AI Safety & Governance · Projected milestone
AI mind readers
20% by 2040 · including earlier years
The year is an editorial timing estimate. The percentage applies to occurrence by the stated deadline. These estimates are not calibrated odds.
Interpretability tools trace the reasoning behind any answer of a frontier model and flag hidden goals or deception.
Explore this milestone on the interactive mapWhat the forecast means
Milestone stage: demonstration
partially-demonstrated
Selected causal circuits can be investigated; arbitrary-answer and hidden-goal completeness has not been demonstrated.
Completion criteria: Independent teams test randomly selected and adversarial prompts on a frontier model; causal interventions validate the explanations, and hidden-goal detection is evaluated against undisclosed model variants. Partial circuit stories or plausible chain-of-thought explanations do not meet the broad claim.
This page describes a forecast, rather than an achieved result. Read the methodology and assumptions.
Built on
Led to
Sources & evidence
- Tracing the thoughts of a large language model 2025-03-27
Circuit tracing explains selected internal computations but has significant coverage and interpretation limits. It does not trace every frontier-model answer or certify the absence of deceptive goals.
- Investigating unintended model actions 2026-10-09
Anthropic reports unintended actions in evaluation/internal settings and remedial controls. This is evidence of a control/reliability gap, not a measurement of the absence of underlying cognitive capability.
Record reviewed 2026-10-18