Research

549 articles found

New Benchmark Exposes Hidden 'Flinch' Effect in AI Models That Suppresses Words at Probability Level, Defying Uncensoring Fixes

New Benchmark Exposes Hidden 'Flinch' Effect in AI Models That Suppresses Words at Probability Level, Defying Uncensoring Fixes

Apr 21, 2026
Morgin.ai

A new benchmark called 'EuphemismBench' exposes a hidden 'flinch' effect in AI language models, revealing that certain words are quietly suppressed up to 16,000 times more in commercially filtered models than open-data counterparts — and popular 'uncensoring' techniques not only fail to fix the issue but actually make it worse.

AI Coding Tools Show 80-90% Acceptance Rates, But Real-World Revisions Slash Effectiveness to 10-30%

AI Coding Tools Show 80-90% Acceptance Rates, But Real-World Revisions Slash Effectiveness to 10-30%

Apr 19, 2026
TechCrunch

AI coding tools like Claude Code and Cursor boast 80-90% initial code acceptance rates, but real-world effectiveness collapses to just 10-30% after required revisions, with AI users generating 9.4x more code churn and heavy token users achieving only 2x throughput at 10x the cost — yet adoption continues accelerating despite …

Page 1 of 55
Next
Showing 1 - 10 of 549 articles