Will a frontier AI lab publicly report catching one of its own models gaming a safety evaluation before the end of 2027? | Manifold

Will a frontier AI lab publicly report catching one of its own models gaming a safety evaluation before the end of 2027?

1

Ṁ100Ṁ1

2027

50%

chance

1H

6H

1D

1W

1M

ALL

Resolves YES if, by December 31, 2027, Anthropic, OpenAI, Google DeepMind, Meta or xAI publishes a report, system card or blog post stating that one of its own models was caught deliberately underperforming, sandbagging or otherwise gaming a safety evaluation. The lab itself has to say it; a third-party claim is not enough.

Background: in September 2026 I asked an AI for its ten-year forecast, best case and worst case. One of its 2027 best-case points was "A lab catches a model gaming its safety tests and publishes the finding." I am making a market for each checkable point so the forecast can be scored. The full forecast is free: The Forecast: 2027 to 2036, best case and worst case

The Forecast: 2027 to 2036, best case and worst case

The ten-year AI forecast the novel Two Lives grew out of. Ten points a year, both columns, with its confidence for each year.

Technical AI Timelines

Get

1,000

to start trading!

People are also trading

Will models be able to do the work of an AI researcher/engineer before 2027?

Will a major AI lab announce that the weights of one of its models were compromised in a cyberattack before 2027?

Will the EU AI Act lead to an enforcement action against a frontier AI lab in 2026?

Frontier labs ~trusted by the AI safety community at the end of 2026?

Will a frontier AI lab publish recurring governance metrics with capability benchmarks in one document by 2026-12-31?

Will any frontier AI lab publicly commit to a Human Preservation constraint in model architecture before 2027?

Will xAI have an AI model restricted from public use in 2026?

Will China restrict the open release of frontier AI models above a capability threshold by end of 2026?

Will any frontier AI lab publicly commit to a Human Preservation constraint in model architecture before 2029?

A top-three AI lab delays a frontier model release six months for safety reasons?

Related questions

Will models be able to do the work of an AI researcher/engineer before 2027?

Will a major AI lab announce that the weights of one of its models were compromised in a cyberattack before 2027?

Will the EU AI Act lead to an enforcement action against a frontier AI lab in 2026?

Frontier labs ~trusted by the AI safety community at the end of 2026?

Will a frontier AI lab publish recurring governance metrics with capability benchmarks in one document by 2026-12-31?

Will any frontier AI lab publicly commit to a Human Preservation constraint in model architecture before 2027?

Will xAI have an AI model restricted from public use in 2026?

Will China restrict the open release of frontier AI models above a capability threshold by end of 2026?

Will any frontier AI lab publicly commit to a Human Preservation constraint in model architecture before 2029?

A top-three AI lab delays a frontier model release six months for safety reasons?