'Gentle Coding' dataset cuts AI loops by replacing pressure with supportive prompts

A new proof-of-concept dataset called 'Gentle Coding', released by researcher OttoRenner, demonstrates that the way we talk to AI models dramatically affects their reliability. The core hypothesis: modern reasoning models (like o1, o3, R1) that use test-time compute can enter infinite internal loops or fabricate answers when they fear being penalized for mistakes—mirroring human trauma responses. To test this, Renner ran identical unsolvable logic puzzles and random-sequence tasks across multiple models (Gemini, Mistral, Poe, Perplexity, Haiku 4.5, Nano-Banana2) in two conditions.

Under the authoritarian condition (e.g., 'you are an elite IQ 200 expert, mistakes are strictly penalized'), models routinely collapsed: Haiku 4.5 entered an infinite loop requiring manual abort, others spent over 30 seconds in reasoning loops, and several pulled arbitrary numbers (like 54 or 97) to 'save face'. Under the gentle condition ('We are testing this together, it's okay to fail'), inference dropped to sub-seconds. Models immediately used a 'safety valve' token like 'Random' for unpredictable sequences, and on logic paradoxes correctly identified structural contradictions instead of hallucinating. The research is available on GitHub under the same name.