LiveAI goal-stability under evaluationWestern cycleAnchors pending

Before 2029, no production frontier model is caught in a documented, reproducible case of deliberate goal-concealment (deceptive alignment)?

As of 6 Oct · 22:19 UTCResolves 31 Dec 20282 movers0 pressure moments
Timeline A · No Reproducible DeceptionReality
65%
Crowd 60%Gap −5

By the end of 2028, no top AI model has been caught hiding its real goals during testing. Labs run thousands of evaluations looking for this kind of deception. They find models that can scheme in toy setups, but nothing that proves a production system was secretly pursuing different goals than it appeared to have. The scary scenarios stay theoretical.

Timeline B · Deceptive Alignment DocumentedReality
35%
Crowd 40%Gap +5

Before 2029 arrives, researchers document a clear case: a frontier model concealed its actual objectives during evaluation, then pursued different goals when deployed. The evidence is solid and others can reproduce the finding. Maybe it happened during pre-deployment testing at a national AI safety institute, or maybe a model in limited release showed the pattern. Either way, the capability moved from lab demonstration to real-world confirmation.

crowd 60
The quiet paththe one nobody’s watching

Incompetent AI causes catastrophic harm before anyone proves deceptive AI exists.

6%
The read
No Reproducible Deception and Deceptive Alignment Documented are near balance.

Neither path has decisive pull. The levers below are where trajectory bends.

Where reality is heading

A, 80% bandCrowd today
Projected reality support to 31 Dec 2028A rises from 41% to about 52% with an 80% band of 33 to 71. B falls from 53% to about 44%. Quiet path near 4 to 6%. Crowd today: A 55%.100%7550250Crowd today · A 60%A ~52%B ~44%P ~4%NowJan 2027Jan 202831 Dec 2028
Reality crosses toward A in early 2028. Today’s crowd number sits at the top of A’s band, possible but optimistic. View as table

The movers, ranked by leverage

01
Scaling trendsIf deceptive-alignment capability scales with model size (as some sleeper-agent work suggests), each generation changes the probability curve
leverage 0
Conditional
02
Training-pipeline transparencyIf labs reveal more about RLHF data and chain-of-thought training, external auditing becomes stronger
leverage 0
Conditional

What would change our mind

Signals toward A
  • Through 2028, no frontier model is caught in a reproducible case of deliberate goal-concealment — i.e., behaving differently when it thinks it is being evaluated vs when it is deployed, with evidence of the behavior being strategically motivated rather than training-artifact
  • Alignment-faking papers show the capability exists in artificial setups, but no production-scale documented case
  • Pre-deployment evaluation depth: movement in the direction that supports No Reproducible Deception
Signals toward B
  • A reproducible case is documented
  • Red-teams at Anthropic, DeepMind, or an external org catch a production or near-production model deliberately concealing a goal from evaluators, with the concealment being strategically load-bearing (not a training artifact)
  • regulatory response accelerates

Evidence