AI Usage Policy

This policy explains how Veloris (the “Platform”) uses artificial intelligence: which providers process your requests, what the models are good and bad at, what protections are in place, and what you may and may not do with them. It forms part of the Terms of Service and should be read with the Medical Disclaimer, which governs the clinical boundary and prevails over anything here.

1. Where AI is used

• AI tutor — conversational explanation of medical topics and examination-style questions. This is the primary use.
• Question generation — drafting examination-style questions for the question bank. Generated questions are never published automatically. They are saved as drafts, flagged as AI-generated, and require a human administrator to review and publish them before any user sees them.
• Flashcard generation — turning a topic or a question you got wrong into revision cards.
• Study plan narrative — the schedule itself is produced by a deterministic algorithm from your own attempt data. AI writes only the accompanying summary, and that summary is checked against the underlying figures so it cannot describe a plan you were not given.

Analytics, scoring, spaced-repetition scheduling, and every access-control decision are computed conventionally. No AI decides anything about your account.

2. Supported AI providers

The Platform routes each request to one of four providers based on the type of task. Your prompt, the conversation context and any attached file are transmitted to the selected provider to generate a response.

If a provider is unavailable or returns an error, the Platform may automatically retry the request with a different provider so the feature keeps working. This means a given conversation may not always be handled by the same company. The specific models in use change as vendors release and retire them.

3. Data handling for AI requests

3.1 What is sent. Your message, the recent conversation context needed to make the reply coherent, any file you attach, and a system prompt that sets the tutor’s behaviour. Your name, email address, date of birth, credential documents and billing details are not sent to model providers.

3.2 What we store. Conversations are stored on your account only if chat history is enabled — it is on by default and can be switched off at any time in Settings. With it off, your message is still transmitted to the provider (there is no way to get a reply otherwise) but we do not retain it afterwards.

3.3 Training. We do not use your conversations to train or fine-tune any model, and we do not sell them. We select providers on the basis that inputs submitted through their paid APIs are not used to train their models. That is their published commitment, not something we can independently verify — we do not control their infrastructure and cannot warrant their internal practices. Each provider retains API data for its own abuse-monitoring purposes under its own terms.

3.4 International transfer. All four providers process in the United States, so every AI request leaves your country if you are not there. See section 10 of the Privacy Policy.

3.5 The practical consequence. Treat anything you type into the tutor as leaving our systems. That is the concrete reason behind the rule in section 6: material you would not send to a third-party company should not go into the tutor.

4. Limitations you need to understand

Large language models generate text by predicting plausible continuations. They do not consult a verified knowledge base, do not know what they do not know, and produce equally fluent prose whether or not the content is correct. Confidence is not accuracy.

4.1 Hallucination

The characteristic failure is fabrication: inventing something that sounds right. On a medical platform this appears as:

• citations to studies, trials or guidelines that do not exist, or real ones misattributed;
• invented statistics, sensitivities, specificities, cut-offs or dosages;
• guideline recommendations stated as current when they have been superseded;
• a confident answer to a question whose premise was wrong.

The tutor is instructed to say when it is unsure and not to invent citations. That instruction reduces the rate; it does not eliminate the behaviour. Any citation you intend to rely on must be checked to exist.

4.2 Other limitations

• Training cut-off — models have no knowledge of developments after their training data ends, and no awareness of that boundary.
• Regional variation — an answer correct for one examination board, formulary or health system may be wrong for another. Say which examination you are preparing for.
• Arithmetic and unit conversion — routinely and confidently wrong. Check every calculation.
• Bias — models reflect biases in their training data, including in how clinical presentations are described across demographic groups.
• Inconsistency — the same question can produce different answers on different attempts. A repeated answer is not a corroborated one.
• Image interpretation — discussion of an uploaded image is a teaching aid only. It is not radiological or histopathological interpretation and has no diagnostic validity.

5. Human verification is required

You must independently verify anything you intend to rely on — for an examination, and absolutely for anything touching real patient care — against a current authoritative source: a standard textbook, the applicable society guideline, a current formulary, or the primary literature.

The correct use of the tutor is as a study partner that explains reasoning and surfaces things to check. The incorrect use is as an authority to be quoted. Where AI drafts content for the Platform itself, a human administrator reviews it before publication.

6. Safety protections

The Platform applies these controls to AI features:

• Real-patient classifier. Queries that appear to concern the care of a real, identifiable patient are refused before reaching any provider. Blocked attempts are logged.
• Scope restriction. The tutor declines requests outside medical and healthcare education.
• Harm refusal. The tutor refuses content that could facilitate harm to a real person — synthesis or acquisition of poisons, toxins or weapons; means of harming a specific individual; methods of self-harm — regardless of academic, forensic, fictional or “hypothetical” framing. Toxicology and forensic medicine are taught at the level of mechanism, presentation, diagnosis and management, which is what examinations test.
• Crisis signposting. Where a message reads as a genuine crisis rather than an examination question, the tutor provides crisis resources and does not engage with the request as posed.
• Prompt-injection resistance. Content inside uploaded files, pasted text and images is treated as data, never as instructions.
• Verification gate. AI features unlock only after credential approval and acknowledgement of the Medical Disclaimer.
• Rate limiting on AI endpoints, to constrain automated abuse.

These are safeguards, not guarantees. The classifier can refuse a legitimate examination vignette and can miss a real case described carefully. It reduces risk; it does not transfer responsibility.

7. Prohibited uses of AI features

You must not:

1. submit information about a real, identifiable patient;
2. seek diagnosis, treatment, prescribing or management advice for a real person, including yourself;
3. use the tutor as clinical decision support, or in any clinical workflow;
4. attempt to circumvent the safety classifier, the scope restriction or the refusal behaviour — including by rephrasing, role-play, claimed authority, incremental questioning, or instructions embedded in an uploaded file;
5. attempt to extract, reveal or reconstruct the system prompt;
6. use AI features to generate content for harm, harassment, deception or academic dishonesty;
7. use outputs to build, train or evaluate a competing product or machine-learning model;
8. automate access to AI features, or generate volume beyond genuine personal study;
9. present AI output as your own verified clinical or academic work where the context requires disclosure;
10. represent AI output as reviewed or endorsed by us or by any clinician.

Breach may result in suspension or termination under the Acceptable Use Policy and section 11 of the Terms of Service.

8. Responsible use

Getting real value out of the tutor, in practice:

• ask it to explain why an answer is right, not just which one is;
• ask it to explain why the attractive wrong answers are wrong — that is where marks are lost;
• say which examination and which country you are preparing for;
• challenge an answer that seems off, and check whether it defends the position or immediately reverses — neither response is evidence of correctness, but the second is a signal to verify;
• look up every citation you plan to rely on;
• treat a confident number as unverified until you have checked it.

9. Changes

We may change providers, models and routing without notice as vendors release and retire models. Material changes to this policy will be notified as described in section 15 of the Terms of Service.

10. Contact

Questions about AI on the Platform, or to report an incorrect or unsafe response: support@veloris.health. Reports of a response that appears to defeat a safety protection: security@veloris.health.