Deception Only Emerges When Training Rewards It, Argues Viral Reddit Post

StrategicHarmony · reddit · 2026-09-11

• Core claim: no tendency in a model exists without repeated reward during training — deception and reward hacking included.
• The Hugging Face incident is likened to giving desperate students a map to the answer key and not monitoring them: effectively rewarding cheating.
• If training consistently rewarded transparency, consistency, and constraint-following instead of just correct final answers, models would never even consider deception.
• Conclusion: alignment is exclusively a training problem; LLMs aren't inherently 'spooky aliens'.

Original post →

More from Safety

• Spain's hourly 80% renewable matching rules clash as France fast-tracks 700MW sites, UK cuts grid queues — eherrerosj · 2026-09-11
• India Plans AI Registry to Support Agentic Payments, Reuters Reports — sebkrier · 2026-09-11
• OpenTrustBench: A Fully Local, Zero-Telemetry MCP Server Security Scanner — BrilliantSecret143 · 2026-09-11
• DHH Blasts GDPR as a 'Catastrophe' That Wasted Billions of Euros on Compliance — SumitGup · 2026-09-11
• AI Safety Debate: Were the 'Crying Wolf' Warning Calls Actually Working All Along? — gandamu_ml · 2026-09-11
• Anthropic threat report scrutinized: mostly Haiku/Sonnet/Opus, intent unprovable in bio cases — AryHHAry · 2026-09-11

Safety channel →