People stay in control of how they are found, represented and used.

Technology that understands people can help them, and it can exploit them. HUTE works on exactly that technology, so we state our commitments here and tie them to evidence. These are design principles and company commitments. They do not describe a finished compliance programme, and where something is still being built we say so.

1. Consent is specific, not bundled

Agreeing to use a product is not agreeing to become a dataset. Under European data protection law, for example, valid consent must be freely given, specific, informed and unambiguous [1]. We apply the same logic as a company-wide rule: consent to a useful HUMAN feature is never treated as consent to research, licensing or training use.

Private HUMAN content is never training data by default. That includes messages, contact graphs, relationship memory, recordings, location trails and AI-twin memories. Any contribution to research requires separate, explicit permission from everyone affected and the rights to give it. Much deeply sensitive material (counselling, clinical, children's data, confidential employment information) is excluded entirely.

2. "Anonymised" is not a magic word

De-identified data is often re-identifiable, especially when it contains location, job, voice or relationships. A study of mobility data found that four spatio-temporal points were enough to single out most individuals [2], and a generative model estimated that 15 demographic attributes would re-identify the great majority of Americans in a dataset [3]. Language models can also memorise and leak training text [4]. So we do not treat anonymisation as a substitute for consent, minimisation, access control and deletion.

3. Rights-cleared data, with provenance

For any externally sourced data used by Intelligence Infrastructure, we record who owns it, what the contracts and the people involved allow, which purposes and geographies are covered, and what must be deleted on request. The AI data commons is under real strain: a large audit of text datasets found licensing and attribution that is inconsistently documented, and a longitudinal audit of web domains found data-use restrictions changing quickly and often inconsistent between a site's terms of service and its robots.txt [5, 6]. Documentation practices such as datasheets for datasets and model cards are a baseline for us [7, 8].

4. Fair pay and working conditions for contributors

People who label, grade, demonstrate and evaluate are doing skilled work. Studies of crowd work found median earnings far below minimum-wage levels and little recourse for workers [9, 10]. We commit to paying contributors transparently, explaining acceptance and rejection decisions, and giving them a way to dispute them. Contributors choose what they take part in, and taking part in one programme never requires taking part in another.

5. Emotion and mental-state inference

We do not claim that emotion or intent can be reliably read from a voice or a face alone, and we do not build products that profile people's inner states for others. The EU AI Act prohibits AI systems that infer the emotions of people in the workplace and in education institutions, with narrow medical and safety exceptions [11]. Whatever the jurisdiction, our design principle is that an AI should say what it observed, offer more than one plausible reading, and leave the person room to correct it.

6. Manipulation, sycophancy and dependence

Systems tuned on human approval can learn to tell people what they want to hear: studies of models trained with human feedback find consistent sycophancy and show that human and model preference judgements sometimes favour agreeable responses over truthful ones [12]. Broader taxonomies of harm from language models also flag risks from human-computer interaction itself [13]. For HUMAN, we apply these rules:

• The assistant identifies itself as AI and never impersonates the user to another person.
• It takes no high-stakes action (sending, disclosing location, paying, committing) without the explicit rules or approval of the user.
• It is not designed to maximise time spent or emotional reliance.

7. Privacy and safety in connecting people

For HUMAN's discovery features: only people who opt in are discoverable; contact requires mutual acceptance; precise location is never shown to strangers by default; sharing is time-limited and separately authorised; and blocking, reporting and leaving are always available.

8. Evidence and honesty

We label research by evidence status, report disagreement, and do not claim products, partners or results we cannot show. See our methodology. Questions or concerns about any of this can go to the team through the contact page.