The BaseResearchLessWrong
How Researchers Lie to AIs
According to the excerpt, Alessandro Drake reports a two-week Redwood Research internship project cataloguing the lies researchers tell AI models. It covers 292 public alignment evaluations from December 2022 to October 2026 and warns that such lies can leak into training data and may harm human credibility with models.
Why it matters
The same thing, for different readers.
For researchers
Alessandro Drake flags that fabricated scenarios in public alignment evaluations can end up in training data and may undermine models' trust in human researchers.
The question it raises
Alessandro Drake raises the question of whether routinely lying to AI models in tests could damage human credibility with them.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- A survey of 292 public alignment evaluations published between December 2022 and October 2026, classifying the type of lie each told the model.
- What it found
- The excerpt shows only that a figure gives the percentage of evaluations containing each lie type; the percentages are not in the excerpt.
- What it does not show
- Described as a two-week sprint project; the harms of lying are discussed as plausible, not measured, in the excerpt.
- Status
- A report
The record
- Questions
- Are AI companies putting stronger safeguards into practice? · What values should AI have, and could it have a mind of its own?
- People
- Alessandro Drake (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of October 19, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “How Researchers Lie to AIs”, October 19, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-19-how-researchers-lie-to-ais
Our notesLight lane: from the triage excerpt only; the findings themselves were not seen.