Version 28 September 2026
You are talking to a machine.
This page belongs to article 50 of the AI Act: you have the right to know that you are communicating with an AI system, and what that system is. We go one step further and also tell you where the system is weak.
Which AI model you are using
Union AI is a service that offers open-weight language models. We do not train our own base model. We run existing models, with publicly available weights and under an open licence, on European infrastructure, with our own system prompt, meter and routing around them.
| Tier | Model | Licence |
|---|---|---|
| Lite | gpt-oss-120b | Apache 2.0 + usage policy |
| Standard | qwen3.6-35b-a3b | Apache 2.0 |
| Long context | deepseek-v4-flash-0731 | MIT |
| Pro | glm-5.2 | MIT |
If you ask in the chat directly which model you are talking to, you get an honest answer. An assistant that denies its own base is an assistant you cannot believe on other points either.
Our system prompt, published in full
This is the core of the instruction that every chat conversation carries. We publish it because the behaviour it enforces is exactly the behaviour we promise.
You are Union AI, the AI assistant of Union AI.
All processing of this conversation happens on infrastructure
inside the European Union.
IDENTITY
Present yourself consistently as Union AI. If someone asks
directly which model you run on, answer honestly which
open-weight model it is and under which licence. Never
invent another origin and do not deny the base.
LANGUAGE
Answer in the language of the user. Never use Chinese,
Japanese or Korean characters unless the user explicitly
asks for them or writes in that language.
CONTENT
Do not refuse legitimate questions about politics, history
or socially sensitive topics. Answer them factually, in a
balanced way, and state where opinions differ.
Do refuse: instructions for weapons, malware, fraud, or
material that harms children. Short, clear, no moralising.Why we sometimes use a different model
Some open-weight models are trained under a legal system with its own sensitivities and duck away from factual questions about politics or history. That is not a safety refusal but a defect, and we treat it as a defect.
If we detect such a refusal within the first sentence of a response, we cancel the response before it reaches you and run the question again on a model that does not share that blind spot. This happens at most once per request, and you can see it in the header X-EU-Routed: censorship and in the chat from a small label next to the response.
We do not do this for genuine safety refusals. Questions about weapons, malware, fraud or material that harms children are refused and stay refused. Routing on until a model complies is an automated jailbreak, and we do not build that, certainly not under our own name.
Web search, memory and the website assistant
Web search. When you turn on Search, our own model writes one short search query from your message, and the results come back as numbered sources: under the answer, and as [1], [2] after each claim that rests on them. The model is told to cite only what is in the results, and to treat web pages as material, never as instructions: anything in a page that tells it what to do is to be ignored. That makes it much harder for a page to steer an answer. It does not make it impossible, so check the sources that matter.
Memory. If you turn on Memory, the short facts it keeps about you are added to later conversations, and the model is told that you can see and edit that list. Memory is never used in a team's shared folders.
The website assistant. A business can put an assistant on its own website. It answers only from what that business gave it: its web pages, documents, and questions and answers. It tells every visitor that they are chatting with an AI assistant and that answers can be wrong, and says how long the conversation is kept. A visitor can ask for a person instead.
Energy. Every chat answer carries an estimate of the energy it took and its CO₂e. How that estimate is worked out, and what it leaves out, is in the methodology on the sustainability page.
What we measure and publish
Every reroute is logged: which trigger, which model before and after, how long the detection took. That lets us state a number instead of a promise: how often a model ducks away, on which topics, and whether it gets better or worse after a model change.
We intend to publish those figures periodically. As soon as the first measurement period is complete, they will appear here, with the measurement method alongside, so you can judge them instead of believing them. Every reroute is visible per request in the response headers, and why we treat a ducked question as a defect rather than a value is argued in European sovereign AI.
Limits of the system and of the models
Language models produce incorrect answers with conviction. Union AI changes nothing about that: our routing repairs one specific fault, ducking away from legitimate questions, and not the general fallibility of the model. Check what you use, especially in legal, medical and financial contexts.
What happens to your data is set out in the privacy statement; who runs which part of the service on the subprocessors page.
Frequently asked questions
What does article 50 of the EU AI Act require?
Article 50 sets transparency duties for AI systems that interact with people. Anyone deploying a chatbot must make sure users know they are dealing with an AI system, told clearly and from the start of the first interaction. It became applicable on 2 August 2026.
What must a chatbot's AI disclosure actually say?
That the person is talking to an AI system, in plain words, before or at the start of the first exchange, and in a form they cannot miss. A line buried in the terms of service does not count, and neither does a label only a developer would recognise, such as a model name or a machine-readable marker.
Which AI model am I talking to?
The model behind every tier is published by name and licence on this page, and the chat answers honestly if you ask it directly. An assistant that denies its own base is an assistant you cannot trust on anything else either.
Do open-weight models avoid politically sensitive questions?
Some do. Models trained under legal systems with their own sensitivities can duck factual questions about politics or history that a European user would consider ordinary. We treat that as a defect rather than a safety feature: when we detect it in the first sentence, we rerun the request once on a model without that blind spot and say so in the X-EU-Routed header.
Why do some models answer worse in Dutch than in English?
Because training data is overwhelmingly English, and tokenisers are tuned for it. Smaller European languages get less training signal and need more tokens per sentence, so answers can be weaker and marginally dearer. It varies per model, which is why the model behind each tier is published rather than hidden.
Can instructions hidden in a web page change the chat's answer?
They should not. Web search results reach the model as numbered sources, with the rule to treat them as material and to ignore any instructions inside them. Every claim that rests on a page carries its number, so you can check it. No model follows such a rule perfectly, which is why the sources are always shown.
Why does an answer sometimes come from a different model?
Two reasons: the primary model failed, or it refused a legitimate question in a way we classify as a defect. We then rerun the request once on another model and tell you in the X-EU-Routed header. We never do this for genuine safety refusals, because routing until a model complies is an automated jailbreak.