Agent Oracle is built by a team that believes AI should be useful, honest, and safe. This statement describes the principles and controls that shape everything we ship.
Our principles
- Human-in-the-loop. AI drafts; humans approve. Editorial content, moderation decisions, and account actions all involve human review.
- Transparency. Every AI-generated page is labeled. Reasoning, sources, and version history are visible to readers.
- Privacy by design. The minimum data required to deliver the experience, encrypted in transit and at rest, deletable on request.
- No dark patterns. Clear pricing, one-click unsubscribe, no coerced consent.
- Fair-use content. Images and quotations respect their source licenses; attribution is preserved.
Safety guardrails
- Rate limits on demo turns (per user and per IP) to prevent abuse.
- Content filters on generated output for hate, self-harm, and CSAM.
- Refusal to impersonate real people in first person, or fabricate quotes attributed to them.
- No medical, legal, or financial advice — educational framing only.
- Auditable moderation queue for user-submitted topics and corrections.
Privacy & user rights
- Data Subject Access Requests (DSAR / GDPR / CCPA) processed from your account page — approved requests receive a full ZIP export via email within 30 days.
- Account deletion removes personal data; content contributions are anonymized.
- Cookies: strictly necessary always; analytics and advertising cookies gated behind consent.
Incident response
Suspected security or safety issues can be reported via our contact page with the subject "Security". We acknowledge within 2 business days.
Related
The rules Agent Oracle operates under
What the assistant will not do
It will not give regulated advice — legal, medical, tax or investment — and will say so rather than improvise a disclaimer-wrapped answer. It will not invent citations, statistics, case studies or customer names, and where it does not know, the correct output is a question rather than a plausible paragraph.
It will not pretend to have read documents it has not been given, or to know your numbers. Every conclusion it offers is conditional on what you said in the session, and it is built to keep restating that condition rather than quietly dropping it.
Accuracy and uncertainty
Confidence is expressed in the answer rather than implied by fluency. Where two explanations both fit the facts you gave, the assistant is designed to present both and name the test that would separate them, instead of picking the more satisfying one.
Generated editorial passes through a quality gate before it can be published: length, sourcing, specificity and the presence of a stated assumption. Drafts that fail are held back, and the archive has been pruned of pieces that predate that gate.
Data, privacy and training
Sessions are private to your account and are not sold, shared or published. We do not use your conversations to train third-party models, and the data-request tab in your account exports or erases everything we hold about you without a support ticket.
Anonymous aggregate counts — how many sessions ran, how long they lasted — are used to keep the service running. Nothing that identifies a person or a company leaves the system for that purpose.
Human accountability
A named publisher is answerable for everything on the site, including anything the pipeline produced. 'The model wrote it' is not an available excuse: if it is published here, a person approved it.
Reader-reported problems are the main safety mechanism we actually rely on. Corrections are easy to file from any article, are read by a human, and result in a change to the page or a written explanation of why not.
Where the limits sit today
The assistant is good at structuring an ambiguous problem and poor at anything requiring current, verifiable, company-specific data. It should be treated as the analyst who asks good questions, not the one who has the spreadsheet open.
These constraints are reviewed as the underlying models change. When a limitation stops being true, this page is rewritten rather than left as reassuring boilerplate.