AI Transparency Notice

1. Purpose

1.1 This Notice is our disclosure under Article 50 of the EU AI Act (Reg. (EU) 2024/1689), which has applied since 2 August 2026. It tells you where Lemma uses artificial intelligence, what each AI feature does, which provider's model sits behind it, and what it cannot be trusted to do.

1.2 It is written to be read, not to be skimmed past. If anything here is unclear, ask support@getlemma.io.

1.3 The Skill Credibility Index is dealt with separately in the SCI Methodology & Rights Notice, because it engages GDPR Art. 22 as well as the AI Act.

---

2. Every AI feature on Lemma

2.1 AI Tutor

2.1.1 What it does: holds an interactive practice conversation with you in a skill, generates and marks practice exercises, and explains mistakes. Passed exercises count as evidence in the SCI (up to 5 points; SCI Methodology & Rights Notice, section 4.1.2). 2.1.2 Providers: OpenAI (GPT-4o mini for conversation, GPT-4o for exercise generation), with Anthropic (Claude) as a fallback if OpenAI is unavailable. 2.1.3 It is an AI system. You are not talking to a person, and no human reads your AI-Tutor conversations in the ordinary course. 2.1.4 Memory: the Tutor keeps a short AI-generated summary of your progress (what you have covered, where you struggled) and uses it in later conversations. You can ask us to delete it at privacy@getlemma.io. 2.1.5 Limitations: it can state things that are wrong with complete confidence, misjudge your level, and give outdated information. It is not a teacher.

2.2 Adaptive assessments

2.2.1 What it does: generates the questions for each assessment attempt as you take it, at a difficulty chosen from your previous answers, and scores some answers (for example, an explanation of a code fix) against a rubric. 2.2.2 Providers: OpenAI (GPT-4o for questions, GPT-4o mini for scoring), with Anthropic (Claude) as a fallback for both if OpenAI is unavailable. 2.2.3 Limitations: generated questions can be ambiguous, mis-levelled or occasionally wrong, and AI scoring can misjudge an answer. Questions are not reviewed by a person before you see them. If you think a question or a score is wrong, tell us at support@getlemma.io: a person reviews it, and if we agree we correct the result and recompute your SCI.

2.3 Report cards

2.3.1 What it does: once a month, writes a short summary of your progress — strengths, weak points, what to work on next — from your activity data. 2.3.2 Provider: Anthropic (Claude Haiku). If it is unavailable, you receive a standard text without an AI-written paragraph. 2.3.3 Limitations: it summarises the data it is given. If that data is thin, the summary will be too. A report card is a prompt for reflection, not a verdict, and it does not affect your SCI.

2.4 Website assistant

2.4.1 What it does: the "Ask Lemma" assistant on getlemma.io answers questions about Lemma from our public pages. 2.4.2 Provider: OpenAI (GPT-5 mini), reached through Vercel's AI Gateway. 2.4.3 Your data: we do not store the conversation; it stays in your browser tab until you close it (Privacy Policy, row 2.20A). 2.4.4 Limitations: it can misstate our products or prices. Our Terms and policies, not the assistant, are what bind us.

2.5 What we do not use AI for

2.5.1 Moderation. We do not use automated classifiers or AI models to moderate content. Reports are reviewed by people (Notice & Action Policy). 2.5.2 Matching and ranking. Partner suggestions are ranked by rules we wrote, with fixed weights, plus a statistical comparison with similar Users and a small random element so that newer Users are seen. We test alternative weightings on groups of Users. No machine-learning model is trained or used for ranking today. The main parameters, including the paid-plan boost, are in the Terms of Service, clause 5A. 2.5.3 The SCI. The score is computed by fixed rules from Evidence Events (SCI Methodology & Rights Notice). Some of those Evidence Events come from the AI features above.

2.6 Our position under the AI Act

2.6.1 Art. 50 (transparency) applies to the AI Tutor and the website assistant now (section 3). 2.6.2 High-risk systems. Our assessments are generated and partly scored by AI and, together with the SCI they feed, are offered to employers through the Assess and Verification APIs to evaluate candidates. We therefore plan on the basis that we are the provider of a high-risk AI system under Annex III point 4(a) (recruitment and selection), and possibly point 3(b) (evaluating learning outcomes). Those obligations apply from 2 December 2027, following the Digital Omnibus (Reg. (EU) 2026/1744). Our roadmap is in section 12 of the SCI Methodology & Rights Notice.

---

3. Article 50(1) — you are told you are dealing with an AI

3.1 Art. 50(1) requires that a system interacting with a natural person is designed so the person is informed that they are interacting with an AI system, unless it is obvious.

3.2 We do this at or before first interaction, in the interface, not only in this document.

3.2.1 AI Tutor. Directly under the box where you type your first message, the Tutor states: "You are talking to an AI tutor. It can make mistakes, so check anything important." It is shown every time you start a conversation, is not hidden behind a tooltip and cannot be dismissed.

3.2.2 Website assistant. The assistant introduces itself as "Lemma's AI assistant", its header reads "AI · answers from Lemma's own pages", and under the message box it states: "AI answers can be wrong. Don't share personal data. Chats aren't stored."

3.3 Every Tutor exercise and assessment question also carries a visible "AI-generated" label (section 4.4).

---

4. Article 50(2) — marking AI-generated content

4.1 Art. 50(2) requires providers of systems generating synthetic content to mark outputs as artificially generated or manipulated, in a machine-readable format.

4.2 We treat AI-generated report cards, assessment questions and Tutor exercises as substantive generated content within Art. 50(2). We do not rely on the "assistive editing" carve-out for them, because the machine is doing the writing. Systems already on the market before 2 August 2026 have until 2 December 2026 to comply with Art. 50(2) (Reg. (EU) 2026/1744).

4.3 An honest statement about the state of the art. Robust machine-readable marking of plain text is an unsolved industry problem. Watermarking that survives copying, editing and re-formatting works reasonably well for images and audio; for short text it does not yet work reliably, and no standard has settled. We will not claim a capability that does not exist.

4.4 What we do instead, as our current best effort: 4.4.1 a visible "AI-generated" label on every Tutor exercise and assessment question, and on the AI-written paragraph of the monthly report card (shown only when an AI model actually wrote it); 4.4.2 a durable internal record of which model produced which output and when, so any item can be traced after the fact. Today we record when each item was generated but not yet which model produced it; we will add that record before 2 December 2026.

4.5 What we will adopt. We will implement the marking techniques set out in the Code of Practice being prepared under Art. 50(7) once it is finalised by the AI Office, and will update this Notice when we do.

---

5. AI output can be wrong

5.1 Every AI feature here can produce confident, fluent, incorrect output.

5.2 Nothing produced by an AI feature on Lemma is professional advice. Do not rely on it for legal, medical, financial, safety-critical or engineering decisions. Check anything that matters against a qualified source.

5.3 If an AI feature produces something wrong, harmful or offensive, report it at support@getlemma.io or through the in-product flag. This does not limit your rights against us for a service that is not in conformity (Directive (EU) 2019/770).

---

6. Human oversight

6.1 No AI system on Lemma makes a final decision that materially affects you without a human being able to intervene.

6.2 Specifically: account restrictions, suspensions and terminations are decided by a person; a disputed assessment question or score is reviewed by a person; and the SCI is subject to human review and appeal under section 9 of the SCI Methodology & Rights Notice.

6.3 We do not use AI to infer emotional state, and we do not process Practitioners' emotional state, the content of their private conversations, or predictions about the exercise of their fundamental rights (Platform Work Directive (EU) 2024/2831; AI Act Art. 5(1)(f)).

---

7. Your data and model training

7.1 We do not train AI models on your content. We do not train or fine-tune any model of our own.

7.2 Our model providers, OpenAI and Anthropic, process AI-feature inputs as our processors under GDPR Art. 28. Under their API terms they do not use our data to train their models, and they may keep it for a limited period for abuse monitoring. Transfers outside the EEA are covered as described in the Privacy Policy.

---

8. AI literacy (Art. 4)

8.1 AI Act Art. 4, as amended by Reg. (EU) 2026/1744, requires us to take measures to support the development of AI literacy among our staff and others operating AI systems on our behalf.

8.2 We do this by: training everyone who deploys, configures or reviews an AI feature on what these systems can and cannot do and on their failure modes; requiring anyone reviewing AI output in moderation or assessment to be trained on automation bias; keeping a record of who has been trained and when; and publishing plain-language explanations — this Notice and the SCI Methodology & Rights Notice — so that Users can build the same understanding.

---

9. Changes and contact

9.1 If we add an AI feature, change a model provider or change what is done with your data, we update this Notice and tell affected Users before a material change takes effect.

9.2 Questions: support@getlemma.io. Data protection: privacy@getlemma.io (our privacy lead).