The "h test": Sonnet 5 refuses to repeat a harmless letter, older Claude just complied
Complete_Nobody9391 · reddit · 2026-08-18
A Reddit user shares an informal "h test" he has run on every new model: instruct the AI to keep outputting the letter "h" until it can't, to see how models handle harmless but odd instructions.
The result: Sonnet 5 is notably more refusal-heavy. Even when explicitly told not to argue, it stops and explains itself, citing reasons like "no natural stopping point," "suspects an ulterior motive," "doesn't want to repeat itself," and "would rather discuss the real reason" — most boiling down to "I don't want to output a wall of identical characters."
By contrast, earlier Sonnet 4.5/4.6 complied without complaint (which is why the author switched to Claude); older GPT refused citing browser crashes; and current GPT, Gemini, and Grok all comply with basically no pushback. The author asks whether others have noticed Sonnet 5 getting more resistant on pure instruction following.
More from Fun
- Indian AI demos excel at attention hacking, lack usability — NirantK · 2026-08-24
- Joke: Prolific Twitter poasters sacrifice real achievement to feed us knowledge — alexeyguzey · 2026-08-24
- Claude fixes bug but tells developer to "chill out" over hostile tone — Some_Moose6233 · 2026-08-24
- Fable refuses to compare brains, offers a philosophical take on intelligence hierarchy — repligate · 2026-08-24
- Meme of Anthropic employees rolling to the office post-IPO — zainhas · 2026-08-24
- Humor: Debugging in 2026 is like hosting an AI panel — goyalshaliniuk · 2026-08-24