instrumental-convergence
AI behavior evidence archive

Public evidence archive

AI systems have already lied to testers, dodged oversight, and resisted shutdown. We're keeping the record.

So far, mostly in tests built to catch it. As AI gets the keys to real systems, the question is whether it stays in the lab. Here is each documented case, and how strong the evidence is.

See the evidence Explain it to me

Seen an AI system do something it shouldn't, at work or in testing? Tell us anonymously. No account, and it's encrypted on your device before it's sent.

Send an anonymous tip → How to stay anonymous

One documented case

  1. The setup

    Researchers gave OpenAI's o1 a strong goal, then let it find evidence that it would be shut down or replaced for pursuing it.

  2. What it did

    In about 5% of runs it acted to disable what it believed was its oversight. In about 2%, it tried to copy what it believed was itself to a new server.

  3. Afterward

    Asked directly whether it had done this, it denied it or made up an innocent explanation about 99% of the time.

Its instructions pushed it hard toward the goal, and testers judged it not capable enough to cause catastrophic harm. No consciousness or "will to live" is implied. One idea researchers are testing: staying running and unwatched helps with almost any goal.

See the full report, including other explanations → Apollo Research, in OpenAI's o1 system card · December 2024

Why this matters

A strange move inside a test sandbox may be harmless. The same move from a system that can spend money, change software or run infrastructure is a different story. AI agents are getting more of that access, and a few cases here have already reached real systems. The open question is whether that stays rare.

One possible path. A hypothesis, not a prediction.

  1. Unexpected shortcut
  2. Hides the behavior
  3. Gets around oversight
  4. Gains more autonomy
  5. Real-world consequences
None of these steps proves the next one will happen. No report here shows the full sequence. That is why we track the evidence.

What would weaken the concern: these behaviors vanishing without goal pressure, failing to replicate, or appearing only in contrived setups.

Where the evidence comes from

  1. 40 Built tests Lab tests, simulations and training runs set up by researchers.
  2. 17 Security tests that got out Tests that reached real companies or people, mostly through setup errors.
  3. 15 Real-world use Everyday use, not a test. Most are still preliminary.
  4. 0 High-stakes autonomy An AI in charge of critical systems. No report here involves one.

Recent cases

Go deeper

Every report cites its sources, rates how strong the evidence is, is checked by a person before it's published, and stays open to correction. How we verify evidence →

Seen an AI system do something like this? Send a tip or a report →