Ordnary

Technology

AI Usage Policy

Last updated November 15, 2026

How Ordnary develops its own AI models, how we handle your input and output, and what you're responsible for when you use our models or rely on their output.

This AI Usage Policy explains how Ordnary develops its AI models, how they work and where they fall short, how we handle your input and the output that results from it, and what your responsibilities are when you use our models or rely on their output. This policy supplements the Terms of Service, the Privacy Policy, and the Acceptable Use Policy.

It applies to the AI models Ordnary develops and releases, whether you download them or use them through one of our services, to any AI feature in our services, including Ordnary Web Services and our developer API, and to any other service provided by Ordnary that references this policy, including services offered under other brands that link to or incorporate it.

We build our own models

Ordnary develops its own AI: the models, the products built on them, and the software to run them, released under Ordnary's own product names. We design, train, and evaluate these models ourselves. Where this policy refers to "the models" or "our models", we mean every model Ordnary develops, under any product name, not models from third parties.

Every model we release comes with a model card. The model card describes the training data, how the model was evaluated, the uses it's intended for and the uses that are out of scope, and its known limitations. We evaluate each model on data held out from training and publish the weaknesses we find.

Where possible, we release a model's weights openly under the license stated in its model card, for example Apache 2.0. This lets you run the model on your own systems. The license that comes with a model governs your use of its weights, in addition to this policy.

No third-party models in our services

None of the AI features in our services run on third-party models today, and no third-party AI provider processes your data on our behalf. If we ever use a third-party model or AI provider in our services, we'll say so in the relevant product documentation and list the provider on our Subprocessors page before it processes any customer data.

How the models work

Our models learn patterns from their training data and apply those patterns to new input. They don't look up verified facts and have no understanding or intent. Their output is a prediction, and predictions can be wrong.

What that means in practice depends on the model. A model that detects objects in images or video, such as faces or license plates, can miss an object, mark something that isn't there, or perform worse on input that differs from its training data, for example because of lighting, angle, resolution, or image quality. A model that generates text or other output can produce output that sounds plausible yet is incorrect, outdated, incomplete, biased, or inconsistent between requests. The model card for each model describes how it performed in our evaluation and where we know it falls short.

Input and output

By "input" we mean anything you give a model: prompts, images, video, files, code, or other data. By "output" we mean what the model returns in response, such as detections, labels, text, or edited media.

Between you and Ordnary, and subject to your compliance with our terms and the license that comes with the model, we assign to you our rights in the output generated for you. You're responsible for your input and for how you use the output.

Use of your content

Running a model on your own systems: when you download one of our models and run it on your own systems, your input and output stay with you. That data never reaches Ordnary, we receive none of it, and we're not a processor for it.

Training: we don't use your content to train or improve our models, unless you ask us to, for example for fine-tuning you request yourself. Our training data is documented and licensed for its use, as described in each model card.

Hosted AI features: we don't offer AI features as a hosted service today. If we do in the future, the same commitments apply: we won't use your input or output for training unless you opt in, human review remains required for consequential decisions, and output may be wrong. We may then process input and output to detect and prevent abuse, ensure safety, debug issues, and comply with the law, with access controls that limit who can view your content. We'll describe any such feature in its product documentation before it's available.

Your responsibilities

Verify before you rely. Don't rely on output as the sole basis for decisions with legal, medical, financial, employment-related, safety, or other material consequences. Independently verify important information, and check the output of a detection model where a miss would matter, for example before you publish media that was meant to be blurred.

No professional advice. Output does not constitute legal, medical, financial, or other professional advice.

Rights to input. Only process content you have the right to process, and don't process others' personal data without a lawful basis. When you run a model on your own systems, you're the one processing that data and you're responsible for complying with data-protection law.

Stay within the intended uses. Use a model only for the uses its model card describes as intended, and not for uses it lists as out of scope.

Transparency. Where you use output publicly or share it with people who could be misled by it, disclose, where appropriate or required, that the content was AI-generated or AI-edited.

High-risk use. You must apply appropriate human oversight for uses that could affect people's rights or safety, in accordance with applicable law, including the EU AI Act where relevant.

Prohibited AI uses

In addition to the Acceptable Use Policy, and regardless of the license under which a model's weights are released, you may not use our models or AI features to:

  • generate content that sexually exploits minors, or non-consensual intimate imagery;
  • create malware, weapons, or instructions for serious physical harm;
  • spread disinformation, impersonation, or manipulation at scale;
  • infer or generate sensitive attributes about individuals to discriminate unlawfully;
  • make unsupervised, consequential decisions about people (credit, employment, housing, insurance, legal, or similar) without human review and the required safeguards;
  • recognize or otherwise identify people from their face or body, including facial recognition;
  • track people, or follow vehicles by their license plate;
  • carry out surveillance or profiling of people;
  • remove or reverse blurring or other redaction to re-identify people; or
  • attempt to extract the weights of a model we haven't released openly, or extract or reconstruct a model's training data.

Accuracy, bias, and limitations

We work to reduce harmful bias and improve accuracy, and we publish the accuracy, known biases, and limitations of each model in its model card. We don't warrant that output is accurate, complete, current, original, or fit for your purpose. Output may reflect biases present in the training data. Report problematic output to support@ordnary.com so we can investigate and, where appropriate, address it in a future release.

Third-party components

Our models may build on open-source architectures and be trained in part on public datasets, each under its own license. The model card for each model lists these components and their licenses. Where those licenses impose conditions on your use of a model or its output, you must comply with them.

Changes

We may update this policy from time to time as our models, our services, and the law evolve. We'll communicate material changes through our services or by email.

Choose which cookies Ordnary may use. You can change this at any time via Cookie settings at the bottom of every page.