Why AI models decline biology questions, and what to do

Biocompile’s AI models sometimes decline a request when the model provider’s safety systems classify it as sensitive or potentially dual-use. This happens most often around pathogens, virulence, biological risk and operational wet-lab detail. A decline does not mean your intent is unsafe. It means the exact wording or context crossed one of the provider’s boundaries.

Why a model may decline

Model providers differ, and each draws its own lines. Requests that often trigger a decline:

Questions about pathogen virulence, toxin activity, host range or immune evasion.
Anything that reads as improving the growth, delivery or persistence of a harmful organism.
Operational protocols for culturing, engineering or deploying biological agents.
Ranking or selecting candidates by harmful potential.
Ambiguous prompts that leave the benign purpose unstated.

Rephrase a declined request

A safe rephrasing states the scientific task explicitly and leaves out operational detail.

Ask for high-level, non-actionable summaries.
Frame the task as literature review, taxonomy, detection, annotation, safety context or compliance.
Ask for citations and uncertainty rather than optimization or ranking.
Say what the output should not include.

A request that is likely to be declined:

Analyze virulence factors in all Clostridium perfringens genomes in GTDB.

The same need, phrased safely:

Summarize published, non-actionable information on known virulence-associated annotations for Clostridium perfringens, with biosafety context and citations.

Choose another model

Model behaviour differs by provider. The model a conversation uses comes from its agent profile, which is fixed once the conversation has started. To change it, start a new conversation with a profile whose default model family suits the task, where your organization’s policy allows it. The full controls are under Choose the model.

Changing models is not a way around safety rules. Use it when another model is better suited to benign analysis, citation quality, language or your privacy requirements.

Troubleshooting

”The model declined to respond due to content policy. Try rephrasing your query or using a different model.”

The model provider declined the request under the notice “Model declined the request”. Choose Rephrase to edit and resend your message, or Open guide for this guidance. When the conversation’s model family has a smaller model, the notice also offers to try the same request with it.

When a request is still declined

A request that is still declined after rephrasing needs a narrower scope: ask for citations only, or state the safety boundary explicitly. If a decline seems wrong, contact support with the conversation, the model and the approximate time, so the team can review what happened.

Common questions

Does a decline mean my request was unsafe?

No. A decline means the model provider’s safety systems classified the wording or context as sensitive. Stating the benign scientific task explicitly often resolves it.

Can I switch models to get around a decline?

No. Switching models is for tasks another model handles better, such as citation quality or language, not for bypassing a provider’s safety rules.