focusing on training AI in relevant fields so as to improve its accuracy in such areas.

Establishing Ethical and Technological Deficiencies

Ethical and technological deficiencies in AI translation will also be listed and include among
others:

 Bias in training datasets consisting of AI which can reinforce stereotypes (Prates et al.,
2019).
 The problem of low resource languages remained poorly addressed.
 Leaking privacy in some sensitive translations (Toral & Way, 2018).

The findings will inspire the creation of translation systems that are more inclusive, context
aware and emphasizing security.

My opinion

Human and AI translation systems exhibit harmonious coexistence, allowing them to strengthen
each other's methods of translation, instead of being at odds with one another. Every method has
certain scopes, and their integration into the processes gives rise to potentials for the
transformation of the translation sector.

AI translation, which includes Google Translate, DeepL, and ChatGPT, amongst other AI’s, has
made it possible for people around the world to communicate seamlessly in different languages.
To translate product descriptions, technical manuals, or even during the initial stage of customer
contact, where lots of materials need to be translated and quality is of secondary importance,

13
their time and cost efficiency is without match. The translation systems are efficient because they
are based on neural networks and transformer models, which means they require the minimum
amount of human participation to generate more accurate outcomes. These system capabilities
only extends to phrases and simple structure based translations, lacking sensitivity within
language and visualization within scope, therefore creating challenges towards more
sophisticated, artistic, or even ethnocentric material. Humor, idioms, and various other contexts
cannot be applied unless it is limited to AI translation, circumstances for whom are impossible to
meet, hence remaining a weak point in the middle of the translation process.

Nevertheless, human translators still possess an understanding, creativity and cultural sensitivity
that is hard to find elsewhere. They are best suited for those areas where attention to detail is of
utmost importance, like legal translation, medical interpretation, or literary work. They have the
capability to comprehend hidden meanings, to modify the voice of the source, which is also the
intent of the author, and to present the material in the proper context. Still, the human translation
is very slow and expensive – so unfit for high, repetitive tasks where the machine excels.

I find that the future will be in fully or partially hybrid workflows. First drafts of AI can be used
and then human professionals can polish the china and a beautiful balance between speed, cost
and quality can be reached. It helps improve both systems – the eyes of human translators are
saved, yet the imagination and contextual vision provided by AI, may not be enough. Integration
is highly hopeful for business and for industries where large amount of materials need translation
for fast turn around time but for time immaterial.

However, the implementation of AI translation systems also incorporates significant ethical

issues such as bias, intrusion of privacy and the possible displacement of human translators.
Overcoming the mentioned challenges of AI systems namely requires design, explainability and
politics for use so that AI complement rather than replace human systems.

To conclude, through translation, speech and communication are not fought between human
beings and AI. It is their joint evolution that enables them to become more sophisticated and
speed up communication and interaction between languages and cultures.

14
Conclusion

The analysis of human and automatic translations indicates that there is an equilibrium which
both systems present with certain advantages and limitations. On the other hand, AI translation
systems have their advantages in terms of speed, scope and costs. This means they are extremely
important in the case of monotone routine and large scale tasks like producing technical
materials or providing support. However, their limitations come into play when such contexts
involve dealing with artistic, imaginative, or culturally deep content where human creates the
dominant advantage because they understand the context, know how to use idioms and such, and
know the culture.

This study addresses the possibilities of AI translations expansion into AI translations’ practical
applicable areas such as the translation of technical documentation and general business

15
correspondence. On the other hand, it emphasizes the need for human translators in legal,
medical and literary translation as the most demanding areas in terms of accuracy, creativity and
cultural context. Surveys and expert analyses confirmed that for the users human translation
means reliable and flexible solution, while AI systems are suitable for inexpensive and quick
translations.

This research points out that combining AI with human professionals can be a way forwards for
the translation. In this instance, O.D. By doing so, they can save time and the end products will
be cheaper: organizations can get high quality translations with fast turnaround, because AI is
used for drafting and the translation is edited by humans.

In addition, the research pinpointed ethical and technological limitations in AI translation,

including AI models' partiality, the risks of invasion of privacy, and lack of advancement in
dealing with underrepresented languages. Tackling these issues would be a matter of constant
development and supervision to foster more inclusive, safe, and efficient AI systems.

References

1. Aljanabi, M., & Mohammed, S. Y. (2024). Advancing translation quality assessment:

Integrating AI models for real-time feedback. Retrieved from ResearchGate
2. Bahdanau, D., Cho, K., & Bengio, Y. (2015). Neural Machine Translation by Jointly
Learning to Align and Translate. arXiv preprint. [Link]
3. Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., ... &
Amodei, D. (2020). Language Models are Few-Shot Learners. Advances in Neural
Information Processing Systems, 33, 1877-1901.

16
4. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers
of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM
Conference on Fairness, Accountability, and Transparency, 610–623.
[Link]
5. Castilho, S., Moorkens, J., Gaspari, F., & Way, A. (2017). The Human Factor in Machine
Translation. Translation Studies Quarterly, 10(2), 157–175.
6. Cho, K., Van Merriënboer, B., Gulcehre, C., Bougares, F., Schwenk, H., & Bengio, Y.
(2014). Learning Phrase Representations Using RNN Encoder-Decoder for Statistical
Machine Translation. arXiv preprint. [Link]
7. Denkowski, M., & Lavie, A. (2014). Meteor Universal: Language Specific Translation
Evaluation for Any Target Language. Proceedings of the Ninth Workshop on Statistical
Machine Translation, 376–380.
8. Gaspari, F., Toral, A., & Way, A. (2015). Challenges in Integrating Machine Translation
into Professional Translation Workflows. Translation Spaces, 4(2), 336–353.
9. Hutchins, W. J. (2005). The History of Machine Translation in a Nutshell. International
Journal of Translation, 17(1-2), 5–16.
10. Koehn, P., & Knowles, R. (2017). Six Challenges for Neural Machine Translation.
Proceedings of the First Workshop on Neural Machine Translation, 28–39.
11. Läubli, S., Toral, A., & Way, A. (2020). Assessing Post-Editing Effort for Neural Machine
Translation. Language Resources and Evaluation, 54(1), 139–165.
[Link]
12. Maggiolino, M. (2024). Antitrust concepts and artificial intelligence: The case of
plausibility. Stanford Computational Antitrust Journal. Retrieved from HeinOnline
13. Petersen, M., Alaa, A., Kıcıman, E., & Holmes, C. (2024). Artificial intelligence–based
copilots to generate causal evidence. NEJM AI. Retrieved from NEJM AI
14. Prince, E. W., Hankinson, T. C., & Görg, C. (2024). A visual analytics framework for
assessing interactive AI for clinical decision support. Proceedings of the World Scientific
Conference. Retrieved from World Scientific
15. Prates, M. O. R., Avelar, P. H., & Lamb, L. C. (2019). Assessing Gender Bias in Machine
Translation: A Case Study with Google Translate. Neural Computing and Applications,
32(1), 6363–6381.

17
16. Papineni, K., Roukos, S., Ward, T., & Zhu, W. J. (2002). BLEU: A Method for Automatic
Evaluation of Machine Translation. Proceedings of the 40th Annual Meeting of the
Association for Computational Linguistics, 311–318.
17. Toral, A., Way, A., & Läubli, S. (2018). Human vs. Neural Machine Translation:
Assessing Acceptability and Adequacy. Proceedings of the 15th Conference of the
European Chapter of the Association for Computational Linguistics (EACL), 68–74.
18. Vashee, K. (2020). Neural Machine Translation and the Future of Translation. The
Journal of Translation, 16(1), 1–15.

18