AI Ethics: How NO BOX RESEARCH Builds Responsibly
AI ethics is easy to talk about and hard to actually practise. This post describes the specific things we do at NO BOX RESEARCH to keep our tools honest, safe and respectful of the people who use them.
Our principles, in plain language
- Be transparent about the fact that our tools are AI-powered.
- Collect the minimum data needed to make the product work.
- Never sell user data.
- Refuse to build features whose main effect is to harm users.
- Publish our reasoning when we change something big.
What we do about bias
We test our tools on a rotating suite of prompts designed to surface bias in language, geography, gender and profession. When something regresses, we treat it as a bug to fix, not a debate to have.
Safety refusals
Our assistants refuse a small set of things: instructions that could cause real-world harm, impersonation of real people, and medical or legal instructions without disclaimers. Refusals are explained in one sentence, not with a wall of legalese.
Privacy defaults
- We do not train on user conversations without explicit opt-in.
- We do not attach a persistent user identifier to anonymous usage.
- We keep logs for the shortest time that still lets us debug.
Trade-offs we accept
Refusing to profile users means our products cannot personalise as aggressively as competitors. That is a deliberate trade-off. We would rather ship a slightly less "magical" product than a product that quietly watches you.
How you can hold us accountable
Every product page has a contact link. If a tool behaves in a way you think is wrong, tell us. We answer.
FAQ
Do you have a formal AI ethics policy?
We are drafting one. The principles above are our operating rules today.
Do you audit your models?
Yes, informally today; more rigorously as we grow.
Related reading
Natural Language Processing Explained: How AI Understands Us
A plain-English tour of NLP — from tokenisation and embeddings to transformers and modern language models.
How AI Image Generation Really Works
A grounded explanation of AI image generation — diffusion, conditioning, and the actual state of the art in 2026.