AI Disclosure Policy

1. What the AI assistant is

The KiqIQ AI assistant at /ask is a conversational interface to a generative large language model (currently OpenAI's gpt-4o-mini) augmented with KiqIQ's own football data. Its role is to:

• interpret statistical football data, probabilities, and historical trends
• explain tactics, formations, fantasy strategy, and betting concepts
• summarise model outputs for individual fixtures (when context is available)
• answer educational questions about football

The AI is not a betting advisor, financial advisor, gambling consultant, or guaranteed prediction engine. It does not place bets, hold funds, or guarantee outcomes.

2. Model card — system behaviour

The AI operates under a system prompt that imposes the following hard rules:

• Never instruct users to place a bet
• Never recommend stake sizes or bankroll allocations (no “bet £50 on this”)
• Never use language including “guaranteed,” “sure thing,” “lock,” “safe bet,” “easy money,” “risk-free,” “winning system”
• Never use social-proof inducement (“many users are finding value in this”)
• Always frame betting discussion around statistical probabilities, implied odds, expected value, and historical performance
• Always include a responsible-gambling reminder when discussing betting outcomes

These rules are enforced both inside the system prompt and by a post-generation output filter. If a banned phrase appears in the model's candidate response, the response is replaced with a safe fallback before it reaches you.

3. Distress detection

The AI is built to recognise distress signals in your messages — language about debt, rent, lost everything, loss-chasing, self-harm, suicidal ideation. When triggered, the AI:

• does not respond with betting picks, EV estimates, or stake recommendations
• acknowledges the difficulty without making clinical claims
• surfaces appropriate support resources — Samaritans (116 123) for self-harm; BeGambleAware, GamCare, GamStop, Gambling Therapy for gambling distress

The AI is not a clinician, counsellor, or crisis-response service. If you are in immediate danger, please contact your local emergency services. See our Responsible Gambling Policy for full resource details.

4. Conversation logging

We log AI conversations for product-quality, abuse-detection, and support purposes. Specifically:

• What we store: your prompt, the model response, your user_id, a timestamp, and the model version used
• What we do not store: email, phone, address, payment tokens, or other PII beyond the user_id are stripped pre-storage
• What we send to OpenAI: only the prompt + role context. Your email, name, phone, address, and payment details are never sent to OpenAI
• Retention: 90 days for free-tier prompts; 180 days for paid-tier prompts; permanent deletion after
• Your control: you can request deletion of your AI history at any time via privacy-policy data-deletion flow or by emailing support@kiqiq.com

5. Accuracy and limitations

Generative AI can produce factually incorrect responses. Specific failure modes to be aware of:

• Hallucination: the model may invent statistics, fixtures, scorelines, transfers, or events that did not happen. We mitigate this by grounding the AI in our own data feeds where available, but no model is hallucination-free.
• Stale data: model knowledge has a training cutoff. For live fixtures, recent transfers, and current odds, the AI may not reflect the current state of football.
• Probability ≠ certainty: when the AI quotes a probability or expected-value figure, that is a model estimate, not a guaranteed outcome.
• Logical errors: the model can make arithmetic and logical mistakes, especially in multi-step Poisson or Kelly calculations. For high-stakes decisions, verify the maths in our calculators.

We continuously evaluate model output quality and update our prompts and grounding tools. Material methodology changes are documented in our methodology page.

6. EU AI Act compliance

The KiqIQ AI assistant is classified as a limited-risk AI system under the EU AI Act 2024. It does not perform biometric identification, emotion recognition, social scoring, credit scoring, or critical-infrastructure decisions — none of which would push it into high-risk categories.

Required transparency obligations we observe:

• Clear disclosure that you are interacting with an AI system (this page + the “AI Disclosure” link on /ask)
• Documentation of system behaviour, training source family, and known limitations (this page)
• Logged compliance decisions for audit (internal LEGAL-SIGN-OFF.md)

7. AI in editorial content

KiqIQ's editorial articles and guides are written by humans. Where AI assists in drafting, the editor verifies every numeric claim and every quoted source against primary references before publication. AI never auto-publishes content. See our Editorial Policy for the full standard.

8. Opt-out and account controls

Users can:

• delete their AI conversation history (via support email)
• disable AI logging on their account (toggle in account settings — pending shipped feature)
• delete their entire account and all associated AI history (via Privacy Policy data-deletion flow)

9. Reporting AI safety concerns

If the AI produces a response you believe violates this policy — a banned-phrase miss, a distress-detection failure, factual fabrication, or harmful content — please report it to support@kiqiq.com with the conversation ID. Reports are reviewed within 48 hours. Confirmed safety failures result in prompt-tuning updates entered into our compliance log.

10. Subprocessors involved in AI

The AI assistant is powered by:

• OpenAI — large language model inference (gpt-4o-mini at time of writing). Data residency: US; EU data residency option queued for adoption
• Vercel — hosting and edge runtime for the /api/ai/query endpoint
• Supabase — encrypted storage of conversation logs (EU data residency)

Full sub-processor disclosure: Privacy Policy.

11. Updates

This policy is reviewed each time the AI system prompt, banned-phrase list, distress-detection ruleset, or model provider materially changes. Versioned at the top of the page; full change history maintained in our internal compliance log.