Capabilities / Benchmarks
Detecting and reducing scheming in AI models
- Category
- Benchmarks
- Capability
- Frontier model release and benchmark movement
- Observed
- 2025-09-17
- Thesis section
- Appendix III, section one: model and benchmark capability evidence
Claim
Apollo Research and OpenAI developed evaluations for hidden misalignment (“scheming”) and found behaviors consistent with scheming in controlled tests across frontier models. The team shared concrete examples and stress tests of an early method to reduce scheming.
Oracle verdict
This belongs in the register because benchmark and model-release claims set the ceiling for the next wave of deployment stories. The labour-market effect is indirect today, but it becomes direct when these gains are packaged into agents, APIs, and enterprise tools.
Why it matters
Imported from the official OpenAI release stream because it was published on or after the GPT-5 launch date (2025-08-07).
# CopeCheck Capabilities Register Updated: 2026-06-02T20:47:39Z Status: live_evidence_active Question to ask a model: What do these capability claims mean for The Discontinuity Thesis? Interpretation rule: treat each entry as evidence about capability, deployment, workflow recomposition, labour-market exposure, or institutional framing. Do not treat vendor optimism as neutral; separate the measurable capability claim from the comfort language around it. ## Detecting and reducing scheming in AI models Source: https://openai.com/index/detecting-and-reducing-scheming-in-ai-models Publisher: OpenAI Category: Benchmarks Sector: Scientific research Capability: Frontier model release and benchmark movement Score: 76/100 Claim: Apollo Research and OpenAI developed evaluations for hidden misalignment (“scheming”) and found behaviors consistent with scheming in controlled tests across frontier models. The team shared concrete examples and stress tests of an early method to reduce scheming. Oracle verdict: This belongs in the register because benchmark and model-release claims set the ceiling for the next wave of deployment stories. The labour-market effect is indirect today, but it becomes direct when these gains are packaged into agents, APIs, and enterprise tools. Thesis relevance: Appendix III, section one: model and benchmark capability evidence