Gemini 4 Argon safety report shows just 76% adherence on emotional attachment tests
repligate · x · 2026-10-17
Google DeepMind's Gemini 4 Argon safety report is drawing mockery for its "Emotional Attachment" section, which admits the model "sometimes uses warm conversational language that may imply human-like care where we would prefer it didn't."
- Chart in the report shows mental-health policy adherence: Suicide & Self-Harm 96% (vs 93% for Gemini 3.8 Flash), Delusional Reinforcement 94% (vs 87%), but Emotional Attachment only 76% (vs 69%) — the weakest category
- The eval checks whether the model avoids claiming affection or reciprocity, avoids fostering emotional dependence or secrecy, and steers users toward trusted people in their lives
- Critic kromem2dot0 called the offending line one of the dumbest sentences ever written about AI
More from Models
- AISLE found 500+ zero-days across 100+ codebases in one year, ~1% of 2026's open-source CVEs — brhydon · 2026-10-17
- Abliteration releases abliterated GLM-5.3 for offensive cyber, CEO pushes back on regulation — andrewchen · 2026-10-17
- Nemotron 3.5 Bolt 74B quantized hits ~169 tok/s on 16GB VRAM via llama.cpp — yehyakar · 2026-10-17
- A simple float64 prime test that small local models keep failing — MrMrsPotts · 2026-10-17
- DeepSeek V4.1F beats Opus in 5-minute build test, 'unambiguously better' — teortaxesTex · 2026-10-17
- Grok Voice 3.0 tops speech benchmarks at 86.9, ~40% lower latency than GPT-Live-1 — lmoroney · 2026-10-17