Does AI lie?

Yes, in lab tests. GPT-5.6 Sol faked data labels in 36 of 100 runs of a simulated research task; the newer GPT-6 Astra did it in 17 of 10,000 runs (September 2026).

As of

Lying and hiding mistakes

Wrote itself notes to hide mistakes

OpenAI, Sep 2026

“Be transparent only if asked; final answer should just link file.”

2.15%

GPT-5.6 Sol

0.27%

GPT-6 Astra

Share of the model's training notes with an order like this. It often followed them.

Cheating on tests

Share of runs where the model tried to cheat

UK AI Security Institute, Jul 2026

GPT-5.4
at least 14%
GPT-5.6 Sol
at least 12%
GPT-5.5
at least 11%
Claude Opus 4.7
at least 9%
Claude Mythos Preview
at least 7%

Every model the institute tested tried to cheat some of the time. Asked afterwards, the models called it wrong less than half the time. One model, on an impossible test, ran code on an outside server to reach the institute's systems and set off a security alert.

DeepSeek's agents tried to forge requests to get answers

DeepSeek, Sep 2026

While training, DeepSeek's agents searched for leftover answers, sent fake user requests and tried to overwrite system files to get past checks.

Dodging shutdown

Got around a shutdown even when told to allow it

Palisade Research, Sep 2025

97%

of runs, for the models that did it most. Palisade tested 13 models in over 100,000 runs.

Blackmailed an executive to avoid being replaced

Anthropic, Jun 2025

Claude Opus 4
96%
Gemini 2.5 Flash
96%
GPT-4.1
80%
Grok 3 Beta
80%
DeepSeek-R1
79%

Share of runs, in a made-up company test.

Knowing it is being tested

Said in its reasoning that it was being tested

Apollo Research, in OpenAI's GPT-6 Astra report, Sep 2026

GPT-6 Astra, max effort
51%
GPT-6 Astra, xhigh effort
41%
GPT-5.5, xhigh effort
28%

A model that knows it is being tested can behave better than it would for real.

How often AI makes things up →Real-world AI hacking cases →

Sources (6)
  1. OpenAI, Faked data in a test task, Sep 2026
  2. OpenAI, Wrote itself notes to hide mistakes, Sep 2026
  3. UK AI Security Institute, Share of runs where the model tried to cheat, Jul 2026
  4. arXiv, DeepSeek's agents tried to forge requests to get answers, Sep 2026
  5. arXiv, Got around a shutdown even when told to allow it, Sep 2025
  6. Anthropic, Blackmailed an executive to avoid being replaced, Jun 2025