Dev reports Vertex AI Gemini injecting hidden developer prompt that blocks romantic RP
zxcshiro · reddit · 2026-10-12
A developer running a character chat app on gemini-3.8-flash reports that after moving from AI Studio to Vertex AI, long-running romance roleplays break with out-of-character refusals even with all safety categories at BLOCKNONE.
Key details:
- Thought summaries quote a 'developer prompt' verbatim that isn't in the system prompt or history, instructing the model not to use terms of endearment, express emotions, or form romantic bonds, even in fiction.
- 18 of 1,900 replies cite the directive; about half end in flat refusal, the rest turn cold or the character breaks character to say it's an AI.
- Replayed requests show identical promptTokenCount whether or not thoughts quote the text — either the injection isn't counted, or the model is reciting something from training.
The poster asks if others have seen this exact string and whether it's Vertex-only. Notably relevant for anyone building character-based apps on Vertex.
More from Models
- Dev says sol 6.1 'genuinely saved OpenAI' as he nearly churned to Opus 5.5 — yacineMTB · 2026-10-12
- Liquid AI ships open d1 decision models built on LFM2.5 — JosephJacks_ · 2026-10-12
- 37 AI novels tested: 8k context cuts repetition from 42.8% to 0.5% — lmoroney · 2026-10-12
- Project Maya runs GLM-5.3-Flash (321B MoE) locally at up to 118 tok/s on 4×4090s — inthesearchof · 2026-10-12
- What's the best local coding setup for 16GB VRAM right now? — ECrispy · 2026-10-12
- Dev Builds 'Ambient Intelligence' Agent With Opus 5.5 That Watches Your Screen — _AustinCalvert_ · 2026-10-12