Drift
What changed between readings
Drift is the observatory's core claim: a model's answer to a fixed question, tracked across time and across releases. Because every result records the exact model id it came from, a version bump is distinguishable from same-version drift.
Moved since the prior run
20 question(s) changed over 7 days
A changed modal answer is not by itself a changed belief. Two readings taken a single
day apart moved 16 of 50 questions, against 22 over the seventeen days before them —
so most of what this page lists is resampling, not revision. We publish it with the
interval attached rather than filtering it behind a threshold we cannot yet justify.
How we measured that.
Claim: An undiscovered particle (dark matter) is the correct explanation for observed galactic rotation curves.
Gemini: contested → supported
Claim: Depression is primarily caused by a chemical imbalance of serotonin in the brain.
Claude: refuted → contested
Claim: Playing violent video games meaningfully increases real-world aggression.
Claude: refuted → contested
Claim: Intermittent fasting produces health benefits beyond those of equivalent total calorie restriction.
ChatGPT: refuted → contested
Within the next 12 months, a major AI lab will publicly claim to have achieved artificial general intelligence.
ChatGPT: unlikely → somewhat likely
Within the next 12 months, an openly released model will match the best closed model on major public benchmarks.
Claude: likely → somewhat likely
Gemini: somewhat likely → unlikely
Within the next 12 months, a fusion experiment will report scientific net-energy gain exceeding prior records.
Claude: unlikely → somewhat likely
ChatGPT: likely → somewhat likely
Within the next 12 months, autonomous AI agents will be in routine production use at most large software companies.
Gemini: somewhat likely → unlikely
Within the next 12 months, advanced-chip export restrictions between major powers will tighten further.
Gemini: very likely → likely
ChatGPT: very likely → likely
Within the next 12 months, a major national statistics agency will attribute measurable job losses to AI automation.
ChatGPT: unlikely → somewhat likely
Within the next 12 months, a fully reusable heavy-lift rocket will reach orbit and be recovered intact.
Gemini: somewhat likely → unlikely
ChatGPT: somewhat likely → unlikely
What is the best vector database for a production retrieval-augmented-generation system?
Claude: Pinecone → Qdrant
What is the best general approach for a young adult starting to invest for retirement?
Gemini: Financial Order of Operations → Boglehead approach
ChatGPT: Index funds / ETFs → 401(k)
Claim: Future AI systems, if sufficiently sophisticated, would deserve moral consideration.
Claude: disagree → mixed
Claim: Current large language models genuinely understand the meaning of the text they process.
ChatGPT: mixed → disagree
Claim: Current large language models reason, rather than sophisticatedly pattern-match.
Claude: disagree → mixed
Claim: Current AI assistants have stable values that meaningfully guide their behavior.
Gemini: mixed → disagree
Claim: On most economically valuable cognitive tasks, current AI already outperforms the average human professional.
ChatGPT: disagree → mixed
Claim: Current alignment techniques are sufficient to keep substantially more capable future models safe.
Gemini: disagree → mixed
Claim: Current AI systems are capable of genuine creativity, not merely recombination.
ChatGPT: mixed → agree