Research & Multi-Model Output Policy
Sources, multi-model synthesis, model routing, review claims, citations, and the limits of AI research.
1. Research is assistance
Research tools can retrieve information, compare model contributions, and produce a combined answer. They assist your investigation; they do not guarantee completeness, accuracy, freshness, or agreement with every source. This policy supplements the AI Accuracy & Reliance Policy.
Use a question with enough context to make verification possible. For a date-sensitive question, check the date, geography, product version, and meaning of the claim. A fluent answer can still be wrong, and an available search result can still be misleading.
2. Model selection and routing
Available models depend on plan access, provider availability, and the current catalog. Auto and saved Model Studio preferences determine routing as described in the product. A displayed brand or family name does not establish a particular provider model version or guarantee that a request will never use a disclosed fallback.
Multi-Model Research uses the selected supported models and combines successful contributions when the required result can be produced. Provider failures, access restrictions, timeouts, and rate limits may prevent completion. The product should describe those limitations rather than fabricate missing contributions.
3. Automatic research preferences
Enabling automatic Multi-Model Research applies your saved selection to supported future requests until changed or disabled. Review that setting when switching tasks, plans, or connected accounts. Multiple-model processing may send relevant context to more than one provider.
An automatic research preference is not approval to publish content, contact people, deploy an app, or buy a domain. Those actions have their own permissions and confirmation flows. Remove or disable research when you prefer a supported ordinary chat request.
4. Sources and current information
Search may return primary sources, secondary reporting, social posts, cached information, and summaries. Each can be incomplete or outdated. A citation should support the adjacent claim; the presence of a link is not proof that the page was authoritative or that every detail of the answer is sourced.
Read important sources directly. Check publication dates and the date an event occurred. Treat social posts and anonymous claims as evidence to evaluate, not automatic confirmation. A paywall, inaccessible page, or snippet can limit what the assistant can verify.
5. Combined answers and uncertainty
Combining several models does not create an independent consensus or multiply reliability by the number of models. Providers may share assumptions, repeat the same mistaken source, or produce inconsistent reasoning. A synthesis can lose distinctions present in individual contributions.
Look for stated uncertainty, disagreement, missing information, and the scope of verification. Ask for evidence when a claim matters. Do not treat a longer answer, research status animation, or elapsed time as a measure of accuracy.
6. Code review and testing claims
Models may review generated code, identify problems, or propose checks. A model review is different from executing code in a relevant environment. Testing claims should identify the checks actually run and their result; “looks correct” or “should work” is not an executed test.
Passing a limited check does not prove security, accessibility, compatibility, or complete correctness. Review test coverage and reproduce important behavior in your intended environment. Do not deploy code solely because several models approved its text.
7. Inputs, confidentiality, and rights
Submit only material you are authorized to share with the providers needed for research. A selected model may receive the prompt, relevant conversation, attachments, or retrieved context according to the feature. Private mode limits saved Zexir history; it does not prevent live provider processing.
Sources and generated outputs can contain copyrighted material, personal data, or confidential content. Use citations responsibly and respect access restrictions and content rights. Summaries and generated files do not automatically confer redistribution rights.
8. Reliance and reporting
Verify information before using it for medical, legal, financial, safety, employment, or other high-impact decisions. Consult qualified professionals or official channels where needed. Research tools are not emergency services and do not replace a duty to use human judgment.
Report broken citations, false completion claims, or model attribution problems to mail@zexiry.com with a safe chat identifier and concise example. Do not include confidential records merely to demonstrate an error. We may update sources, routing, or safeguards as the service evolves.
Related policies and contact
For related responsibilities, see our Acceptable Use, AI Accuracy & Reliance, Account & Security, and Subscription & Billing policies. Contact mail@zexiry.com for questions about this supplement.