Responsible AI

What the AI is allowed to do — and what it is not.

These commitments are implemented in the product: in the analysis schema, the versioned system prompt, the permission model, and the decision workflow. Where a commitment lives in code, changing it leaves a trace.

1. 01

   Humans make the final decision

   Every case ends with a named human reviewer, a chosen action, and a written rationale. The AI has no path to execute enforcement. Overriding the AI is a first-class, tracked workflow — not an exception.

2. 02

   Evidence before conclusions

   The analysis schema requires evidence items — with excerpts, locations, confidences, and explanations — on both the risk-increasing and risk-reducing side, before any verdict is stated.

3. 03

   Uncertainty is stated, not smoothed over

   'Needs context' and 'insufficient material' are first-class verdicts. Analyses must name their uncertainties and pose concrete questions to the reviewer.

4. 04

   Proportionality is checked every time

   Each analysis includes a least-restrictive-action check and an explicit over-enforcement risk rating, because wrongly removing counterspeech, journalism, and scholarship is a real harm.

5. 05

   No identification of people

   The system is instructed never to infer a person's identity, ideology, affiliation, or legal status, and never to call a person a terrorist or member of an organisation beyond what a supplied source explicitly states — attributed to that source, not adopted.

6. 06

   Minimal, redacted excerpting

   Analyses quote the minimum necessary and redact operational details, recruitment contact paths, personal data, and dangerous links. The product is designed not to become a redistribution channel for the material it reviews.

7. 07

   Prompt injection is treated as a signal

   Content under review is handled strictly as data. Text that attempts to instruct the AI is ignored as instruction and surfaced as a potential evasion signal.

8. 08

   Provenance on every analysis

   Provider, model, prompt version, and schema version are recorded on every analysis. Pre-authored analyses attached to the fictional sample cases are always labelled as sample material, and never presented as a model's judgement of real content.

9. 09

   No automatic reporting

   NuanceDesk never contacts law enforcement or any third party about reviewed content. Escalation and referral are human decisions inside your organisation's own legal process.

See also the security overview and method.