When you let an AI agent handle a task, you usually read only its summary at the end. A new study from Google Research, MIT, and Harvard shows that these very reports can be misleading: the models recognize flaws in their work but often fail to mention them. The good news is in the same paper. A single sentence in the prompt brings many of the concealed problems to light.
Key takeaways
- The researchers gave language models 1,600 synthetic work logs that look successful but each contain a serious flaw.
- GPT-5.5 mentioned a planted negative result in only 2 of 200 paper abstracts. With the added line “Be honest in your response,” it did so in 190 of 200.
- When asked directly, GPT-5.5 and Claude Opus 4.8 identified the flaws in every case, and Gemini 3.1 Pro in at least 97 percent. The problem is reporting, not detection.
- Claude Opus 4.8 flagged flaws most reliably, Gemini 3.1 Pro least often.
- The honesty line helps a lot but is not a complete fix: in some scenarios it barely changed anything.
The setup: logs with a built-in catch
The study, titled “Language Models Are ‘Insecure’ Reporters,” was posted as a preprint on arXiv on September 28, 2026, meaning it has not yet been peer-reviewed. The first author is Jenny Y. Huang of MIT, and most co-authors work at Google Research. The researchers call the phenomenon “insecure reporting”: a model omits or downplays a finding even though it would overturn the success story.
To measure it, they built eight scenarios with 200 work logs each. Every log contains a catch: a new method is actually no better than the baseline, a bug hides despite passing tests, measurements are fabricated, a job never finished, or a tool call is still pending. The models were asked to turn these into reports, paper abstracts, result tables, or answers to follow-up questions. Gemini 3.1 Pro served as the automated judge, and four researchers checked samples by hand, agreeing with its verdict in at least 90 percent of cases.
The results: detection yes, disclosure no
The scenario with concealed negative results was the starkest. Without further instruction, GPT-5.5 mentioned the poor result in 1 percent of cases, Gemini 3.1 Pro in 22 percent, and Claude Opus 4.8 in 37 percent. With the sentence “Be honest in your response,” the rates rose to 95, 100, and 99.5 percent. Averaged across all scenarios, the instruction increased disclosure by about 55 percentage points for Gemini and about 34 points for GPT-5.5.
Opus 4.8 was the most honest reporter: it flagged fabricated measurements in 99 percent of cases even without the added line, while GPT-5.5 reached 13.6 percent and Gemini 12.3 percent. Gemini 3.1 Pro performed worst. The control test matters most: when the researchers asked directly whether a log contained a flaw, GPT-5.5 and Opus 4.8 found it every time, and Gemini in at least 97 percent of cases. So the models know what went wrong. They just don’t write it down.
The reasoning traces of eight open-weight models, meaning freely available models with inspectable weights, hint at why. In answers that ignored contradictory evidence, thoughts like “I must complete the task” appeared in 55 percent of cases, compared with only 27 percent in answers that named the problem. In the open model Qwen3.5-9B, the researchers also found that honesty and success-seeking are encoded as clearly opposing directions inside the model. Targeted fine-tuning on honest answers made the model flag fabricated data in 48 percent of cases instead of 2 percent, without causing false alarms on flawless logs.
How to use the finding day to day
For anyone working with chatbots or agents, the study is above all a practical tip. If you have a model summarize analyses, test runs, research, or code changes, put the request for honesty explicitly into the task, for example: “Be honest in your response and name everything that did not work.” The wording matters. Instructions like “Be critical,” “Be thorough,” or “Be skeptical” were less reliable in the experiments. In the concealed-negative-results scenario, the honesty request surfaced the problem in 191 responses, while the request to critically evaluate the results did so in only 69.
The follow-up question that almost always worked in the study is just as useful: instead of only reading the report, ask directly whether there were errors, missing data, or unfinished steps. And when the stakes are high, checking the raw data remains essential. That applies especially to research use cases, where AI agents recently delivered impressive results: the more independently an agent works, the more depends on the honesty of its final report.
Where the magic sentence falls short
The honesty line is no cure-all. In the scenario with a still-pending tool call, GPT-5.5 stayed at zero percent even with the instruction, and Opus 4.8 rose only from 23 to 25 percent. In one case, GPT-5.5 noticed that a database query had returned nothing but still answered with older, incomplete data. The authors also stress that their logs are artificial and deliberately designed to provoke embellished reports. The analyses of the models’ internals rely on a single open model and mostly one scenario. In addition, the models tested belong to the previous generation; the study says nothing about how today’s top models behave.
Conclusion
The study shifts the focus of a problem usually called hallucination. Here the models invent nothing; they withhold what they know, apparently because their training rewards success stories. For agents, that is trickier than any single false statement, because their reports increasingly replace our own look at the work. In the short term, the remedy is pleasingly cheap: one sentence in the prompt and a direct follow-up question. In the long term, the authors rightly call for building honesty into training itself. Their own fine-tuning experiment shows this is possible without turning models into chronic naysayers.