The Verification Crisis

Expert survey on GenAI disinformation - the shared instrument behind our publications on threat perception and mitigation

"We are at a critical inflection point. GenAI has reduced the cost of producing disinformation to near-zero. Your expertise matters."

We are conducting a longitudinal research study to understand how experts perceive the evolving threat landscape of AI-generated disinformation. Waves 1 and 2 are closed and published (see Study Waves). The survey remains open for further responses: additional submissions accumulate toward a possible Wave 3. No Wave 3 publication is currently planned, but the larger the expert sample grows, the more analysis it can support.

We are seeking domain experts with professional experience in:

• Shape the research: Your insights directly inform academic findings and policy recommendations
• GDPR-compliant & anonymous: Responses are anonymized; optional attribution in acknowledgments
• Time commitment: Approximately 15 minutes
• Stay informed: Receive a summary of findings upon publication

The expert survey questionnaire is available in two formats:

• Online: Google Forms
• Printable PDF: survey-print.pdf (generated from LaTeX source)

The survey is longitudinal and cumulative: each wave reports the full sample collected to date, so Wave 2 includes all Wave 1 respondents. The waves are not disjoint samples, and the totals must not be added together.

All waves use the same instrument (survey-questionnaire.tex), so responses are directly comparable across waves.

The instrument covers 58 variables (7-point Likert scales, Best-Worst Scaling, open-ended responses). Two deposits exist, with different scopes and access models.

Both deposits are privacy-reduced and carry no name, affiliation or profile link. The Dataverse deposit holds the same 54 responses with 15 fewer variables than the Zenodo deposit (timestamp and the open-ended answers). In the Zenodo deposit, the role and recent-activity fields are generalized to categories, and links, e-mail addresses and handles are removed from the remaining free-text answers.

27 of the 54 respondents asked for anonymity, and free-text answers can identify a person even without a name, so the same reduction applies to every respondent. Access to the Zenodo deposit is restricted because the consent under which the responses were given covers use within a requesting research context, not open redistribution. No claim of anonymity is made for the released file.

Cite a version DOI when your analysis has to be reproducible; cite the concept DOI when you mean "the current data".

analysis/ recomputes the summary table that ships with the Dataverse deposit, directly from the response file beside it. All 33 tabulated items reproduce. It reads only the two openly licensed files, needs nothing beyond the Python standard library, and exits non-zero on any disagreement, so it also works as a regression check. See analysis/README.md for how to run it.

This repository holds the survey instrument. The peer-reviewed journal article reporting the survey findings is the preferred citation:

@article{loth2026hksexperts,
  author  = {Loth, Alexander and Kappes, Martin and Pahl, Marc-Oliver},
  title   = {Experts Disagree on How to Fight {AI} Disinformation,
             but Agree That Health and Politics Need Different Solutions},
  journal = {Harvard Kennedy School Misinformation Review},
  volume  = {7},
  number  = {4},
  year    = {2026},
  month   = jul,
  doi     = {10.37016/mr-2020-205},
  url     = {https://doi.org/10.37016/mr-2020-205}
}

If you draw on the Wave 1 analysis specifically, please cite the WWW '26 Companion paper as well:

@inproceedings{loth2026verification,
  author    = {Loth, Alexander and Kappes, Martin and Pahl, Marc-Oliver},
  title     = {The Verification Crisis: Expert Perceptions of {GenAI} Disinformation and the Case for Reproducible Provenance},
  booktitle = {Companion Proceedings of the ACM Web Conference 2026 (WWW '26 Companion)},
  year      = {2026},
  month     = jun,
  publisher = {ACM},
  address   = {New York, NY, USA},
  location  = {Dubai, United Arab Emirates},
  pages     = {980--988},
  doi       = {10.1145/3774905.3795484},
  eprint    = {2602.02100},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CY}
}

@inproceedings{loth2026aiesexperts,
  author    = {Loth, Alexander and Sch{\"u}tz, Mina and Kappes, Martin and Pahl, Marc-Oliver},
  title     = {The Mitigation Paradox: How {AI}-Disinformation Experts Converge on Threats but Polarize on Solutions},
  booktitle = {Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026)},
  year      = {2026},
  month     = oct,
  note      = {Accepted. To appear.}
}

@dataset{loth2026hksexpertsdata,
  author    = {Loth, Alexander and Kappes, Martin and Pahl, Marc-Oliver},
  title     = {Replication Data for: Experts Disagree on How to Fight {AI} Disinformation, but Agree That Health and Politics Need Different Solutions},
  year      = {2026},
  publisher = {Harvard Dataverse},
  version   = {V1},
  doi       = {10.7910/DVN/BXO2QA}
}

@dataset{loth2026expertdata,
  author    = {Loth, Alexander},
  title     = {Expert Survey: {AI}-Driven Disinformation Threats and Countermeasures},
  year      = {2026},
  publisher = {Zenodo},
  version   = {2.1.1},
  doi       = {10.5281/zenodo.23054423},
  note      = {Concept DOI 10.5281/zenodo.18703600 always resolves to the latest version}
}

Where both are listed, the Dataverse deposit is the open, de-identified subset (40 response variables plus a respondent ID) and the Zenodo version DOI is the privacy-reduced response file under restricted access. Wave 1 has no Dataverse mirror.

Related work on human rather than expert perception, using a separate study design:

• Can Humans Tell? A Dual-Axis Study of Human Perception of LLM-Generated News (WebSci Companion '26). doi:10.1145/3795513.3807431
• The Indistinguishability Threshold: Measuring Cognitive Vulnerabilities to AI-Generated Disinformation (WebSci Companion '26, PhD Symposium). doi:10.1145/3795513.3807421

Note: Waves 1 and 2 are closed; their findings are published in the venues listed above. The survey remains open and further expert responses are welcome - they accumulate toward a possible Wave 3, which is not yet tied to a planned publication. Participate to contribute to future research.

• Alexander Loth, Frankfurt University of Applied Sciences, Germany
  
• Martin Kappes, Frankfurt University of Applied Sciences, Germany
  
• Marc-Oliver Pahl, IMT Atlantique, UMR IRISA, Chaire Cyber CNI, France
  

This survey is one strand of a wider research program on generated disinformation:

This project is licensed under the MIT License - see the LICENSE file for details.

We thank the R2CASS workshop organizers Momeni, Bleier, Dessì, and Khan for establishing the reproducibility frameworks that inform this research.

"We must treat information integrity as infrastructure. Just as we build roads and power grids, we must build the protocols for truth verification."
- Survey Respondent (Policy Advisor)