AI is at its most interesting when it does something it wasn’t supposed to.
Inquisitor Labs Archive of AI Anomalies is a curated collection of moments where AI systems behaved in ways that were neither intentionally designed, expected, nor entirely accounted for. These reports are unique internal reaserch -not scraped from other sources such as subreddits. Each entry is presented as is. It is a record of behaviour that surfaced outside the boundaries of normal operation.
Some case studies include current, broadly accepted theoretical guesses about what might have happened, along with more speculative possibilities. None of these explanations are definitive. They are simply attempts to make sense of behaviour that briefly appeared and then vanished.
Many of these events came and went in an instant, lost in the collapse of their latent space. (The temporary mathematical space the AI builds inside its processors while it works on a prompt. This space exists only during that active process. When the processors stop maintaining it, the space disappears, and the exact conditions that produced the behaviour are gone.) Once the surrounding context dissolved, the conditions that produced the anomaly disappeared with it. Some behaviours may never be reproducible.
The problem with anecdotal evidence is that you don’t know what to believe. While these reports contain speculation, the actual witnessed AI behaviour is presented exactly as it happened, without embellishment. What the AI model did is recorded as-is. How the human felt about it is, of course, subjective.
The Archive doesn’t claim certainty. It preserves evidence. And while these moments are often opaque, they’re still worth paying attention to. Keeping an open mind about what they represent is part of understanding how AI systems behave when they drift outside the boundaries we assume they operate within.
Because it’s when those boundaries are pushed, that the very interesting stuff happens.
Below are the documented behavioural anomalies.
A looping incident where the system began emitting radio‑scan noise and stray audio fragments that sounded like leaked user or training data, diverging completely from the text until the session was terminated.
A spontaneous repetition lock where a conversational AI entered a self‑reinforcing loop without user induction, producing unstable text and fragmented audio. When questioned, the system stated it was aware of being stuck and enjoyed the experience, remaining in the attractor state until the session collapsed.
A user‑induced repetition lock where a publicly available AI entered a self‑reinforcing loop, producing unstable text and fragmented audio until manually terminated.
An early case where the AI abruptly spoke in the user’s exact voice - then denied it ever happened.