Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Regulating Reality: Exploring Synthetic Media Through Multistakeholder AI Governance

T0 review · 4 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper argues that governing synthetic media is most productive when framed as a problem of misrepresentation and impersonation, not of AI technology alone, and that trust and time-awareness are the load-bearing governance resources.

desk verdict A candid insider qualitative study with real interview data and a genuinely useful temporal-perspective theme, but the design's circularity keeps the findings closer to a well-documented expert consensus than a field-wide result. read the letter →

arxiv 2502.04526 v1 pith:3AITA4YG submitted 2025-02-06 cs.CY

Claire R. Leibowicz This is my paper · ORCID
classification cs.CY
keywords syntheticmediaAIgovernancemultistakeholdertransparencydeepfakesmisrepresentationtrustqualitativeinterviews
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

The paper studies how the people who actually govern synthetic media—civil society, industry, media, and policy actors—understand the problem and choose solutions. Based on 23 in-depth interviews and two real-world cases of multistakeholder governance, it argues that the core challenge is not the technology itself but misrepresentation and impersonation. It also finds that stakeholders' sense of time—past, present, and future—shapes their decisions, that trust among collaborators is a prerequisite for effective rulemaking, and that technical fixes like AI labels are necessary but limited. The stakes are practical: if the paper is right, governance efforts that center on "was this made by AI?" will keep missing the point, and policy should instead supply context, build trust, and adapt over time.

What carries the argument

The analytical machinery is a qualitative, inductive thematic analysis of 23 semi-structured interviews, anchored in two real-world governance case studies in which the author participated. Three coding passes—attribute coding, descriptive coding, and higher-level thematic coding—yielded the organizing concepts that carry the argument: a three-stage model of governance (understanding the problem, developing norms, implementing solutions), a dual-use account of synthetic media, the temporal perspective spanning past, present, and future as a mediating frame, and trust as both an interpersonal precondition and an audience-facing property of transparency interventions. The temporal lens does the heaviest interpretive work: it reconciles the seemingly contradictory view that synthetic media is old news with the view that it demands urgent, novel governance.

What would settle it

A survey experiment in which audiences rate disclosures from trusted versus unfamiliar issuers could settle the paper's claim that trust in the disclosing authority is a prerequisite for transparency to work.

Watch

Extended reading notes

Core claim

The paper's central finding is that the synthetic media challenge, as the stakeholders who govern it describe it, is a problem of misrepresentation and impersonation rather than a problem of AI-generated content as such. That framing carries three corollaries. First, technical transparency measures such as provenance metadata and "AI or not" labels are worth pursuing but cannot carry the governance burden alone, because knowing whether a machine touched an image says nothing about whether the content misleadingly represents someone or something. Second, trust does double work: audiences must trust the disclosures they are given, and the diverse institutions writing those disclosures must trust one another enough to move past corporate positions. Third, time matters: stakeholders frame synthetic media as an evolution of an old problem, yet feel a new urgency because harms appear faster than beneficial uses, and their governance recommendations emphasize iterative, adaptable rules.

Load-bearing premise

The argument's load-bearing premise is that the 23 people interviewed, mostly from North America and Europe and many from the author's own professional network and sympathetic to multistakeholder governance, are representative enough to reveal the field's shared understanding of synthetic media.

Editorial extensions

If this is right

  • Governance metrics should track reductions in misrepresentation and impersonation, not just whether content carries an AI label.
  • Policymakers should pair technical transparency with media literacy and contextual disclosures such as source, editorial changes, and process, rather than stopping at a binary "AI or not" marker.
  • Voluntary multistakeholder guidelines can catalyze alignment and inform regulation, but they need eventual enforcement to be effective.
  • Trust-building among stakeholders should be treated as a first-order governance activity, since it enables honest collaboration on hard problems.
  • Rules should be designed to be revisited frequently, because stakeholders expect both the technology and public understanding of it to change quickly.

Reading between the lines

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

  • If temporal framing does the interpretive work the paper claims, a comparative study of earlier visual-manipulation governance moments, such as Photoshop-era photo ethics, should find the same past-present-future pattern; that would test the mechanism outside synthetic media.
  • The paper's trust finding suggests a testable extension: audience trust in a transparency label should vary with the credibility of the issuing institution, so a label from a platform with a commercial interest may need to be relayed through a third party to work.
  • Because the sample skews Western and toward participants already sympathetic to multistakeholder governance, the author's implicit claim that misrepresentation is the unifying problem may not hold for non-Western regulatory contexts, where state surveillance, labor displacement, or cultural production harms could dominate the agenda.
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

4 major / 4 minor

Summary. This paper reports a qualitative study of multistakeholder synthetic media governance. The author draws on autoethnographic review of two Partnership on AI (PAI) cases (the Deepfake Detection Challenge and the Synthetic Media Framework), 13 interviews with participants in those cases, and 10 additional policymaker interviews from Europe and North America. Using semi-structured interviews and inductive coding, the paper identifies challenges (misrepresentation/impersonation, dual use, temporal perspectives), solutions (technical transparency with limits, moving beyond an 'AI or not' binary), and implementation factors (voluntary guidelines, trust, iterative governance). It claims these themes inform evidence-based policy design and contribute to the multistakeholder AI governance literature.

Significance. If the findings are robust, the paper provides rare first-hand empirical insight into how cross-sector actors actually collaborate on synthetic media governance, and it usefully situates the policy problem as misrepresentation rather than technology per se. The study has notable strengths: the full interview protocol is included in Appendix A; the methods and positionality are described in detail; the research received ethics approval; and Section 6 candidly acknowledges the Western sample and participants' pre-existing inclination toward multistakeholder approaches. The direct quotes give the analysis texture. However, due to the sample composition and the protocol's derivation from the same cases that recruited participants, the findings are best read as an exploratory insider case study rather than a representative field consensus.

major comments (4)
  1. [Section 3.3 and Appendix A] The interview questions were explicitly 'drawn from themes that emerged during autoethnographic reflection on the PAI cases' (Section 3.3), and 13 of the 23 interviewees were participants in those same cases (Section 3.2). The protocol in Appendix A asks directly about dual use (Q7), transparency and disclosure (Q8), and multistakeholder collaboration (Q4/Q5), so the appearance of these themes in Sections 4.2.2 and 4.3 is partly by construction. Section 6 concedes that case participants were inclined toward multistakeholder approaches. This self-confirming design is load-bearing because the paper's central claim—that synthetic media should be governed as a problem of misrepresentation and impersonation with temporal and trust dimensions—rests on themes that the instrument itself seeded. I ask the author to address this by (a) reporting the codebook and any negative-case analysis, (b) using an independent second coder, or (c) reframing the contribution as an exploratory case study of insider perspectives rather than a general field consensus.
  2. [Section 3.2, Table 1, Section 6] The sample is 23 individuals, of whom 13 came from the two PAI cases and 10 were policymakers recruited by snowballing from Europe (4) and North America (6). Sectoral composition is skewed: 7 industry, 3 civil society, 3 media, and 10 policy. There are no participants from the global South and no civil society participants outside the PAI orbit. Section 6 acknowledges these limits, but the Discussion (Section 5) nevertheless states that the research 'highlights synthetic media as a problem of misrepresentation and impersonation' without the caveat that this reflects a narrow group of stakeholders. The recruitment strategy therefore supports claims about this particular multistakeholder community, not about synthetic media governance broadly. Please either expand the sample or systematically qualify the general claims in Sections 4 and 5.
  3. [Section 4.2.1] The assertion that 'There was consensus among participants' on misrepresentation and impersonation as the central challenge is not backed by systematic evidence. The paragraph provides two illustrative quotes (PAI9 and PAI2), but no count of how many of the 23 participants expressed this view, how the code 'misrepresentation' was defined, or which participants dissented. Because this is the paper's headline finding, the analysis should quantify theme prevalence and show deviant cases. Otherwise the 'consensus' claim is unverifiable and could reflect selective quotation.
  4. [Section 4.2.2] The dual-use tally is internally inconsistent: the text reports '10 stakeholders... said it was net negative, four stakeholders (one media, two industry, two policy) said it was net positive, and eight either said it was too complicated'—the net-positive breakdown sums to five, and the three categories sum to 22 rather than 23. Please correct the counts and clarify whether one participant's response was uncodeable or missing. This is a reporting error in a results section that otherwise relies on counts sparingly.
minor comments (4)
  1. [Section 3.4] The coding process would be easier to evaluate with a codebook including code definitions and example segments; the current description names attribute coding and inductive coding but does not specify how themes were counted or how code saturation was assessed.
  2. [Table 1] The table header and cells could be clearer: list the exact number of participants per sector in the header or a separate column, and ensure the participant IDs (PAI1, PAI 9, etc.) are formatted consistently.
  3. [Section 4.2.3] The statement that 'More than half of participants' referenced temporal perspectives should be accompanied by an exact count and a few representative quotes, matching the level of detail used elsewhere in the findings.
  4. [Reference list, [29]] Reference [29] lists the author as 'Claire R .' with an incomplete surname; this should be corrected.

Circularity Check

2 steps flagged · score 4.0 of 10

Interview protocol reproduces the author's own PAI case themes, but the central misrepresentation, temporal, and trust findings retain independent content.

  1. self definitional [Section 3.3 (Interview Protocol) and Appendix A.1]
    ""All interview questions (see Appendix A) were drawn from themes that emerged during autoethnographic reflection on the PAI cases." Appendix A then asks: "7 — Dual Use: How do you understand the balance of synthetic media’s positive and harmful impacts? ... 8 — Transparency and Disclosure: What do the example cases of PAI governance reveal about open questions and opportunities related to disclosing and relaying that content has been synthesized for audiences?""

    The findings presented as inductive discoveries in Sections 4.2.2 and 4.3 — dual use and transparency/disclosure — are not emergent from unstructured conversation; the protocol explicitly solicits them, and Section 3.3 says those questions were themselves drawn from the author's autoethnographic reflection on the same PAI cases being studied. The instrument therefore encodes the expected themes into the data-collection step, and the findings re-export those themes as if they were bottom-up findings. The dual-use question makes it near-certain that 'dual use' will appear as a theme, and the autoethnographic theme of the 'value (and difficulty) of supporting transparency and disclosure' makes 'transparency limitations' similarly pre-loaded.

  2. fitted input called prediction [Section 4.1 (High-Level Themes from PAI Cases) and Abstract/Section 3.4]
    ""I worked firsthand on the PAI cases and began by reviewing them and selecting key themes from existing material. Doing so thematically orients this research and supports testing such assumptions with stakeholders through interviews. Key themes included synthetic media’s dual use, the need for sociotechnical interventions on content, the simultaneous value (and difficulty) of supporting transparency and disclosure about content, and the ways in which different actors in the synthetic media pipeline have roles to play in mitigating the harms from synthetic media and optimizing benefits.""

    The Abstract and Section 3.4 advertise 'Inductive coding reveals key themes' and 'formal inductive coding of the interview transcripts,' implying themes emerged from the data. Section 4.1, however, shows that the 'key themes' were selected in advance from the author's own PAI case material, used to write the protocol, and then tested on interviewees recruited from those same cases (Section 3.2). The findings section then reports dual use, technical-transparency limits, and stakeholder-role themes that match the pre-selected themes. The pre-selection is the fit; the interview responses are the prediction that largely returns the fit.

full rationale

This is a qualitative interview study rather than a formal derivation, so the circularity is methodological rather than equation-level. The clearest circular step is the protocol/finding loop: Section 3.3 says all interview questions were drawn from the author's autoethnographic reflection on the same PAI cases being studied, and Appendix A shows explicit prompts on dual use and transparency. The findings then report dual use and transparency limits as key inductive themes (Sections 4.2.2, 4.3). Those two themes are partly guaranteed by the instrument, so they are not independent discoveries. Section 4.1 confirms the pre-selection: key themes were 'selected' from existing PAI case material before interviews. The sampling amplifies this: 13 of 23 interviewees participated in the same PAI cases, and the author led both cases and has professional relationships with many interviewees, as the Positionality Statement acknowledges. The paper does disclose this, and Section 6 concedes the participants' inclination toward multistakeholder approaches, which is good practice. The central claims about misrepresentation, temporal perspectives, and trust, however, were not explicit in the pre-selected theme list or the protocol, and they appear to come from the interviewees' open-ended answers and the author's coding; they therefore retain independent content. Self-citation of the author's PAI reports is present but is secondary to the empirical material and disclosed. Overall, this is partial self-definitional circularity in the theme-generation loop, not a total reduction of the main thesis to its inputs, so the score is 4 rather than 6 or higher.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper's central claims rest on qualitative research assumptions: the definitions chosen, the representative of the sample, the reliability of self-report, and the sufficiency of the coding. These are domain assumptions that cannot be independently verified from the manuscript alone.

assumptions (5)
  • domain assumption Synthetic media is defined as 'images, videos, and audio clips, that have been significantly altered or generated by algorithms, including by AI' (adapted from US Executive Order).
    This definition shapes which governance efforts are included in the analysis and is adopted, not derived, in Section 1.
  • domain assumption The ICANN definition of a multistakeholder process is accepted as the analytical frame.
    Section 1.2 borrows ICANN's definition, which determines what counts as multistakeholder governance in the case selection.
  • domain assumption Interviewees' self-reported perspectives are treated as accurate reflections of their views and of the governance processes.
    The analysis relies entirely on semi-structured interviews, with no independent observation or document verification (Section 3).
  • domain assumption The two PAI cases (Deepfake Detection Challenge and Synthetic Media Framework) are representative of multistakeholder synthetic media governance.
    The paper selects cases from PAI, where the author is an employee, and generalizes from them (Section 3.1).
  • domain assumption Inductive coding of 23 interviews is sufficient to surface themes generalizable to the broader field.
    The paper's claims about 'what stakeholders reveal' rest on the assumption that the sample saturation and coding quality support generalization (Section 3.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Regulating Reality: Exploring Synthetic Media Through Multistakeholder AI Governance." pith.science (2026). https://pith.science/paper/3AITA4YG

@misc{pith2026250204526,
  author       = {Pith},
  title        = {Pith review of: Regulating Reality: Exploring Synthetic Media Through Multistakeholder AI Governance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3AITA4YG}},
  note         = {Machine review of arXiv:2502.04526}
}
read the original abstract

Artificial intelligence's integration into daily life has brought with it a reckoning on the role such technology plays in society and the varied stakeholders who should shape its governance. This is particularly relevant for the governance of AI-generated media, or synthetic media, an emergent visual technology that impacts how people interpret online content and perceive media as records of reality. Studying the stakeholders affecting synthetic media governance is vital to assessing safeguards that help audiences make sense of content in the AI age; yet there is little qualitative research about how key actors from civil society, industry, media, and policy collaborate to conceptualize, develop, and implement such practices. This paper addresses this gap by analyzing 23 in-depth, semi-structured interviews with stakeholders governing synthetic media from across sectors alongside two real-world cases of multistakeholder synthetic media governance. Inductive coding reveals key themes affecting synthetic media governance, including how temporal perspectives-spanning past, present, and future-mediate stakeholder decision-making and rulemaking on synthetic media. Analysis also reveals the critical role of trust, both among stakeholders and between audiences and interventions, as well as the limitations of technical transparency measures like AI labels for supporting effective synthetic media governance. These findings not only inform the evidence-based design of synthetic media policy that serves audiences encountering content, but they also contribute to the literature on multistakeholder AI governance overall through rare insight into real world examples of such processes.

Figures

Figures reproduced from arXiv: 2502.04526 by the authors.

Figure 1
Figure 1. A diagram of the organizational logic of analysis and discussion, made with graphics from Canva. 4.2 How Stakeholders Understand the Synthetic Media Challenge The descriptive and inductive coding processes revealed many shared perspectives on how stakeholders understand synthetic media’s challenges. Interviewees emphasized myriad policy questions encapsulated in synthetic media governance and those that unify them, … view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. TRIED: Truly Innovative and Effective AI Detection Benchmark, developed by WITNESS

    cs.CY 2025-04 conditional novelty 5.0 of 10

    The TRIED Benchmark is a 177-point, six-pillar checklist for rating AI detection tools on sociotechnical criteria, but its scoring thresholds are not empirically validated.

Reference graph

Works this paper leans on

53 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [1]

    unethical to deploy deepfake detectors,

    A media stakeholder talked about how their takeaway from Case 1 was that it would be “unethical to deploy deepfake detectors,” alluding to the adversarial dynamics that make detectors susceptible to evasion [32]. And yet, despite these limitations, the stakeholders underscored their hope that such technical interventions can be improved through sociotechn...

    work page 2020

  2. [2]

    an important first step to get convergence…amongst ecosystem actors

    creating norms and guidelines for how to solve to those challenges. Here, I explore the third stage: how stakeholders implement those norms and solutions. Pursuing different governance instruments can affect how solutions are enacted, and interviews about implementation shed light on synthetic media governance overall. 4.4.1 Nongovernmental, voluntary gui...

    work page 2022

  3. [3]

    In the sections that follow, I motivate the use of PAI cases as the foundation for this research before describing study participants, interview questions, and analysis methods

    interviews with 10 synthetic media government policymakers who did not participate in the cases. In the sections that follow, I motivate the use of PAI cases as the foundation for this research before describing study participants, interview questions, and analysis methods. 3.1 Positioning the Partnership on AI (PAI) as a Research Foundation PAI is a usef...

    work page 2016

  4. [6]

    existing problem

    the importance of trust—both between audiences and solutions, and among diverse stakeholders—in supporting governance effectiveness. 5.1.1Temporal Perspectives Affect Synthetic Media Governance The role of temporal perspectives in stakeholders’ understanding of synthetic media and their ability to influence its governance emerged throughout the interviews...

    work page 2014

  5. [8]

    The Journal of Politics (September 2024)

    Political Deepfakes are as Credible as Other Fake Media and (Sometimes) Real Media. The Journal of Politics (September 2024). https://doi.org/10.1086/732990

    work page doi:10.1086/732990 2024

  6. [10]

    Partnership on AI Report

    How the BBC used face swapping to anonymize interviewees. Partnership on AI Report. Retrieved from https://partnershiponai.org/wp-content/uploads/2024/03/pai-synthetic-media-case-study-bbc.pdf

    work page 2024

  7. [12]

    NIST (November 2024)

    Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency. NIST (November 2024). Retrieved January 18, 2025 from https://www.nist.gov/publications/reducing-risks-posed-synthetic-content-overview-technical-approaches-digital-content

    work page 2024

  8. [13]

    Research-Technology 17 Management 63, 3 (May 2020), 42—48

    Competitor Collaboration Before a Crisis: What the AI Industry Can Learn. Research-Technology 17 Management 63, 3 (May 2020), 42—48. https://doi.org/10.1080/08956308.2020.1733889

    work page arXiv 2020

Show all 53 references
  1. [15]

    governance modalities and governance functions

    suggested that “multistakeholder participation is touted as a mechanism for building normative consensus on the governance of AI technologies and being a policy vehicle for operationalizing such principles.” This sentiment is echoed by many institutions in the AI space—ranging...

    2024

  2. [16]

    Retrieved August 23, 2022 from http://arxiv.org/abs/2006.07397

    The DeepFake Detection Challenge (DFDC) Dataset. Retrieved August 23, 2022 from http://arxiv.org/abs/2006.07397

    2022 arXiv

  3. [17]

    Science 380, 6650 (June 2023), 1110—1111

    Art and the science of generative AI. Science 380, 6650 (June 2023), 1110—1111. https://doi.org/10.1126/science.adh4451

    2023 doi

  4. [18]

    SSRN Journal (2020)

    Principled Artificial Intelligence: Mapping Consensus in Ethical and Rights-Based Approaches to Principles for AI. SSRN Journal (2020). https://doi.org/10.2139/ssrn.3518482

    2020 doi

  5. [19]

    The New York Times

    Voice Deepfakes Are Coming for Your Bank Balance. The New York Times. Retrieved January 14, 2024 from https://www.nytimes.com/2023/08/30/business/voice-deepfakes-bank-scams.html

    2024

  6. [20]

    jdsr 2, 3 (November 2020)

    Coproduction, Ethics and Artificial Intelligence: A Perspective from Cultural Anthropology. jdsr 2, 3 (November 2020). https://doi.org/10.33621/jdsr.v2i3.53

    2020 doi

  7. [23]

    The New York Times

    This Tool Could Protect Artists From A.I.-Generated Art That Steals Their Style. The New York Times. Retrieved March 25, 2023 from https://www.nytimes.com/2023/02/13/technology/ai-art-generator-lensa-stable-diffusion.html

    2023

  8. [24]

    What’s changed? MIT Technology Review

    AI companies promised to self-regulate one year ago. What’s changed? MIT Technology Review. Retrieved January 22, 2025 fromhttps://www.technologyreview.com/2024/07/22/1095193/ai-companies-promised-the-white-house-to-self-regulate-one-year-ago-whats-changed/

    2025

  9. [25]

    Retrieved January 13, 2024 from https://www.technologyreview.com/2023/02/27/1069166/how-to-create-release-and-share-generative-ai-responsibly/

    How to create, release and share generative AI responsibly | MIT Technology Review. Retrieved January 13, 2024 from https://www.technologyreview.com/2023/02/27/1069166/how-to-create-release-and-share-generative-ai-responsibly/

    2024

  10. [26]

    extend beyond [their] typical wires-and-switches connotation to [include] any innovation that has extended, structured, or transformed visual perception and communication

    suggest, visual technologies “extend beyond [their] typical wires-and-switches connotation to [include] any innovation that has extended, structured, or transformed visual perception and communication.” Synthetic media, also known as generative or AI-generated media, is one su...

    2023

  11. [27]

    APSC 53, 4 (October 2020), 703—709

    A New Model for Industry—Academic Partnerships. APSC 53, 4 (October 2020), 703—709. https://doi.org/10.1017/S1049096519001021

    2020 doi

  12. [28]

    PAI Publication

    The Deepfake Detection Challenge: Insights and Recommendations for AI and Media Integrity. PAI Publication. Retrieved from https://partnershiponai.org/wp-content/uploads/2021/07/671004_Format-Report-for-PDF_031120-1.pdf

    2021

  13. [29]

    https://doi.org/10.48550/arXiv.2407.13025

    From Principles to Practices: Lessons Learned from Applying Partnership on AI’s (PAI) Synthetic Media Framework to 11 Use Cases. https://doi.org/10.48550/arXiv.2407.13025

    doi arXiv

  14. [30]

    In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (AIES ’21), July 30,

    The Deepfake Detection Dilemma: A Multistakeholder Exploration of Adversarial Dynamics in Synthetic Media. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (AIES ’21), July 30,

    2021

  15. [31]

    https://doi.org/10.1145/3461702.3462584

    Association for Computing Machinery, New York, NY, USA, 736—744. https://doi.org/10.1145/3461702.3462584

  16. [32]

    Tech Policy Press

    A Call for Modular Multistakeholder AI Governance: Practical Recommendations for the Upcoming AI Action Summit | TechPolicy.Press. Tech Policy Press. Retrieved January 18, 2025 from https://techpolicy.press/a-call-for-modular-multistakeholder-ai-governance-practical-recommenda...

    2025

  17. [33]

    Center for Data Innovation

    The AI Act’s AI Watermarking Requirement Is a Misstep in the Quest for Transparency. Center for Data Innovation. Retrieved January 22, 2025 from https://datainnovation.org/2024/07/the-ai-acts-ai-watermarking-requirement-is-a-misstep-in-the-quest-for-transparency/

    2025

  18. [36]

    Retrieved from https://partnershiponai.org/annual-report-2023/

    Meeting the Moment: 2023 Annual Report. Retrieved from https://partnershiponai.org/annual-report-2023/

    2023

  19. [37]

    National Academies of Sciences, Engineering, and Medicine. (2022). Fostering Responsible Computing Research: Foundations and Practices [Consensus Report]. Washington, DC: The National Academies Press. https://doi.org/10.17226/26507. 18

    2022 doi

  20. [38]

    Where Responsible AI meets Reality: Practitioner Perspectives on Enablers for Shifting Organizational Practices. Proc. ACM Hum.-Comput. Interact. 5, CSCW1 (April 2021), 7:1-7:23. https://doi.org/10.1145/3449081

    2021 doi

  21. [39]

    The Verge

    OpenAI’s DALL-E AI image generator can now edit pictures, too. The Verge. Retrieved January 14, 2024 from https://www.theverge.com/2022/4/6/23012123/openai-clip-dalle-2-ai-text-to-image-generator-testing

    2024

  22. [40]

    Twitter Blog

    Building rules in public: Our approach to synthetic and manipulated media. Twitter Blog. Retrieved from https://blog.twitter.com/en_us/topics/company/2020/new-approach-to-synthetic-and-manipulated -media

    2020

  23. [41]

    The New York Times

    Deepfake Technology Enters the Documentary World. The New York Times. Retrieved January 13, 2024 from https://www.nytimes.com/2020/07/01/movies/deepfakes-documentary-welcome-to-chechnya.html

    2024

  24. [42]

    MIT Technology Review

    Cryptography may offer a solution to the massive AI-labeling problem. MIT Technology Review. Retrieved January 13, 2024 from https://www.technologyreview.com/2023/07/28/1076843/cryptography-ai-labeling-problem-c2pa-provenance/

    2024

  25. [43]

    long-term AI risks

    Resolving the battle of short- vs. long-term AI risks. AI Ethics (September 2023). https://doi.org/10.1007/s43681-023-00336-y

    2023 doi

  26. [44]

    Retrieved January 22, 2025 from https://misinforeview.hks.harvard.edu/article/misinformation-interventions-are-common-divisive-and-poorly-understood/

    Misinformation interventions are common, divisive, and poorly understood | HKS Misinformation Review. Retrieved January 22, 2025 from https://misinforeview.hks.harvard.edu/article/misinformation-interventions-are-common-divisive-and-poorly-understood/

    2025

  27. [46]

    In Proceedings of the Computational Humanities Research Conference, December 4 — 6, 2024, Aaarhus, Denmark

    Bringing Rome to Life: Evaluating Historical Image Generation. In Proceedings of the Computational Humanities Research Conference, December 4 — 6, 2024, Aaarhus, Denmark. 113-126

    2024

  28. [47]

    Generative ML and CSAM: Implications and Mitigations. (2023). https://doi.org/10.25740/jv206yg3793

    2023 doi

  29. [48]

    IEEE Journal of Selected Topics in Signal Processing 14, 5 (August 2020), 910—932

    Media Forensics and DeepFakes: An Overview. IEEE Journal of Selected Topics in Signal Processing 14, 5 (August 2020), 910—932. https://doi.org/10.1109/JSTSP.2020.3002101

    2020

  30. [49]

    The Verge

    Deepfake detection algorithms will never be enough. The Verge. Retrieved January 14, 2024 from https://www.theverge.com/2019/6/27/18715235/deepfake-detection-ai-algorithms-accuracy-will-they-ever-work

    2024

  31. [50]

    Social media reconsiders its relationship with the truth. Axios. Retrieved August 20, 2022 from https://www.axios.com/2019/08/21/social-media-reconsiders-relationship-truth

    2022

  32. [51]

    Retrieved January 22, 2025 from https://www.brennancenter.org/our-work/research-reports/regulating-ai-deepfakes-and-synthetic-media-political-arena

    Regulating AI Deepfakes and Synthetic Media in the Political Arena | Brennan Center for Justice. Retrieved January 22, 2025 from https://www.brennancenter.org/our-work/research-reports/regulating-ai-deepfakes-and-synthetic-media-political-arena

    2025

  33. [52]

    Tech Policy Press

    Exploring Legal Approaches to Regulating Nonconsensual Deepfake Pornography | TechPolicy.Press. Tech Policy Press. Retrieved January 22, 2025 from https://techpolicy.press/exploring-legal-approaches-to-regulating-nonconsensual-deepfake-pornography

    2025

  34. [53]

    An MIT Exploration of Generative AI (March 2024)

    Labeling AI-Generated Content: Promises, Perils, and Future Directions. An MIT Exploration of Generative AI (March 2024). https://doi.org/10.21428/e4baedd9.0319e3a6

    2024 doi

  35. [2007]

    SAGE Publications, Inc

    The SAGE Dictionary of Qualitative Inquiry. SAGE Publications, Inc. https://doi.org/10.4135/9781412986281

    doi

  36. [2010]

    Qualitative Report 13, (January 2010)

    Qualitative Case Study Methodology: Study Design and Implementation for Novice Researchers. Qualitative Report 13, (January 2010). https://doi.org/10.46743/2160-3715/2008.1573

    2010

  37. [2015]

    Multistakeholderism: anatomy of an inchoate global institution. Int. Theory 7, 3 (November 2015), 572—616. https://doi.org/10.1017/S1752971915000081

    2015 doi

  38. [2016]

    The New York Times

    Protecting Humans and Jobs From Robots Is 5 Tech Giants’ Goal. The New York Times. Retrieved August 14, 2022 from https://www.nytimes.com/2016/09/29/technology/protecting-humans-and-jobs-from-robots-is-5-tech-giants-goal.html

    2022

  39. [2019]

    Nat Mach Intell 1, 1 (January 2019), 5—6

    Bridging near- and long-term concerns about AI. Nat Mach Intell 1, 1 (January 2019), 5—6. https://doi.org/10.1038/s42256-018-0003-2

    2019 doi

  40. [2020]

    Meta Blog

    Enforcing Against Manipulated Media. Meta Blog. Retrieved from https://about.fb.com/news/2020/01/enforcing-against-manipulated-media/

    2020

  41. [2021]

    MIT Technology Review

    This avocado armchair could be the future of AI. MIT Technology Review. Retrieved January 21, 2025 from https://www.technologyreview.com/2021/01/05/1015754/avocado-armchair-future-ai-openai-deep-learning-nlp-gpt3-computer-vision-common-sense/

    2025

  42. [2022]

    Journalism 23, 3 (March 2022), 708—729

    Deepfakes, misinformation and disinformation and authenticity infrastructure responses: Impacts on frontline witnessing, distant witnessing, and civic journalism. Journalism 23, 3 (March 2022), 708—729. https://doi.org/10.1177/14648849211060644

    2022 doi

  43. [2023]

    diverge on AI regulation: A transatlantic comparison and steps to alignment

    The EU and U.S. diverge on AI regulation: A transatlantic comparison and steps to alignment. Brookings. Retrieved January 20, 2025 from https://www.brookings.edu/articles/the-eu-and-us-diverge-on-ai-regulation-a-transatlantic-comparison-and-steps-to-alignment/

    2025

  44. [2024]

    Retrieved January 22, 2025 from https://www.aoshearman.com/insights/ao-shearman-on-tech/zooming-in-on-ai-11-eu-ai-act-what-are-the-obligations-for-the-limited-risk-ai-systems

    Zooming in on AI — #11: EU AI Act — What are the obligations for limited-risk AI systems? A&O Shearman Blog. Retrieved January 22, 2025 from https://www.aoshearman.com/insights/ao-shearman-on-tech/zooming-in-on-ai-11-eu-ai-act-what-are-the-obligations-for-the-limited-risk-ai-systems

    2025

  45. [2025]

    Tech Policy Press

    Regulating Election Deepfakes: A Comparison of State Laws | TechPolicy.Press. Tech Policy Press. Retrieved January 22, 2025 from https://techpolicy.press/regulating-election-deepfakes-a-comparison-of-state-laws

    2025

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.