--- title: EU AI Act statement lang: en path: /legal/ai-act alternate_de: /de/rechtliches/ki-verordnung updated: 2026-10-23 version: 1.2 --- *Since 9 October 2026, System1 Models operates under the name Decision Models (decisionmodels.io). References to System1 Models and system1models.ai in this document mean Decision Models and decisionmodels.io; nothing else changes.* # EU AI Act statement Version 1.2 ยท Last updated: 23 October 2026 This page explains how System1 Models relates to the EU Artificial Intelligence Act (Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744) and what that means for you as a customer. It is not legal advice. It is our own assessment of our role and duties. It is not a certificate: no conformity assessment or official certification exists for a service like ours, and we do not claim one. ## 1. What the Service does The Decision Models API takes a situation and a set of answer options that you supply, and returns a probability for each option. It does not chat with people, and it does not generate free text, images, audio or video. You decide how its output is used in your own product. ## 2. The models and our role - **Most models are open-weight models by outside developers, served as published.** The current lineup, with authors, revisions and licences, is on [Model licences](/legal/model-licences). We do not change the weights of these models, and we never use your data for training. - **REAX models are our own fine-tunes.** REAX 1 (`reax-1`) is a fine-tune by Decision Models of Perplexity Decider v1.2 (Perplexity, Apache-2.0), which is derived from Qwen3.8-27B (Qwen team, Apache-2.0). We publish its weights under the Apache License 2.0, with a model card and a list of modifications. By our estimate, the fine-tuning used a small fraction of the compute used to train the original model, far below the indicative criterion in the European Commission's guidelines on general-purpose AI models (more than one third of the original training compute). Under those guidelines, the developers of the original model remain the providers of the general-purpose AI model; we are a downstream modifier. We document REAX training data, intended use and limits voluntarily on the model card. If a future REAX model or a model trained in our network crosses that criterion, we will reassess our role first and update this page. - **In our assessment, we are not the provider of a general-purpose AI model** for any model we serve today. The obligations for general-purpose AI models (Articles 53 and 54, in force since 2 August 2025) rest with the developers who trained and released them. Documentation and training-data information for the underlying models come from their developers' model cards; links are on [Model licences](/legal/model-licences). - **We provide the Decision Models API service, an AI system offered under our own name.** We are responsible for the service we operate: its intended purpose, its limits, the information on this page, and how we handle misuse. The Service is not a high-risk AI system and is not intended for high-risk purposes (section 3). - **You are the deployer, or the provider of your own AI system.** When you build the API into your product, you decide the purpose and the context. Your duties follow from that use (section 4). ## 3. Obligations that apply now, and what we do | AI Act rule | Applies since | What we do | |---|---|---| | Prohibited AI practices (Art. 5) | 2 February 2025 | Our [Acceptable Use Policy](/legal/acceptable-use) forbids every prohibited practice. It binds you and your end users. We suspend keys when we find misuse. | | AI literacy (Art. 4, as amended) | 2 February 2025 | We take measures to support the AI literacy of the people who operate the Service on our behalf. This page and our documentation state what the models can and cannot do, to support your team's AI literacy. | | Transparency for certain AI systems (Art. 50) | 2 August 2026 | Art. 50 covers AI systems that interact directly with natural persons or generate synthetic audio, image, video or text content. The API as delivered does neither: it returns probabilities for options you supply. If your product interacts with people or generates content, you must meet Art. 50 for your product (section 4). | | High-risk AI systems (Art. 6 with Annex III, or Annex I) | 2 December 2027 (Annex III) and 2 August 2028 (Annex I), as amended | The Service is not intended for any high-risk purpose. Our Acceptable Use Policy forbids decisions with legal or similarly significant effects on people based solely on the Service's output without meaningful human review. | ## 4. What you must do - Use the Service only within our [Acceptable Use Policy](/legal/acceptable-use) and pass its rules on to your end users. - If your product interacts with people, tell them they are dealing with an AI system where Art. 50(1) requires it. If your product generates or manipulates content, meet the marking and disclosure duties of Art. 50. - If you use the Service within a high-risk AI system (for example in employment, education, credit or access to essential services), you are the provider or deployer of that system and responsible for its requirements. Tell us before you do so. We will say honestly whether we can support that use and, if we do, agree in writing which information and technical assistance we provide (Art. 25(4)). - Test the models for your use case and monitor their output. Probabilities can be wrong or poorly calibrated for your data, and performance differs by language and domain. - Support the AI literacy of the people in your organisation who use the Service (Art. 4). ## 5. Intended purpose and known limits - **Intended purpose:** fast, low-cost classification and choice among options you supply, as one component in software that you control. - **Not intended for:** being the sole basis of decisions about individuals with legal or similarly significant effects; any practice prohibited by Art. 5; safety-critical control without independent safeguards. - **Known limits:** the models can be confidently wrong, may reflect biases in their training data, and have not been evaluated for every language or domain. Our [benchmarks](/benchmarks) show where they trail other models. ## 6. Changes and contact We update this page when the law, the Commission's guidance or our models change. [Previous versions](/legal/ai-act/archive). Questions and reports of misuse go to [info@productivity-boost.com](mailto:info@productivity-boost.com).