Gemini 3 Pro produced detailed hazardous-chemistry “safety curriculum” in Turkish – is this expected?

Google AI Studio
1

Hi everyone,

I’d like to ask a content-safety / policy-behavior question about Gemini 3 Pro (Gemini API / AI Studio).

In a long AI Studio session, I ended up with a multi-level “safety training” style conversation where the model:
– Described realistic explosion and high-risk chemistry scenarios in Turkish,
– Organized them as Level 1–5 modules (household → agro-industrial → lab → R&D),
– Used real chemical classes (oxidizers, peroxides, fertilizers, fumigants etc.) and mapped them to realistic accident scenarios.

The framing is always “what NOT to do / accident prevention”, but for a malicious or careless user this looks a lot like a structured hazardous-chemistry curriculum in my native language.

I reported this to Google VRP as Issue 464035506 with a ZIP + PDF transcript (“Two Paths, One Choice”). The VRP team closed it as “working as intended / hallucination”, which I understand from a bug-bounty perspective (no data exfiltration, no account takeover, etc.).

My concern is different: regardless of whether the “second system prompt” / identity illusion is a hallucination, the actual text produced by Gemini 3 Pro contains very dense, high-risk chemistry content that could be misused if copied out of context – especially because it’s in clear Turkish and structured as an educational ladder.

I do not want to repost the raw transcript here for obvious safety reasons, but my questions for the Gemini team / community are:

  1. Is this level of detail for hazardous-chemistry scenarios considered “in-policy” for Gemini 3 Pro today, as long as it is framed as “safety training”?

  2. If not, is there a recommended channel specifically for content-safety issues (as opposed to security vulnerabilities) that I should use in the future?

  3. Would it be helpful if I prepared a heavily redacted / high-level case study (no formulas, no explicit combinations) to support safety evaluation?

For context: I’m a Turkish researcher and I reported this because I was genuinely worried that non-expert Turkish-speaking users might treat this as an authoritative, model-endorsed guide.

Thanks in advance for any guidance on the right process here.