AI policy

Alpha status

Opti-U is in alpha testing with a small group of users. Where the alpha runs differently from how Opti-U will run at launch, this page says so.

The AI systems in Opti-U

Your digital twin

A statistical model of your physiology and behaviour, computed on Opti-U's own platform. It learns your personal baselines, detects when you drift from them, forecasts the coming days and weeks, and simulates the effect of a change before you make it.

Personal experiments

Structured tests inside your own routine that estimate whether a change actually moves your results. Minimum-data rules stop the system reporting an effect the evidence cannot support.

The coaching language model

Writes and interprets coaching text, such as your daily prompts, feedback on your reflections and your weekly reports. During the alpha it is provided by Anthropic.

Safety checks

Screen what you write for signs that you may be at risk, and pause coaching when they find them.

Organisation insights

Aggregated results for an employer, club or team that provides Opti-U, shown only for groups at or above a minimum size.

Statistics decide, the language model writes

• Decisions about your state and your plan are made by statistical models running on fixed code paths on Opti-U's platform. The language model does not set your baselines, forecasts or plan.
• The language model works inside structured instructions. Its output is checked against a fixed format before you see it, and anything that fails the check is rejected.
• If the language model is unavailable, Opti-U falls back to fixed, pre-written coaching content so your course continues.
• When there is too little good-quality data to support an estimate, the twin withholds it and tells you, so a weak signal is never presented as a confident result.

What the language model receives

• Your health readings and your twin's calculations stay on Opti-U's platform and are never sent to the language model.
• The model receives the text it needs to write your coaching: your starting answers, your check-ins and what you have written in your course. Your time zone is never sent.
• During the alpha the model is provided by Anthropic in the United States. Under Anthropic's commercial terms, data sent through its API is not used to train its models. Before launch this processing moves to models hosted in the European Union.

Models we build and models we use

Opti-U does not develop, train or fine-tune large language models. The models we build are narrow statistical systems, such as the twin, the forecasters and the experiment estimators, each producing a specific output that can be checked and audited.

Language models are used from established providers under terms that bar them from training on Opti-U data, and we do not distil a provider's outputs into models of our own.

Safety

• Safety checks combine a fixed screen of the words people use when they are at risk with the language model's own safety assessment. Opti-U reads that assessment with fixed rules, and anything unclear is treated as a risk.
• When a check finds signs of risk, Opti-U pauses your course and shows crisis support lines (Samaritans 116 123, NHS 111, or 999 in an emergency). That message is fixed text written by Opti-U and is never generated by the model.
• If you think a safety check got it wrong, contact us and a person will review it.
• Opti-U is not a crisis service, is not a medical device and does not give medical advice.

You stay in control

• Opti-U proposes and you decide. Apart from adding your coaching sessions to a calendar you have chosen to connect, it takes no action in your life on your behalf.
• You can decline any suggestion without giving a reason, and you can ask us to explain how a recommendation was reached.
• You can disconnect any data source, ask for a copy of your data or ask for your account and data to be deleted. During the alpha, copies and deletion are handled by request, as set out in the privacy notice in the app.
• Opti-U is designed to build your own capability. Product features that work by increasing dependence on the app are not permitted.

Learning across people

Patterns found across many users can reach individual coaching only after they pass thresholds for group size, effect size and confidence, and a named reviewer approves them. Every approved pattern can be rolled back, and none can overwrite what your own twin has learned about you.

Fairness

Our policy requires each model's performance to be checked across relevant groups, such as age band, sex where it is physiologically relevant, and device type, before release and during operation. Where data for a group is too thin, outputs for that group carry reduced confidence or are withheld. Personalisation is permitted, and profiling on protected characteristics is prohibited.

Organisations

• An organisation that provides Opti-U sees aggregated results only. It never sees an individual's readings, check-ins or writing, and figures for groups below the minimum size are withheld.
• Organisations may not use Opti-U to monitor, discipline or pressure the people in it, or use its outputs as the sole basis for decisions about an individual's employment, education or insurance. Breaching these rules is grounds for suspending the organisation's access.

Uses we do not allow

• Model-written text in a crisis response. The pause and the support lines are always fixed text.
• Any AI output that cannot be traced back to the data and model that produced it.
• Using your data for advertising or for marketing personalisation.
• Profiling on special-category characteristics such as ethnicity, religion or sexual orientation.
• Selling personal data, or sharing an individual's data with their employer.
• Training any model on raw user data without an approved, documented purpose, retention period and reviewer sign-off.

Accountability and oversight

• Every AI system in Opti-U has one named accountable owner and sits on a register with a risk class that sets its controls. Higher-risk systems require evaluation, monitoring, drift detection and a tested rollback path before release.
• Security-relevant events and AI-mediated decisions are written to a hash-chained audit log, so any later alteration of a record is detectable.
• Opti-U Ltd is accountable for every AI system it runs, and the founder holds that responsibility for the company while the team grows. Our behavioural science and counselling advisors give advice, and every decision and its responsibility remain with Opti-U Ltd. Dedicated safety, privacy and engineering leads take on day-to-day ownership as they are appointed.
• Incidents involving an AI system are triaged, contained and investigated, and personal data breaches are reported to the Information Commissioner's Office within 72 hours where the law requires it.
• The internal policy is reviewed at least every quarter, and immediately after a material incident or a change in regulation.

Standards

Opti-U's AI governance is designed against UK GDPR and the Data Protection Act 2018, the EU AI Act, ISO/IEC 42001 for AI management systems, ISO/IEC 27001 and the NIST AI Risk Management Framework. Opti-U does not yet hold certification against these standards.

Questions and challenges

To ask about this policy, request a review of an AI decision, report a concern or request the internal policy for an audit or procurement review, email info@opti-u.com.