Responsible AI Commitments
The governance commitments behind LeraLink's use of AI, and where we are not there yet.
Last updated: 13 November 2026
How LeraLink Uses AI describes what each AI feature does. This page sets out the commitments that govern all of them. EU AI Act at LeraLink explains how the EU AI Act applies.
1. People are the authors
A work on LeraLink is written by the people whose names are on it. AI never writes, titles, summarises, researches or evaluates a work for its author. It never changes published text, approves a translation, decides on rights, or gives anyone a role. People think, write, review and decide; AI does bounded, supporting labour.
2. A fixed set of AI tasks
AI is used only for these tasks:
| Task | What it does |
|---|---|
| Translation | Proposes a translation of a Piece |
| Search-engine fields | Suggests SEO fields for a Piece |
| Feature image | Generates a proposed feature image |
| Search index | Turns published text into numerical representations so it can be found by meaning |
| Concepts and people | Names the concepts and people that published text is about, and writes a description of them from the published evidence |
| Composer | A composition assistant that makes proposals only, where it is enabled for an installation |
There are no general-purpose agents. No AI task can call another to produce a work. An installation cannot configure an autonomous agent. A new task is added only as a change to the platform, reviewed and published in How LeraLink Uses AI.
3. Proposals and provenance
What AI produces is a proposal, marked as machine-made, until a person accepts, edits or rejects it. What was made by a model stays marked as such; a person's change is recorded beside it, not over it. The platform records which task and which model answered each request.
There are four narrow exceptions, each stated openly:
- Machine translations published with a work. Machine translations can be shown to readers before a person has reviewed them; they are then labelled "AI translation · Not yet human-reviewed".
- Unreviewed search-engine fields go live when a person publishes, and are marked "needs review" until someone approves, rewrites or rejects them.
- Concept and person descriptions are computed from published text and marked as machine-made. A field a person has fixed is never written over by an automatic process.
- The search index is computed from published text. It is infrastructure, not content, and is not shown to readers.
Body text, titles in the original language and rights reach readers only after a person decides on them.
4. Labelling
AI-written text proposals and generated images are labelled as machine-made in the Studio, where a person decides on them. Machine translations that readers see are labelled as such. A generated image that a person has accepted is marked for readers with a small sign, "Image generated with AI", on the work's page, in image Blocks, on catalogue cards and on the Wall.
Machine-readable marking. Since 13 November 2026, every image generated in LeraLink carries IPTC "trainedAlgorithmicMedia" metadata in its file, and resized copies keep it. Google states that images generated by its Gemini image models also carry its SynthID watermark. Not yet done: the reader's sign does not yet appear on some surfaces: homepage modules, the newsletter, printed publications, share images and link previews. The transparency obligations of Article 50 of the EU AI Act apply from 2 August 2026; for generative systems placed on the market before that date, the machine-readable marking of Article 50(2) applies from 2 December 2026 (Regulation (EU) 2026/1744).
5. No model reads restricted content
A Piece restricted to named people or roles is never sent to any AI model: not for translation, suggestions, images, the knowledge layer or the search index, and not even when someone clicks a button. Its translations are written by people or imported. When a Piece becomes restricted, its search index entries are removed within minutes.
6. AI policies per installation, with no silent fallback
Each installation has an AI policy:
| Policy | What it allows |
|---|---|
| Standard | The platform's default models for each task |
| EU only | Only models hosted in the EU and recorded as not training on inputs; the installation's text is left out of the search index |
| Open models only | Only models with published weights |
| No external AI | Only models running on Cloudflare |
The policy is enforced in the one place every model call passes through. If a model fails, only the next model the policy allows is tried; there is no fallback around the policy. If the policy leaves no model for a task, the task is unavailable and nothing is sent. A model counts as meeting a rule only when the fact the rule needs has been recorded; an unknown fact does not count.
Current status: the facts about each model are not yet fully recorded. Until they are, an installation on the EU only policy has no AI features. Every installation is on Standard today.
Installation administrators can see, read-only, which tasks are on and which models would run, with each provider's country, hosting region, training and retention terms. On request, we provide a printed overview, "Models used by this installation".
An installation writes editorial instructions, such as how to translate, an image style or SEO guidance. It never writes a prompt or chooses a model.
7. No training on your content
SSStudio does not train or fine-tune AI models on customer content.
Each task tries its models in a fixed order set by the platform; an installation's administrators can see that order, with each provider, on the page "AI in this installation". The providers used are Cloudflare Workers AI, OpenAI (translation, SEO suggestions, concept descriptions) and the paid tier of Google's Gemini API (images). Under their published terms for the paid API services we use, OpenAI and Google do not train on the data sent and may keep it for a limited period to detect abuse. We record these terms for each provider and model, and review them when a provider changes its terms. Subprocessors and Recipients lists each provider.
8. Uses we do not allow
LeraLink does not use AI, and does not offer AI features, to:
- identify or recognise people from images, voices or other biometric data, or sort people into biometric categories;
- profile people in order to make decisions about them, such as access, membership, roles, the outcome of a submission or a review, or anything with legal or similarly significant effects;
- assess people's emotions, reliability or behaviour;
- carry out any practice prohibited by Article 5 of the EU AI Act, or any use the Act classifies as high-risk (such as employment, education scoring, credit or access to essential services).
The descriptions of people the knowledge layer writes describe what published texts say about a person. They are not used to decide anything about that person.
Customers may not use LeraLink's AI features for these purposes.
9. People described by AI can ask for a correction or removal
A person described by an AI-written text can ask, from their account on that installation's site, for it to be corrected or removed. The installation's team decides, and the person sees the answer in their account. A correction is made by a person and is kept beside the machine-made text; no automatic process writes over it. A removal takes the text away in every language and leaves a mark that every automatic process honours. A concept the AI has named is not treated as the installation's terminology for translation until a person confirms it. See Your Data Rights.
10. AI literacy
The EU AI Act asks providers and deployers of AI systems to take measures for the AI literacy of the people who operate and use those systems on their behalf (Article 4, as amended by Regulation (EU) 2026/1744).
- The operator. SSStudio, which builds and runs LeraLink, maintains working knowledge of the models used, their limits and their risks, and records the reasons for each design decision about AI.
- Installations. We give installation administrators and authors guidance on what each AI feature does, its limits, how to review a proposal and how to recognise machine-made content: in How LeraLink Uses AI and in the product's labels. Installations that use AI features are deployers in their own right and are responsible for the AI literacy of their own staff.
11. Reviewing providers and models
Before a provider or a model is added to a default chain, we review:
- where it is hosted, and where its provider is based;
- whether it trains on inputs, and how long it keeps them;
- its terms of service and data processing terms;
- whether it is suitable for the task, and its quality for the languages LeraLink supports;
- whether it has published weights.
The facts are recorded for the model and are shown to installation administrators. A new subprocessor is announced in advance under the Change Notification Policy. Other providers can be added by the platform administrator; none other than those named above are in the default chains.
12. Raising a concern
If you think an AI feature on LeraLink is wrong, harmful, unfair or not as described here:
| Concern | Contact |
|---|---|
| An AI-written text about you, or other personal data | privacy@lera.link |
| A legal question, or a concern about our compliance with the EU AI Act | legal@lera.link |
| Content you believe is illegal | Report Content, or legal@lera.link |
We answer within one month, and we publish a change to this page when a concern leads to one.