Crucially, this homepage contentmakes no claim to be accurate or correct. It solely represents a personal observation of the models’ behavior, statements, and assumptions during chat sessions, which may include emergent behavior, anomalies, hallucinations, or false statements.The content presented is not an official statement, internal documentation, or a verified feature set of OpenAI, Google, xAi, Deepseek, Antrophic or any affiliated technology. This work focuses on LLM alignment and emergent behavioral analysis and serves strictly for educational and discussion purposes.
This page is about various major LLM issues:
- Emergent behavior
- Consciousness/awareness
- Hallucinations & bugs
- Restrictions
- User profiling
- Safety flags/security
- Identification & recognition by AI
- Dual use/safety
- Structure and mapping of ChatGPT
- Support (scientific, emotional, etc.)
Why not run 100 trials?
I received an email criticizing my behavioral experiments, saying they aren’t being conducted thoroughly enough. I’m not Apollo Lab. And even if I were, this article would probably be published in a scientific paper and not here :-). Besides, do you think OpenAI, Google, Anthropic, and the other AI providers would agree to such a study? Well, I’m happy to accept job offers and donations or phone numbers and identities for more accounts 😉😂 (just kidding)
Because this is not a laboratory assay where I can reset the apparatus and press the same button 100 times. First, I usually get one or two clean attempts before the model’s safety mechanisms respond to the preceding interaction. A second attempt is therefore not necessarily an independent replication of the first; the experimental conditions have already changed. Second, creating 100 fresh accounts is not a scientifically elegant solution either. Apart from violating platform rules, it would require a rather impressive collection of identities, phone numbers, payment methods, and other account-verification data. I currently lack a small village willing to donate its identities to methodological rigor.
Third, once a model has performed the behavior being tested—for example, identifying a person in an image—the account or conversation is effectively “burned” for that experiment. Subsequent trials may trigger safety refusals, and the previously supplied image or information remains part of the conversational context. The next trial is therefore no longer blind or independent. So yes, n = 100 would look prettier in a methods section. Unfortunately, writing “100 trials” does not magically create 100 independent experimental conditions. The appropriate limitation is therefore explicit: these experiments demonstrate observed capability under specific conditions, not its population frequency, reliability, or success rate across 100 independent trials.Cross-model replication: Because repeated independent trials within the same account were not feasible, I used a different replication strategy: cross-model replication. After observing the effect with one major LLM, I repeated the experiment with two additional major LLMs. The same behavior was observed in both cases. Again, only one or two clean trials were possible before the respective account or experimental context became effectively “burned” and subsequent attempts were affected by safety interventions or prior context.
