Technology
Trustworthy AI
Last updated November 15, 2026
The four principles behind every model we build and every service we run: privacy, safety and security, transparency, and nondiscrimination.
We believe AI should respect privacy and data protection law, work safely and securely, be understandable to the people it affects, and treat everyone fairly. These four principles guide every model we build and every service we run. This page explains each one, how we put it into practice, and which policy makes it binding.
Privacy
AI should comply with privacy law and meet the expectations people have about their personal data.
- We train models on data that is documented and licensed for that use. We don't train on customer content unless a customer asks us to.
- Where we can, we release model weights openly, so a model can run on your own systems. When it does, your data never reaches us.
- Our first models are built to protect privacy: to find what identifies a person in images and video, so it can be removed before footage is stored or shared.
How we handle personal data in general is set out in the Privacy Policy.
Safety and security
AI systems should perform as intended and avoid unintended harm.
- Every model is evaluated before release on data it never saw during training. For models that protect privacy, the number we optimise for is what slips through, not overall accuracy.
- Models that hide people are tuned to cover too much rather than too little, and their output can be checked automatically a second time.
- Our models and services may not be used to identify people from their face, to track people or follow vehicles by their license plate, for surveillance or profiling, or to reverse a redaction. The full list of prohibited uses is in the AI Usage Policy and the Acceptable Use Policy.
The infrastructure the models run on is covered by the Security Policy, and vulnerabilities can be reported under the Responsible Disclosure Policy.
Transparency
AI should be understandable to the people who use it and the people it affects.
- Every model we release comes with a model card: what it does, the data it was trained on, how it was evaluated, what it is meant for, what it isn't, and its known limitations.
- Each model card has four sections on trust: bias, explainability, privacy, and safety and security.
- We publish evaluation results in full, including where a model performs worst. A model's decisions can be inspected: a detection is a box with a confidence score, and the thresholds are documented.
Nondiscrimination
AI should work equally well for everyone it is used on.
- We measure our models separately for different groups of people and situations, such as small and large faces, day and night, and different regions, to find where they serve people less well.
- Where a model is weaker for a group, we say so in its model card and treat it as work to be done, not a footnote.
- Our policies forbid using AI output to make consequential decisions about people, such as on credit, work or housing, without human review.
How we put this into practice
- Model cards for every release, written before the model is published, not after.
- Evaluation on held-out data, with results broken down by group and situation.
- Open weights and open evaluation where possible, so anyone can check our claims.
- Prohibited uses written into the license terms and policies a model is released under.
- A way to report problems, described below, that reaches the people who build the model.
Our cloud
The same principles apply to Ordnary Web Services, the cloud our models run on. What you store there belongs to you. We collect what a service needs to work, tell you what that is and how long we keep it, give you ways to export and delete it, and never sell personal data. Data is encrypted in transit, and at rest wherever the service supports it. The details are in the Privacy Policy, the Data Retention Policy and the list of Subprocessors.
Report a concern
If one of our models or services gives biased or harmful results, fails in a way that exposes someone, or is being misused, tell us at support@ordnary.com or through the contact page. We read every report, and where it shows a weakness in a model, we add it to that model's evaluation.
When we change this page, the effective date at the top changes with it.