Pith. sign in

REVIEW 2 major objections 8 minor 24 references

AI in Support of Diversity and Inclusion

T0 review · 2 major / 8 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This white paper argues that AI systems can and should be built to support diversity and inclusion, not merely to avoid harm.

desk verdict A clear, honest white paper on AI for diversity and inclusion, but it is not a research contribution and should not be peer-reviewed as one. read the letter →

arxiv 2501.09534 v1 pith:6TZE6GLJ submitted 2025-01-16 cs.AI

Çiçek Güven , Afra Alishahi , Henry Brighton , Gonzalo Nápoles , Juan Sebastian Olier , Marie Šafář , Eric Postma , Dimitar Shterionov
show 2 more authors
Mirella De Sisto Eva Vanmassenhove
This is my paper · ORCID
classification cs.AI
keywords AIethicsdiversityandinclusionbiasmitigationtransparentlargelanguagemodelssigntranslationdisinformationmonitoringsociallyresponsible
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This white paper argues that AI systems can and should be built to support diversity and inclusion, not merely to avoid harm. It claims that transparency, bias-aware datasets, and collaboration with affected communities are the practical routes to that goal. The paper supports the argument by reviewing projects that detect and mitigate bias in structured data, monitor search-engine disinformation about LGBTQ+ communities, estimate child malnutrition from images, and translate between sign and spoken languages. If the argument holds, it gives researchers, policymakers, and funders a concrete agenda for socially responsible AI.

What carries the argument

The paper's argument is carried by four mechanisms: transparent algorithms that expose internal representations; fuzzy-rough sets, a mathematical framework for quantifying inconsistent patterns in data, used to measure explicit bias; fuzzy cognitive maps, a graph-based model of causal pathways among features, used to uncover implicit bias; and co-creation, a cyclic process that interleaves user feedback into development. These are joined by large-scale monitoring infrastructure for search-engine behavior and by deliberately diversified datasets, including a children's image dataset for malnutrition screening and parallel signed-spoken language corpora for translation evaluation.

What would settle it

An independent replication would settle the load-bearing examples: running the proposed bias-mitigation algorithm on a public benchmark and measuring implicit bias with the same fuzzy-cognitive-map method should reproduce a reduction near 76%, and an audit of the collected 1.5 million search results should confirm that major US search engines consistently rank legacy media above decentralized outlets for LGBTQ+-related queries. Failure of either check would remove the concrete evidence the advocacy depends on.

Watch

Extended reading notes

Core claim

The central claim is that AI's social impact is not fixed by its accuracy; the same technologies that risk reinforcing inequality can be redirected to challenge stereotypes and close communication gaps. The paper maintains that this requires treating fairness and transparency as design requirements rather than post-hoc corrections. Its evidence is a set of research projects: algorithms that quantify and mitigate explicit and implicit bias in classification data, with a reported reduction of up to 76% in simulated cases; a large-scale monitoring system that collected over 1.5 million search results and found major US search engines prioritizing legacy media over decentralized outlets in LGBTQ+ disinformation; a mobile-image system for detecting childhood malnutrition; and a sign-language translation framework built through co-creation with deaf and hard-of-hearing users.

Load-bearing premise

The advocacy rests on the accuracy of the project results it cites, especially the reported up-to-76% bias reduction and the search-engine ranking findings; these quantitative claims are asserted rather than demonstrated within the paper itself.

Editorial extensions

If this is right

  • AI development should treat bias mitigation and transparency as core design requirements, not post-hoc repairs.
  • Evaluation of AI systems should include inclusion metrics alongside accuracy, such as bias-quantification scores on protected features.
  • Large-scale automated monitoring of search engines and platforms can reveal structural biases in web infrastructure and inform regulation.
  • Diversified datasets and co-creation with affected communities are practical conditions for accessible technologies like sign-language translation.
  • Synthetic data should be used cautiously, since it can worsen bias against minority languages and underrepresented groups.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper implies a research agenda in which bias is an operational, measurable quantity; a natural next step would be standardized benchmarks that compare bias-mitigation algorithms across datasets.
  • If the search-engine finding replicates, it suggests that algorithmic neutrality is a property of specific infrastructures and that independent, non-US engines deserve regulatory attention.
  • The co-creation model used for sign-language translation could be tested as a general method: measuring whether co-created assistive technologies achieve higher trust and adoption than equivalent systems built without user feedback.
  • The 76% mitigation figure is reported from preliminary simulations; treating it as a stable property would require testing across more datasets and against competing mitigation methods.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 8 minor

Summary. This white paper from the Department of Cognitive Science and Artificial Intelligence at Tilburg University argues that AI can and should be developed to support diversity and inclusion when transparency, bias-aware data, and accessibility are central. It organizes the department's research into three themes—transparency of AI, identification and resolution of biases, and accessibility and empowerment—and illustrates each with concrete projects: interpretability work on large language models, gender-bias analysis in machine translation, fuzzy-rough-set and recurrent-network algorithms for bias mitigation, the Child Growth Monitor, the Search Guardian project on LGBTQ+ disinformation, the SignON sign-language translation project, and an AI-based media analysis of migrant portrayals. The paper closes by advocating for socially responsible AI and for fairness and inclusivity as priorities in AI development.

Significance. If evaluated as a position paper or white paper, the manuscript is a coherent overview of a research program and its ethical commitments. Its main strength is its explicitly conditional and self-reflective framing: it repeatedly cautions that technology should not be the preferred solution for every case and that inclusion should not be enforced merely for its own sake. The emphasis on co-creation and community collaboration, particularly in SignON and Search Guardian, is a valuable feature. The paper does not, however, provide original experimental evidence, and its quantitative claims are carried by citations or by single-sentence summaries that cannot be independently assessed from the text. The contribution is therefore programmatic rather than evidential; its significance will depend on the journal's acceptance of white-paper-style contributions.

major comments (2)
  1. [Section 3, 'Mitigating implicit and explicit biases'] The sentence 'Preliminary simulations reported that the algorithm is able to mitigate implicit bias in the data to a large extent (up to 76% in some cases) with minimum information loss' lacks the experimental setup, dataset, baselines, and definition of 'information loss' needed to assess it. The adjacent claim that the algorithm 'is unique in the field and solved a mathematical problem that had remained an open challenge for the scientific community' is similarly unsupported in this text. Because these statements are the concrete evidence for the paper's assertion that AI can mitigate bias, I request that the authors either provide the experimental details and error analysis, or rephrase the statements as project-specific outcomes with explicit pointers to the cited papers and a caveat that they have not been independently verified in this white paper.
  2. [Section 4, 'The dynamics of LGBTQ+ disinformation across Europe'] The monitoring study reports '178k search engine interactions' and '1.5 million search results' and concludes that major US search engines 'prioritize mass or legacy media outlets' while demoting peer-to-peer outlets. The paper does not give the query set, the list of search engines, the collection time window, or the statistical procedure behind the comparison, so the reader cannot distinguish a robust measurement from an anecdotal finding. I request a short methods paragraph or a reference to a publicly available project report, together with a statement of the caveats of the measurement approach.
minor comments (8)
  1. [Section 1 and Section 4] The manuscript contains several typographical errors: 'transparancy' in Section 1, 'courseshave' (missing space) in Section 4, and 'Dimitar Shre-tionov' in the Section 4 heading, which misspells Shterionov.
  2. [Section 4, 'The dynamics of LGBTQ+ disinformation across Europe'] The text says 'Figure 4 shows the location...' but the figure is labeled Figure 1; the cross-reference should be corrected.
  3. [Section 4, 'SignON project'] The project name is spelled inconsistently as 'Gost-Parc-Sign' in the text and 'GoSt-ParC-Sign' in reference [3]; please standardize the spelling.
  4. [Section 4, 'SignON project'] The sentence 'It was funded under the ELE (European Language Equality 2) project hat; then other SignON members contributed...' is garbled: 'hat' should be 'that' and the punctuation should be revised.
  5. [Throughout] The paper alternates between 'LGBTQ+' and 'LGBTI+' (e.g., Section 4 and reference [17]); choose one convention and define it at first use.
  6. [Section 4, 'SignON project'] The white paper 'Sign Language Technology: Do's and Don'ts...' by Rijckaert and Van Landuyt is cited in the text but not listed in the references; please add it.
  7. [Title and author list] The author names in the PDF contain corrupted combining characters (e.g., 'C ¸ i¸ cek G¨ uven'); the production system should render diacritics correctly.
  8. [Figure 2] Figure 2 (labeled 'Co-creation process') has no caption or explanatory description; please add one to relate it to the SignON discussion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity identified: the paper is an expository white paper whose advocacy claims are conditional and do not reduce to their cited supporting examples.

full rationale

This manuscript is a white paper that surveys research projects conducted by CSAI researchers and advocates, in a conditional and normative way, that AI systems can and should support diversity and inclusion when developed with transparency, bias-aware data, and accessibility in mind. The central claim is not derived from any formal model, equation, or fitted parameter, so there is no derivation chain whose conclusion is equivalent to its inputs. The paper cites its own prior work for specific project outcomes, such as the claim that an algorithm mitigates implicit bias 'up to 76% in some cases' (Section 3) and the Search Guardian finding based on 'over 1.5 million search results' (Section 4). These citations are self-citations, but they are presented as illustrative examples of ongoing research rather than as load-bearing premises that force the advocacy conclusion. Even if those quantitative results were overstated or non-reproducible, the paper's normative recommendation would remain intact as a research agenda, and the paper itself includes explicit caveats such as the SignON caution that 'technology should not be the preferred solution for every use case' and that 'inclusion should not be enforced merely for its own sake.' Because no step in the argument reduces by construction to a fitted input, a self-citation chain, or a definitional identity, the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

No free parameters or invented entities. The central claim depends on a domain assumption about AI bias and mitigation, and on the validity of cited project outcomes.

assumptions (1)
  • domain assumption AI systems reflect and can amplify historical biases present in training data; transparency and dataset diversification can mitigate these biases.
    The paper's advocacy depends on this causal chain, which is asserted in Sections 1 and 5 but not demonstrated in the text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AI in Support of Diversity and Inclusion." pith.science (2026). https://pith.science/paper/6TZE6GLJ

@misc{pith2026250109534,
  author       = {Pith},
  title        = {Pith review of: AI in Support of Diversity and Inclusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6TZE6GLJ}},
  note         = {Machine review of arXiv:2501.09534}
}
read the original abstract

In this paper, we elaborate on how AI can support diversity and inclusion and exemplify research projects conducted in that direction. We start by looking at the challenges and progress in making large language models (LLMs) more transparent, inclusive, and aware of social biases. Even though LLMs like ChatGPT have impressive abilities, they struggle to understand different cultural contexts and engage in meaningful, human like conversations. A key issue is that biases in language processing, especially in machine translation, can reinforce inequality. Tackling these biases requires a multidisciplinary approach to ensure AI promotes diversity, fairness, and inclusion. We also highlight AI's role in identifying biased content in media, which is important for improving representation. By detecting unequal portrayals of social groups, AI can help challenge stereotypes and create more inclusive technologies. Transparent AI algorithms, which clearly explain their decisions, are essential for building trust and reducing bias in AI systems. We also stress AI systems need diverse and inclusive training data. Projects like the Child Growth Monitor show how using a wide range of data can help address real world problems like malnutrition and poverty. We present a project that demonstrates how AI can be applied to monitor the role of search engines in spreading disinformation about the LGBTQ+ community. Moreover, we discuss the SignON project as an example of how technology can bridge communication gaps between hearing and deaf people, emphasizing the importance of collaboration and mutual trust in developing inclusive AI. Overall, with this paper, we advocate for AI systems that are not only effective but also socially responsible, promoting fair and inclusive interactions between humans and machines.

Figures

Figures reproduced from arXiv: 2501.09534 by the authors.

Figure 1
Figure 1. This map shows the location of residential internet connections we [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Co-creation process marginalised in the past. We refer the interested reader to the white paper “Sign Language Technology: Do’s and Don’ts, A Guide to Inclusive Collaboration Among Policymakers, Researchers, and End Users.” by Jorn Rijckaert and Davy Van Landuyt (under review). Through SignON it became clear that sign language data is not enough for the requirements of state-of-the-art neural models. Through additio… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages

  1. [1]

    School dropout pre- diction and feature importance exploration in malawi using household panel data: machine learning approach

    Hazal Colak Oz, C ¸ i¸ cek G¨ uven, and Gonzalo N´ apoles. School dropout pre- diction and feature importance exploration in malawi using household panel data: machine learning approach. Journal of Computational Social Science, 6(1):245–287, 2023

    work page 2023

  2. [2]

    XSL-HoReCo and GoSt-ParC-Sign: Two New Signed Language - Written Language Parallel Corpora

    Mirella De Sisto, Vincent Vandeghinste, Caro Brosens, Myriam Vermeer- bergen, and Dimitar Shterionov. XSL-HoReCo and GoSt-ParC-Sign: Two New Signed Language - Written Language Parallel Corpora. In Selected papers from the CLARIN Annual Conference 2023 , pages 23–33, 2024

    work page 2023

  3. [3]

    GoSt-ParC-Sign: Gold Standard Parallel Corpus of Sign and spoken language

    Mirella De Sisto, Vincent Vandeghinste, Lien Soetemans, Caro Brosens, and Dimitar Shterionov. GoSt-ParC-Sign: Gold Standard Parallel Corpus of Sign and spoken language. In Proceedings of the Annual Conference of the European Association for Machine Translation , pages 503–504, 2023

    work page 2023

  4. [4]

    Playing with gender

    Mar ´ ıa Isabel Rivas Ginel, Sarah Theroine,´Eric Poirier, Lo ¨ ıc Barrault, Felix Hoberg, Oliver Czulo, Eva Vanmassenhove, Izabella Thomas, and Lucie 12 Bernard. Playing with gender. all-inclusive games machine translation: The all-ingmt project. In Book of Abstracts, page 144, 2022

    work page 2022

  5. [5]

    Measuring implicit bias using shap feature importance and fuzzy cognitive maps

    Isel Grau, Gonzalo N´ apoles, Fabian Hoitsma, Lisa Koutsoviti Koumeri, and Koen Vanhoof. Measuring implicit bias using shap feature importance and fuzzy cognitive maps. In Intelligent Systems Conference , pages 745–764. Springer, 2023

    work page 2023

  6. [6]

    Mitigating implicit and explicit bias in structured data without sacrificing accuracy in pattern classification

    Fabian Hoitsma, Gonzalo N´ apoles, C ¸ i¸ cek G¨ uven, and Yamisleydi Salgueiro. Mitigating implicit and explicit bias in structured data without sacrificing accuracy in pattern classification. AI & SOCIETY , pages 1–20, 2024

    work page 2024

  7. [7]

    Bias quantification for pro- tected features in pattern classification problems

    Lisa Koutsoviti Koumeri and Gonzalo N´ apoles. Bias quantification for pro- tected features in pattern classification problems. In Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications: 25th Iberoamerican Congress, CIARP 2021, Porto, Portugal, May 10–13, 2021, Revised Selected Papers 25, pages 351–360. Springer, 2021

    work page 2021

  8. [8]

    Reduc- ing gender bias in nmt with fudge

    Tianshuai Lu, No¨ emi Aepli, Annette Rios, Eva Vanmassenhove, Beatrice Savoldi, Luisa Bentivogli, Joke Daems, and Jani¸ ca Hackenbuchner. Reduc- ing gender bias in nmt with fudge. InIn: 1st Workshop on Gender-Inclusive Translation Technologies (GITT). University of Zurich, 2023

    work page 2023

Show all 24 references
  1. [9]

    Predicting human body dimensions from single images: a first step in au- tomatic malnutrition detection

    Hezha MohammedKhan, Marleen Balvert, C ¸ i¸ cek G¨ uven, and Eric Postma. Predicting human body dimensions from single images: a first step in au- tomatic malnutrition detection. In CAIP 2021: Proceedings of the 1st In- ternational Conference on AI for People: Towards Sustaina...

    2021

  2. [10]

    Image-based body shape estimation to detect malnutrition

    Hezha MohammedKhan, C ¸ i¸ cek G¨ uven, Marleen Balvert, and Eric Postma. Image-based body shape estimation to detect malnutrition. In Proceedings of SAI Intelligent Systems Conference , pages 577–590. Springer, 2023

    2023

  3. [11]

    Aran: Age-restricted anonymized dataset of children images and body measure- ments, 2024

    Hezha MohammedKhan, Cascha van Wanrooij, Eric Postma, C ¸ i¸ cek G¨ uven, Marleen Balvert, Chenar Omer Ali AL JAF, and Heersh Raof Saeed. Aran: Age-restricted anonymized dataset of children images and body measure- ments, 2024. submitted

    2024

  4. [12]

    Modeling implicit bias with fuzzy cog- nitive maps

    Gonzalo N´ apoles, Isel Grau, Leonardo Concepci´ on, Lisa Koutsoviti Koumeri, and Jo˜ ao Paulo Papa. Modeling implicit bias with fuzzy cog- nitive maps. arXiv preprint arXiv:2112.12713 , 2021

    2021 arXiv

  5. [13]

    A fuzzy-rough uncertainty measure to discover bias encoded explicitly or implicitly in features of struc- tured pattern classification datasets

    Gonzalo N´ apoles and Lisa Koutsoviti Koumeri. A fuzzy-rough uncertainty measure to discover bias encoded explicitly or implicitly in features of struc- tured pattern classification datasets. Pattern Recognition Letters, 154:29– 36, 2022. 13

    2022

  6. [14]

    Stereotypes, dispropor- tions, and power asymmetries in the visual portrayal of migrants in ten countries: an interdisciplinary ai-based approach

    Juan Sebastian Olier and Camilla Spadavecchia. Stereotypes, dispropor- tions, and power asymmetries in the visual portrayal of migrants in ten countries: an interdisciplinary ai-based approach. Humanities and Social Sciences Communications, 9(1):1–16, 2022

    2022

  7. [15]

    How we’ve taught algorithms to see identity: Constructing race and gender in image databases for facial analysis

    Morgan Klaus Scheuerman, Kandrea Wade, Caitlin Lustig, and Jed R Brubaker. How we’ve taught algorithms to see identity: Constructing race and gender in image databases for facial analysis. Proceedings of the ACM on Human-computer Interaction , 4(CSCW1):1–35, 2020

    2020

  8. [16]

    The signon project: a sign language translation framework

    Dimitar Shterionov, Vincent Vandeghinste, Horacio Saggion, Josep Blat, Mathieu De Coster, Joni Dambre, Henk Van den Heuvel, Irene Murtagh, Lorraine Leeson, and Ineke Schuurman. The signon project: a sign language translation framework. In 31st Meeting of Computational Linguist...

    2021

  9. [17]

    Disinformation campaigns about LGBTI+ people in the EU and foreign influence

    Cecilia Strand, Jakob Svensson, Roland Blomeyer, and Margarita Sanz. Disinformation campaigns about LGBTI+ people in the EU and foreign influence. European Parliament, Policy Department for External Relations, 2021

    2021

  10. [18]

    From impoverished to enriched brains: A lifespan perspective on the link between poverty, nutrition, cognitive functioning and underlying brain networks

    Tilburg University. From impoverished to enriched brains: A lifespan perspective on the link between poverty, nutrition, cognitive functioning and underlying brain networks. https: //www.tilburguniversity.edu/about/digital-sciences-society/ projects/impoverished-enriched-brain...

    2024

  11. [19]

    An interdisciplinary analysis of gender- based discrimination in translation technology

    Tilburg University. An interdisciplinary analysis of gender- based discrimination in translation technology. https://www. tilburguniversity.edu/about/digital-sciences-society/ projects/gender-based-discrimination-translation-technology ,

  12. [20]

    Report on europe’s sign languages

    Vincent Vandeghinste, Mirella De Sisto, Maria Kopf, Marc Schulder, Caro Brosens, and Lien Soetemans. Report on europe’s sign languages. ele project deliverable 1.40. Report on Europe’s Sign Languages , 2023

    2023

  13. [21]

    Gender bias in machine translation and the era of large language models

    Eva Vanmassenhove. Gender bias in machine translation and the era of large language models. arXiv preprint arXiv:2401.10016 , 2024

    2024 arXiv

  14. [22]

    Neutral rewriter: A rule-based and neural approach to automatic rewriting into gender-neutral alternatives

    Eva Vanmassenhove, Chris Emmery, and Dimitar Shterionov. Neutral rewriter: A rule-based and neural approach to automatic rewriting into gender-neutral alternatives. arXiv preprint arXiv:2109.06105 , 2021

    2021 arXiv

  15. [23]

    Proceedings of the 1st workshop on gender-inclusive translation technologies

    Eva Vanmassenhove, Beatrice Savoldi, Luisa Bentivogli, Joke Daems, and Jani¸ ca Hackenbuchner. Proceedings of the 1st workshop on gender-inclusive translation technologies. In First Workshop on Gender-Inclusive Trans- lation Technologies at the European Association for Machine...

    2023

  16. [2024]

    online, accessed at 24 04 2024

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.