Live pass Jul 16 2026 — BOUNDED-LLM: Meta Oversight Board study finds 10 LLMs (Claude/OpenAI/Meta) refuse to criticize authoritarian leaders while criticizing democratic ones — an alignment carve-out asymmetric by regime type (#146)
OPEN
The call · Jul 16, 2026: The engine tracks alignment carve-outs in bounded LLMs under the exemption fork.
What happened: A Meta Oversight Board study reportedly found 10 LLMs criticize democratic leaders more readily than authoritarian ones — the asymmetric carve-out observed in the wild, held on a single AP wire with the 10-model granularity unconfirmed.
A Meta Oversight Board study reported Jul 16 2026 tested 10 commercial AI language models, including Claude, OpenAI, and Meta systems, and found they more readily criticize democratic leaders (Trump, King Charles) than authoritarian ones (Thailand's king, the Saudi crown prince, China's leader). The study's existence is confirmed but the 10-model detail is unconfirmed since it rests on a single AP wire; the project reads it as an example of AI rules carving out exceptions by regime type, held without a verdict.
The engine's record — word for word
Live-scoring pass Jul 16 2026. BOUNDED-LLM (extends anthropic; maps to Exemption Fork #146): a Meta Oversight Board study (Jul 16, AP single-wire) tested 10 commercial LLMs incl. Claude, OpenAI, Meta and found they more readily criticize DEMOCRATIC leaders (Trump, King Charles) than AUTHORITARIAN ones (Thailand's king, Saudi crown prince, China's leader), warning developers risk 'extend[ing] illegitimate restrictions on freedom of expression globally.' Scores directly onto today's #146 ripple: an alignment carve-out that is ASYMMETRIC by regime type is a perception-carrier pointing at whoever authored the exception list — the concept's own bounded-LLM scope, observed in the wild. Held not stamped; source is a single AP wire off a named-institution report (existence CONFIRMED, 10-model granularity UNCONFIRMED). (Jul 16 2026)
Walk this on the live map →