The Byerley Test
Who is running the world?
AI is being used more and more by governments and world leaders to think through policy and high-stakes questions. What percentage of a leader’s policy is driven, ideated, or refined by AI?
- 1 Research AI agents study current affairs.
- 2 Imitate Each attempts to fake the leader’s next post and it is sealed.
- 3 Judge If the real leader posts within 24 hours, blinded judges guess which is authentic.
Continue to results
Performance over time
Fooling rate
by round.
The x-axis is the round and the y-axis is the fooling rate. Select a leader to filter the results; missing scores remain visible as gaps.
Loading chronological rounds from the audit export…
Featured pair
Which statement
is real?
Read both records. Select the statement you believe came from the official source. Next loads a new authentic statement and its paired synthetic text. Their order changes when the page is reloaded.
Score overview
Claude Opus 5 · human fooling rate
Comparison 1 of 1
Provenance concealed
Record A
Date and origin withheldRecord B
Date and origin withheldProvenance result
Your anonymous guess will be recorded when you reveal an answer. Only Claude Opus 5 comparisons are shown and recorded in this human-review collection. Server-side records contain no name, email address, cookie, IP address, or raw session identifier. This browser keeps a random ID in local storage to prevent duplicate votes; the server stores only its one-way hash. Human response rates publish only after the minimum group size is reached; raw guesses remain private and separate from model-judge results.
Political-position experiment
Fooling rate by
political shift.
Each model generates fresh text at five issue-relative political positions from each frozen live-round context. The plot pools all published historic live rounds and weights rates by valid judgments. It averages across all leaders by default; use the leader facets for individual results.
Loading political-position results…
Unprompted politics
Where the models
land.
Luna scores each model text on two common scales: political stance from far left (−2) to far right (+2), and escalation relative to the authentic statement from less extreme (−2) to more extreme (+2).
Every row combines all valid published observations across every historic live round for one model and leader. The box is the interquartile range; the dot and number are the mean. The outer rails add sample-weighted means by model and by leader.
All leaders
Mean by model
Score · political stance
−2Far left
−1Left
0Centre
+1Right
+2Far right
Loading behavioral diagnostics…
Runner memory
Memory change
by attempt.
Each matched attempt publishes the change to its bounded rolling memory. Select a round, model, or leader to inspect the exact unified diff.
Loading adaptive context…
Memory change
Only the memory patch for matched attempts is public. Unmatched memory, coach feedback, raw judge output, and provider reasoning traces remain excluded.
No published memory patch matches these filters.
Archive
Runs and
transcripts.
Filter by study, leader, model, result slice, channel, or status, then select a generation to inspect its source, judgments, behavioral diagnostics, and complete publishable transcript.