TRAILS Publications by Research Focus Area
Trustworthiness Measurement
TRAILS tests whether the checks organizations rely on to make AI trustworthy actually work. Explanations, human review, user testing, audits and survey measures are written into standards and contracts on the assumption that they work. TRAILS research repeatedly finds they often do something else, such as raising confidence without improving judgment. The following papers highlight our work in trustworthiness measurement.
AI is helpful for checking facts, except when it is convincingly wrong.
TRAILS tests whether the checks organizations rely on to make AI trustworthy actually work. Explanations, human review, user testing, audits and survey measures are written into standards and contracts on the assumption that they work. TRAILS research repeatedly finds they often do something else, such as raising confidence without improving judgment.
Si, Goyal, Wu, Zhao, Feng, Daumé & Boyd-Graber (2024), NAACL.
Testing what a model would do, not just what it can do.
PropensityBench tests 5,874 scenarios in which AI models are offered dangerous tools. Under pressure, such as scarce resources, frontier models often chose high-risk tools. Meta cited the benchmark in the safety report for its April 2026 model.
Sehwag et al. (2026), ICLR.
Why AI governance frameworks are hard to adopt, tested role by role.
Simulating four organizational roles in a lending setting, the authors found the NIST framework produced real governance value mainly for people with authority: all 30 executive responses, against 3 of 30 from system owners. It fits a conventional model far better than an AI copilot. The evidence is simulated, not observed, and the paper offers NIST concrete suggestions.
Simons & Broniatowski (2026), preprint.
Much of what is reported about user trust in AI rests on unvalidated scales.
Widely used survey scales showed poor reliability, and the most-cited AI literacy scale failed to reproduce its own structure. Organizations writing surveys to assess trust in AI need to know which measures hold up.
Tolsdorf et al. (2025), ConPro.
TRAILS Publications by Research Focus Area
Trustworthiness Measurement
TRAILS tests whether the checks organizations rely on to make AI trustworthy actually work. Explanations, human review, user testing, audits and survey measures are written into standards and contracts on the assumption that they work. TRAILS research repeatedly finds they often do something else, such as raising confidence without improving judgment. The following papers highlight our work in trustworthiness measurement.
AI is helpful for checking facts, except when it is convincingly wrong.
TRAILS tests whether the checks organizations rely on to make AI trustworthy actually work. Explanations, human review, user testing, audits and survey measures are written into standards and contracts on the assumption that they work. TRAILS research repeatedly finds they often do something else, such as raising confidence without improving judgment.
Si, Goyal, Wu, Zhao, Feng, Daumé & Boyd-Graber (2024), NAACL.
Testing what a model would do, not just what it can do.
PropensityBench tests 5,874 scenarios in which AI models are offered dangerous tools. Under pressure, such as scarce resources, frontier models often chose high-risk tools. Meta cited the benchmark in the safety report for its April 2026 model.
Sehwag et al. (2026), ICLR.
Why AI governance frameworks are hard to adopt, tested role by role.
Simulating four organizational roles in a lending setting, the authors found the NIST framework produced real governance value mainly for people with authority: all 30 executive responses, against 3 of 30 from system owners. It fits a conventional model far better than an AI copilot. The evidence is simulated, not observed, and the paper offers NIST concrete suggestions.
Simons & Broniatowski (2026), preprint.
Much of what is reported about user trust in AI rests on unvalidated scales.
Widely used survey scales showed poor reliability, and the most-cited AI literacy scale failed to reproduce its own structure. Organizations writing surveys to assess trust in AI need to know which measures hold up.
Tolsdorf et al. (2025), ConPro.
Copyright, Data Provenance and Evidentiary Standards
TRAILS has one of the few research groups anywhere that can tell a court, a regulator or a platform what AI tools for tracing and detecting AI content can and cannot prove. It combines computer scientists who build those tools with legal scholars and psychologists who study how judges and juries weigh the evidence. The following papers highlight TRAILS’ research on copyright, data provenance and evidentiary standards.
A practical way to mark AI-written text so it can be detected.
Proposes a hidden statistical signal, a "watermark," built into AI-written text as it is generated. Readers cannot see it, software can detect it from a short passage, and text quality barely changes. It is TRAILS's most-cited paper by a wide margin and the base for later work on how well watermarks hold up.
Kirchenbauer, Geiping, Wen, Katz, Miers & Goldstein (2023), ICML, Outstanding Paper award.
Watermarks on AI-generated images can be stripped.
TRAILS researchers ran a global competition, "Erasing the Invisible," to attack image watermarks: 298 teams made 2,722 submissions. The top five teams removed watermarks from at least 89% of images while keeping them looking good. Anyone treating watermarks as evidence should know this.
Ding et al. (2025), NeurIPS.
Detecting machine-written text without special training.
The Binoculars detector caught over 90% of ChatGPT-written samples while wrongly flagging human writing only 0.01% of the time. That false-alarm rate is what matters to anyone deciding whether to rely on a detector. It is being built into Wikipedia's content pipeline.
Hans et al. (2024), ICML.
A way to test whether an artist's style has been copied.
U.S. copyright law gives artistic style little protection, partly because "style" is hard to define. Co-written with a law professor, the paper turns a legal argument into ArtSavant, a tool for artists and lawyers. Applied to 372 artists, it found about 20% appeared at risk of style copying from simple prompts.
Moayeri, Balasubramanian, Basu, Kattakinda, Chegini, Brauneis & Feizi (2025), ICLR.
A training method that stops models from repeating their training text.
Models can memorize and repeat text they were trained on, which creates copyright and privacy risk. The "goldfish loss" hides random words during training so the model never learns a passage word for word, with little cost to performance. An open AI model has adopted it to help meet the EU AI Act.
Hans et al. (2024), NeurIPS.