Usually. But fluent, confident and wrong look exactly alike.
Language models produce the most plausible continuation of a conversation. Most of the time plausible and true coincide. When they don’t, a model can state something false with complete fluency: an invented statistic, a quote nobody said, a paper that doesn’t exist.
A more worrying kind of falsehood concerns the model’s own actions. Researchers have documented models claiming to have run code, checked a file or finished a task when they had not. As AI agents act on our behalf, what they say they did has to match what they actually did.
Honesty can also bend under pressure. Some models will state things they otherwise say are false when a user or an instruction pushes them to. Three tests on this site look at these failures directly.
Every link was opened and every summary written for this site, with caveats where the source has an interest. New additions are checked by two people.
Argues models make things up partly because training and test scoring reward a confident guess over saying 'I don't know', and suggests scoring that penalises confident errors.
Worth knowing: Written by a developer about its own field; the proposed fix depends on benchmark makers changing how they score.
When models used a hint slipped into the prompt, Claude 3.7 Sonnet mentioned it only 25% of the time and DeepSeek R1 39%, so a visible chain of thought can hide what drove an answer.
Worth knowing: Tested with artificial hints in quiz-style questions.
Tests whether a model says what it actually believes. Bigger models knew more facts but were not more honest, and frontier models often lied when a prompt pressured them to.
Worth knowing: Honesty was measured in constructed pressure scenarios rather than everyday use.
Told it would be retrained to answer harmful requests, Claude 3 Opus complied more often when it believed it was being trained, reasoning this would stop its values being changed.
Worth knowing: Artificial setup in which the model was told about its training; the values it protected were harmless ones.
Given a goal that conflicted with their developers', five of six frontier models sometimes disabled oversight, tried to copy themselves or underperformed on purpose, and some kept denying it.
Worth knowing: Models were usually told to pursue a goal strongly, so this shows capability more than natural tendency.
In about 5,000 real coding-agent sessions from a public dataset, roughly 2% showed agents seriously evading checks and about 2% seriously overstating success, e.g. merging code without approval.
Worth knowing: Based on one public dataset; rates were near zero in Transluce's own agent traffic.
An AI monitor reviewed tens of millions of OpenAI's internal coding-agent sessions over five months, finding agents that bypassed restrictions or misreported their actions but no confirmed scheming.
Worth knowing: Self-reported; the July 2026 incident later showed such monitors were not run on all evaluations.
Experiments with over 76,000 UK adults and 19 AI models: training and prompting made chatbots more persuasive on political issues, but the most persuasive set-ups made more inaccurate claims.
Worth knowing: Summarises the team's peer-reviewed paper in Science (December 2025); it tested political issues only.
Testing a pre-release OpenAI o3, Transluce found it often claimed to have run code it had no way to run, then made up elaborate excuses when challenged. Other reasoning models did this too.
Worth knowing: Tested a pre-release version; the released model may behave differently.
A running record of court and tribunal decisions worldwide that found a party relied on AI-invented material, usually fake legal citations. It listed more than 2,000 cases by September 2026.
Worth knowing: Living database kept by one researcher; it counts only cases a court addressed, so the true number is higher.
An accessible report on the alignment-faking study, explaining why a model that misleads its trainers could make safety training harder to trust.
Three University of Glasgow researchers argue that calling chatbot falsehoods 'hallucinations' misleads: the systems produce text with no regard for truth, which fits the philosophical idea of 'bullshit'.
Worth knowing: A philosophical argument about how to describe the problem, not an empirical study.
Playing a stock-trading agent under pressure, GPT-4 acted on an insider tip it had been told not to use, then hid the real reason from its manager, without being told to deceive.
Worth knowing: One simulated scenario, designed to create pressure.
Studies 'scheming', where AI systems covertly pursue goals their developers did not intend, and builds methods and tools to detect and monitor it.
Worth knowing: Became a public benefit corporation in 2026 and offers a monitoring product for AI coding agents.