Notes on AI

In short

• KI können uses AI in the app for three tasks: matching your job, evaluating your instructions and, in the AI period, supporting you in your own projects (project help) (built, not yet in operation). The app does not build tools for you; you build them in your own AI.
• We have an AI create examples for professions, example tools and the answers in “Before and after” in advance. Nothing about you goes to the AI for this, and a person approves every result before it appears in the app (built, not yet in operation).
• The feedback on the builder, the AI finder and the mini-games work by fixed rules, without AI. (Mini-games: Entwurfsänderung 2026-11-02)
• What you enter into your own AI, for example ChatGPT or Claude, goes to its provider, not to us.
• The AI comes from Anthropic. What you enter for these tasks goes to Anthropic and is stored there in the USA, usually for no more than 30 days.
• AI can make mistakes. Check results before you rely on them.
• Do not enter names, contact details, health data or secrets.
• The AI's feedback helps you learn. It is not a grade, not an exam and not advice.
• The weekly podcast and the weekly overview are made with AI, and the podcast's voices are synthetic. Both are always labeled as such (built, not yet in operation).

1. Where the AI works

Matching your job (built, not yet in operation)

• When: when you describe your job in your own words and the app cannot find it with certainty in its list of professions.
• What the AI does: It assigns your entry to a profession and a field. The comparisons and examples in the lessons and, once there is one, the design of your field are based on this.
• What goes to Anthropic: your text about your job, the language and our list of professions. Before sending, we replace any email addresses and phone numbers we detect with placeholders.
• Without AI: If you pick a suggestion from the list while typing, the app classifies your job without AI. The app asks the AI at most three times in 7 days per account; after that it matches by fixed rules.
• New professions: If your profession is missing from the list, we note a general job title as a suggestion for the list. The list does not record who entered it.
• Changes: You can change your profession at any time in the settings.

Evaluating your instructions (built, not yet in operation)

• When: when you have the instruction from a lesson's builder evaluated ("Get feedback"), improve an instruction under "Write it yourself" in the extra round or choose "Have it evaluated" in a project of your own. This is available in the AI period of the course, and with a free account once, up to three times, in the trial lesson; without an account never. In the builder, the app offers the evaluation only if your instruction contains words of your own (a line of your own or a filled-in gap). Having your own projects evaluated is available in the AI period only. (Entwurfsänderung 2026-11-02: "or built", "Build tool" deleted.)
• What the AI does: It compares your instruction with the task's evaluation criteria. You get an evaluation for each criterion, marked passages in your text, short notes and a suggestion for improvement.
• What goes to Anthropic: your instruction, the task with its evaluation criteria and model answer, and a short summary of your profile: profession, field, work situation and experience with AI. For a project of your own, the task is a general text of ours; the name and goal of your project are not sent for the evaluation. Before sending, we replace any email addresses and phone numbers we detect with placeholders.
• What the evaluation means: It evaluates your instruction, not you. The result feeds into your skill bars. It has no other consequences: no grade, no certificate, no decision about price, access or contract. [Prüfen: keine automatisierte Entscheidung nach Art. 22 DSGVO? – Die Bewertung wirkt nur auf die Fortschrittsanzeige]
• Without AI: For every instruction from the builder, you first get feedback by fixed rules: at most three points on what is missing, and what is already good, plus the model answer. Where the AI does not evaluate, for example in the free version after the trial lesson, that is what you get; under "Write it yourself" the checklist helps you (self-check).

Examples for professions (created in advance; built, not yet in operation)

• When: For the professions on our list of professions, we have examples, exercise tasks and comparisons created in advance, once per lesson, profession and language. You see them with the course and in the trial lesson.
• What the AI does: From the lesson and our details on the profession, it writes examples, exercise tasks and comparisons that fit this profession.
• What goes to Anthropic: only the lesson and details from our list of professions, for example the typical tasks of the profession. No details about you and not your pseudonymous identifier: the examples are not created while you learn, but beforehand and the same for everyone.
• Review by a person: A person reviews every example and approves it before it appears in the app. If a lesson changes, the app no longer shows the old example until a new one is approved.
• The same for everyone: Everyone with the same profession and language sees the same examples. An example for your job carries the note "Example for [profession], created by an AI".
• If nothing is ready yet: If there is no approved example for your profession and a lesson, for example because your profession is not on the list, you see the lesson's general example and, where there is one, the comparison for your field. Today this is the case for everyone, because no examples for professions have been approved yet. [TODO Betreiber: den letzten Satz nach der ersten Freigabe streichen]
• Later: project suggestions that build on your goals and your own projects (planned, not yet in operation). These details would then go to the AI as well. We will update this page beforehand.
• [Prüfen: Gehören die vorab erstellten und von einer Person freigegebenen Beispiele zu den KI-Funktionen der App (Art. 50 Abs. 2 KI-VO, maschinenlesbare Markierung) oder zu den Lerninhalten (Abschnitt 2)? – Die App kennzeichnet sie ohnehin sichtbar und im HTML]

Example tools (built in advance; built, not yet in operation)

Entwurfsänderung 2026-11-02 (Umbau): new.

• When: In the trial lesson, an example tool for your idea opens. With the course, the workshop shows the examples for your profession and your field under "Example tools".
• How they are made: An AI (Claude Sonnet 5.5) builds them in advance from instructions that we wrote, once for everyone. No details about you go to the AI for this. Automated checks (nothing loaded from outside, nothing stored in the browser, language, accessibility, no errors on opening) run before approval; then a person reviews and approves every example.
• Labeling: "Example – built by an AI from a similar instruction", with model and date, plus the instruction it was made from. Until the first real examples are approved, the app shows placeholders built by hand with the note "Placeholder example – made by hand until the real examples are ready".
• Security: An example runs in isolation at a separate address, without access to the app, without internet access and without storage in the browser. What you type stays in this window and goes neither to us nor to the AI.
• Rights: The terms and conditions set out how you may use the examples (section 9.4). [Prüfen: Rechte an KI-Ausgaben und Haftung für generierte Werkzeuge in den AGB regeln – Anthropic überlässt die Rechte an den Ausgaben dem Betreiber, nicht den Nutzerinnen und Nutzern; seit dem Umbau betrifft das die vorab gebauten Beispiel-Werkzeuge (AGB 9.4) und Ergebnisse, die Lernende in ihrer eigenen KI erzeugen (AGB 9.3)]

Before and after (created in advance; built, not yet in operation)

Entwurfsänderung 2026-11-02 (Umbau): new.

• What you see: In some lessons, an AI's answers to a weak and to a good instruction, side by side, and the question of what makes the difference. With the course, once there are any, with instructions from your field.
• How they are made: An AI (Claude Haiku 5.5) answers the two instructions in advance like an ordinary AI chat, without knowing which one is the good one. The question with it ("What makes the difference?") is also written in advance by an AI (Claude Haiku 5.5). A person reviews and approves both. No details about you go to the AI.
• Labeling: every answer with "AI-generated answer · model, date". Today there is only one placeholder pair written by hand; it is called "Sample answer, written by hand (placeholder)" and is never shown as an AI answer.

Building tools (switched off for learners)

Entwurfsänderung 2026-11-02 (Umbau): The app no longer builds tools for learners (switch tool_build off). Nothing goes to Anthropic for this, and no tools are created in your workshop. You put the instruction to work in your own AI (below). If tool building is ever switched on again, the following applies again beforehand: your instruction goes to Anthropic exactly as you wrote it. Here we do not replace email addresses and phone numbers. This page will be updated beforehand.

Your own AI (not a function of KI können)

Entwurfsänderung 2026-11-02 (Umbau): new.

• What happens: With "Use it in your AI" you copy your instruction. "Open in ChatGPT" and "Open in Claude" then open the service's start page in a new tab, without writing the instruction into the address and without partner or tracking parameters. "Other or company AI" only copies.
• Who processes it: What you enter there goes from your device to the provider of the service, for example OpenAI or Anthropic, under its terms. We receive none of it. Whether a provider uses your inputs for training depends on your plan and your settings; the lessons show you where to find this.
• Labeling: Answers from your own AI are labeled by its provider, not by us.
• The same rules for everyone: Anthropic, our AI provider, also offers Claude. In guides, tips and the AI finder, the same rules apply to Claude as to all other services.

AI finder (without AI; built, not yet in operation)

Entwurfsänderung 2026-11-02 (Umbau): new.

• The AI finder recommends one or two AI services for what you want to do, by fixed rules, without AI and without rating the quality of answers. The details about the services are shown with their source and an as-of date, and assessments are marked as such; no prices, no logos, no partner links. Your answers are only in the page's address and are not stored.

Project help (built, not yet in operation)

• When: in the AI period, on the page of a project of your own in your workshop, when you send project help a message (switch project_help, off by default). Up to 10 messages a day and 30 a month, in the AI extension 20 (terms, section 4.3).
• What the AI does: It helps you phrase the next instruction yourself, explains why a result may not have worked and suggests test steps. It is not a general-purpose chatbot and does not write the instruction for you: at most two sentences as an example for one passage. The app does not let answers with code, links or a whole instruction through.
• What goes to Anthropic: your message, your last messages and the answers in the project (at most 20 in all), the name and goal of the project, the newest version of your instruction with the summary of its evaluation and the short summary of your profile as for the evaluation; only if the project had a tool built in the app, also its name, the AI's note on it and how the last build ended (since the rebuild, none are created). Before sending, we replace any email addresses and phone numbers we detect with placeholders.
• Notice: Above the conversation the app always says that you are writing with an AI, not with a person. Every answer is labelled "written by an AI". [Prüfen: Reicht der dauerhafte Hinweis über dem Gespräch als Information „zu Beginn der Interaktion“ nach Art. 50 Abs. 1 KI-VO, oder braucht es zusätzlich einen Hinweis vor der ersten Nachricht?]
• Storage at our end: We keep only the last 20 messages of a project and delete older ones as soon as a new answer arrives. With "Delete history" you delete all messages of a project at once; otherwise they stay until your account is deleted. [Prüfen: Speicherdauer der Projekthilfe-Nachrichten – technisch auf die letzten 20 begrenzt, sonst bis zur Löschung durch dich oder des Kontos; reicht das?]

AI news (planned, not yet in operation)

• News: The AI selects reports from news sources. It writes its own headline and a short summary. It also explains what the report means, per field and work situation, not for you personally. Every report names its source with a link and the date.

Weekly podcast (built, not yet in operation)

• Script: Every week the AI (Claude Sonnet 5.5) writes a script from the published news of the week and the skill of the week: news, skill, practice task, outlook. No data about you goes into it.
• Review pass: A second AI pass (Claude Haiku 5.5) checks every statement of the script against the news and the lesson it relies on and marks statements without support. A news item about which the script makes no supported statement is marked too.
• Approval: By default a person reads the script with its marks, can edit it and approves it; if there are marks, the person has to confirm having read them. Only then is the episode voiced and published. Automatic approval can be switched on but is off by default; it applies only when the review pass finds no mark at all. [TODO Betreiber: manuelle oder automatische Freigabe festlegen] [Prüfen: Automatische Veröffentlichung (SPEC §18 Nr. 5) ausschließen? – § 19 MStV verlangt, Nachrichten vor der Verbreitung auf Inhalt, Herkunft und Wahrheit zu prüfen]
• Voices: The voices are synthetic: fixed standard voices from a speech service, no imitations of real people's voices. [TODO Betreiber: TTS-Anbieter]
• Notice at the start of every episode: The app itself inserts the first sentence of every episode, not the AI. After approval by a person it reads: "This episode was made with AI and reviewed by our editors; the voices are synthetic." After automatic approval: "This episode was made with AI and checked automatically against its sources, not by a person; the voices are synthetic." The same sentence is on the episode's page and in the show notes in the feed. The transcript of every episode is on its page. [Prüfen: Reicht dieser Satz am Anfang der Folge als Hinweis auf künstliche Stimmen (Art. 50 Abs. 4 KI-VO, Abgrenzung zum Deepfake, Leitlinien Rn. 114 und die Beispiele dazu)?]
• Sources: The app builds the links to the sources and to the lesson in the show notes from the news and the lesson, not from the AI's text.

Weekly overview (built, not yet in operation)

• An email with three news items of the week from the same selection, the skill of the week and a link to the current episode. Above the news it says: "The news items are summarised by an AI."

2. How our learning content is made

• We created the lessons, questions, model answers, comparisons, the texts of the mini-games and the list of professions with AI, following our content guide and our specifications. (Mini-games: Entwurfsänderung 2026-11-02)
• We also created the facts and assessments in the comparison of AI services (AI finder) with Claude, a model by Anthropic. Anthropic is itself one of the providers compared there. We checked every fact against the provider's official source; the source is shown with it. (Entwurfsänderung 2026-11-02)
• Several AI-assisted review passes checked them for teaching quality, facts and language. [TODO Betreiber: ob und wie Menschen die Inhalte vor der Veröffentlichung prüfen]
• Content on law and data protection is general information, not advice for your individual case.
• [Prüfen: Fallen Lektionen mit Rechts- und Datenschutzthemen unter Art. 50 Abs. 4 UAbs. 2 KI-VO („Angelegenheiten von öffentlichem Interesse“, Leitlinien Rn. 131)? – dann kennzeichnen oder menschlich prüfen; die Offenlegung hier ist freiwillig]

3. Images, videos and voices

• Images: The app shows AI-generated images of made-up people in everyday work scenes. They are not real people, and no user data went into them. (Entwurfsänderung 2026-11-02: previously "nine images".)
  • images on public pages (sign-in pages, fields, questions), created with the image service Higgsfield (model GPT Image 2.5);
  • 123 images on the learning cards of the lessons, created via Higgsfield with the model Seedream (seedream_v5_pro); only approved ones appear in the app. For all 123 image slots (basics and modules S1 to S5 since 2 November 2026, modules S6, P1 and P2 since 3 November 2026), the app shows drawn building-block scenes instead of these photos (see “Drawn lesson images” below); the photos remain only as a fallback in the files and appear, with the label, only if a scene is withdrawn. (Entwurfsänderung 2026-11-03: previously 82 image slots, basics and S1 to S5.) (Entwurfsänderung 2026-11-02: previously the screenshots of the trial lesson on the start page showed one of these photos with the label; now they show the drawn scene without a label.) [Prüfen: Nutzungsrechte an den Bildern und eine etwaige Pflicht zur Namensnennung nach den Bedingungen von Higgsfield – die AGB (Abschnitt 10) beanspruchen die Rechte an den Bildern für uns oder unsere Lizenzgeber]
• Labeling: Each of these images visibly carries the label "AI-generated image". Its alt text starts the same way; only where a picture is pure decoration of a tile (the tiles of the welcome) is the alt text empty, and the tile's name says what it is. (Entwurfsänderung 2026-11-03: exception for decorative pictures added; the app now also puts the label in front of the alt text of lesson photos that step in as a fallback.) The image files also contain a machine-readable statement (IPTC "trainedAlgorithmicMedia").
• Mascot (Entwurfsänderung 2026-11-02): The building-block phoenix on "Today", on the learning path, at the end of a lesson, in the profile, in the trial lesson and (as still images) on the start page and the profession pages is a drawn 3D figure made of rounded building blocks. The app draws it with program code (three.js), transferred from a design that was created with the help of an AI; we render still images from the same code. It does not look like a photo and therefore carries no "AI-generated image" label (decision of 1 November 2026: the label applies to AI images that look realistic or photo-like). It has a descriptive text for screen readers. [Prüfen: Einordnung des Maskottchens – kein Deepfake im Sinne von Art. 3 Nr. 60, Art. 50 Abs. 4 KI-VO, weil es nicht echt wirkt; Urheberrecht an Figur und Code (Schutz nur bei menschlicher Schöpfung) und Markenanmeldung (AGB Abschnitt 10)] [Prüfen: Sind fotorealistische Bilder ausgedachter Personen Deepfakes (Art. 3 Nr. 60, Art. 50 Abs. 4 KI-VO)? – Die Kennzeichnung erfolgt ohnehin]
• Drawn lesson images (Entwurfsänderung 2026-11-02): On the learning cards of all lessons (every image slot: basics and all modules; lessons added since 3 November 2026 only have drawn scenes and no photo; Entwurfsänderung 2026-11-03: previously basics and S1 to S5, 82 images, then 123; also in the trial lesson, on “Today”, when choosing the field in the welcome and in the screenshot of the trial lesson on the start page) the app shows drawn 3D scenes made of rounded building blocks in the style of the mascot, with simple figures without faces. Each scene is set down as a description in data (furniture, objects, figures, camera angle); program code (three.js) builds it from that and renders still images on our own computer. No AI image model creates a picture, and no image service is used; we wrote the program code and the scene descriptions with the help of an AI coding assistant. The scenes do not look like photos and therefore carry no “AI-generated image” label (like the mascot, decision of 1 November 2026); their alt text describes the scene. [Prüfen: Einordnung der gezeichneten Lektionsbilder wie beim Maskottchen – kein KI-Bild und kein Deepfake im Sinne von Art. 3 Nr. 60, Art. 50 Abs. 4 KI-VO, weil ein Programm sie aus Code zeichnet und sie nicht echt wirken; Urheberrecht an Szenen und Code (Schutz nur bei menschlicher Schöpfung)]
• Demo videos (built, not yet in operation): Short videos on the "How it works" page show the real app with made-up example profiles: "The trial lesson" and "The feedback", recorded from the app without the in-app tool building, with captions and a text version. They contain no AI-generated pictures. The AI evaluation in them is a prepared example; the videos carry the note "Example with prepared answers" in the picture throughout. The example tool in the video is a placeholder built by hand and labelled as such. (Entwurfsänderung 2026-11-02: new videos without the switched-off tool building; the old ones are removed.) [Prüfen: Reicht dieser Hinweis, um eine Irreführung nach UWG zu vermeiden?]
• Social media (built, not yet in operation; nothing is published automatically): Clips with AI-generated people and scenes show "AI-generated" for the whole clip, plus a note in the caption and the platform's label. Carousels and slides contain no AI pictures. Carousels about AI news say "Summarised with AI" and only say "checked by an editor" after a person has reviewed them. The clips have no spoken voices yet; section 1 describes the podcast's synthetic voices. [Prüfen: Art. 50 Abs. 4 KI-VO – genügt die Kennzeichnung im Bild, in der Bildunterschrift und über die Plattform für ausgedachte, fotorealistische Personen?] [TODO Betreiber: Musik in Clips stammt aus eigenen Suno-Titeln; nur Titel aus einem bezahlten Suno-Abo verwenden]

4. Provider and models

• AI provider: Anthropic Ireland, Limited, 6th Floor, South Bank House, Barrow Street, Dublin 4, D04 TR29, Ireland. Anthropic processes the data on our behalf. [Prüfen: Vertragspartner laut Commercial Terms und Auftragsverarbeitungsvertrag (DPA vom 24.02.2025) vor dem Start prüfen und abschließen – ohne Vertrag keine Übermittlung]
• Models (as of October 2026; built, not yet in operation):
  • Claude Haiku 5.5: matching jobs, evaluating instructions, project help (since 31 October 2026, previously Sonnet), creating examples for professions and "Before and after" in advance, the review pass of the podcast scripts; later also news (planned, not yet in operation).
  • Claude Sonnet 5.5: the example tools we have built in advance, and podcast scripts. (Entwurfsänderung 2026-11-02: "building tools" for learners dropped; project help now Haiku.)
• Changes: If we switch to a newer Anthropic model, we will name it here. Before we change provider, we will update this page and the privacy policy.
• Other AI services for our own content: Higgsfield for images and video clips (video clips built, not yet in operation), a speech service for the podcast's voices (built, not yet in operation; [TODO Betreiber: TTS-Anbieter]). These services receive no user data.

5. What happens to your data

• What goes to Anthropic: only what the function in question needs (section 1), and a pseudonymous identifier. It lets Anthropic detect misuse but not find out who you are. We do not send your account's email address or your IP address. Whatever you write into a text yourself is sent as described in section 1. The requests go from our server to Anthropic, not from your device.
• What does not go to the AI: your other profile details, such as tasks, time wasters, goals and interests. Your answers to questions with fixed answer options do not go to the AI either.
• Names: The app does not detect names automatically. If they are in your text, they go to the AI with it.
• No training: Anthropic may not use content from our requests to train its models.
• Location: The app itself is to run on a server in the EU, at Hetzner in Helsinki, Finland (planned, not yet in operation). Our contracting party, Anthropic Ireland, is based in the EU, but stores the data in the USA and may process it anywhere in the world. Anthropic needs appropriate safeguards for this onward transfer; the privacy policy (section 7) says which ones. [Prüfen: Transfer-Folgenabschätzung für die Weiterübermittlung durch Anthropic in die USA (Art. 28 Abs. 4, Art. 44 S. 1 DSGVO); Verarbeitung auf „us“ statt „global“ festlegen? – Anthropic ist nicht nach dem EU-US Data Privacy Framework zertifiziert (Liste geprüft am 20.10.2026)]
• How long Anthropic keeps it: Anthropic deletes inputs and outputs within 30 days. Content that Anthropic's safety systems flag as a breach of its usage policy is kept for up to 2 years, and the results of this automated check for up to 7 years. Statutory retention duties take precedence. [Prüfen: Angaben vor Veröffentlichung mit DPA und Privacy Center abgleichen – frühere Aussagen von Anthropic waren widersprüchlich]
• With us: We store your instructions, evaluations, check results and your own projects in your account until it is deleted; of project help only the last 20 messages per project (section 1). Examples created in advance, example tools and "Before and after" are content of the app without any link to a person. (Entwurfsänderung 2026-11-02: tools and builds dropped.)
• More: Legal bases, recipients, retention periods and your rights are in the privacy policy.

6. What you should not enter

What you write into the AI functions goes to Anthropic and stays stored in your account. So do not enter:

• names and contact details of other people, for example customers, patients or colleagues;
• health data and other specially protected information, for example about religion, political opinions or sexual orientation, not even about yourself;
• passwords, login details, keys for programming interfaces (API keys), bank or card details;
• confidential information from your business or your customers. Follow your employer's rules and your agreements with your customers.

You can still practice: use placeholders such as "Customer A" or "Mr. M." and made-up example values.

If the app finds an email address or phone number in your instruction, it points this out. It blocks nothing; you decide.

7. What the AI cannot do

• It can be wrong. Answers can be incorrect, incomplete or made up, even when they sound convincing. Check facts, figures and calculations before you rely on them.
• Tools can contain errors, the example tools just as much as the ones you build in your own AI. A calculator can calculate wrongly, a form can miss something. Test a tool with values whose result you know; "Check the result" at the end of a lesson helps you with this.
• No advice. The AI does not replace legal, tax, financial or medical advice from a professional.
• The evaluation can be off. It is feedback for practice. You decide what you take on board.
• Results can differ. If you enter the same instruction into an AI again, the result may be different from the example.
• Not always available. AI use is limited per day and, in the AI period, also per month (terms, section 4.3). There are further limits to protect against high costs and misuse; in the AI period they do not apply within your quotas (terms, section 4.5). If the AI is not available at the moment, the app tells you what you can still do without AI. Examples and comparisons in the lessons are created in advance and stay. [Prüfen: Kostengrenzen im KI-Zeitraum – die monatliche Grenze (Startwert seit 31.10.2026: 1 €, in der KI-Verlängerung 0,50 €, harte Grenze, ein Aufruf startet nur, wenn sein schlimmster Fall noch darunter passt) liegt über den höchstmöglichen Kosten der Kontingente (seit 01.11.2026 höchstens 0,62 €, mit einer Free-Phase im selben Monat und gleichzeitigen Aufrufen 0,73 €; vorher 40 € gegen 39,60 €), und das gemeinsame Tagesbudget gilt nicht für die bezahlte Nutzung innerhalb der Kontingente (AGB 4.5); reicht das, damit die Grenzen bezahlte Kontingente nicht vorzeitig stoppen? Notschalter und Ausfälle des KI-Anbieters bleiben Fälle von AGB 13]

8. How we label AI

• In the app: At every AI function, the app tells you that an AI is at work, at the latest when you first use it. AI results carry a label. (built, not yet in operation)
• Machine-readable: We mark AI-generated text so that software can recognize it as AI-generated. (built, not yet in operation) [Prüfen: Die App markiert jeden Bereich mit KI-Text im HTML mit data-ai-generated="true"; reicht das?] [Prüfen: Technik der Markierung nach Art. 50 Abs. 2 KI-VO und dem Verhaltenskodex der Kommission vom Juni 2026 – die Werkzeuge sind als Quellcode ausgenommen]
• Example tools: a visible label above the example with model and date (section 1). The HTML file itself carries no machine-readable label; it cannot be downloaded in the app. In our understanding a marking is not required: the tools are source code, and source code is exempt from the duty of machine-readable marking. (built, not yet in operation; Entwurfsänderung 2026-11-02: previously "Downloaded tools" of learners, which no longer exist) [Prüfen: Verlangt Art. 50 Abs. 2 KI-VO für die vorab gebauten HTML-Beispiel-Werkzeuge eine maschinenlesbare Kennzeichnung? – derzeitige Annahme: nein, Quellcode ist ausgenommen (Leitlinien der Kommission, Rn. 68; research.md §12); offen ist, ob Texte, die die KI in ein Werkzeug schreibt (Überschriften, Beispieltexte), als KI-Text gelten. Bis zum 02.11.2026 bezog sich der Punkt auf den Download der Werkzeuge der Lernenden, der entfallen ist]
• "Before and after": "AI-generated answer · model, date" on every answer (section 1).
• Images: visible label, alt text and a statement in the file for AI images that look realistic; the drawn mascot without a label (section 3).
• Podcast: a spoken notice at the start of every episode, the same sentence on the episode's page and in the show notes; it also says whether a person reviewed the episode and that the voices are synthetic (section 1; built, not yet in operation).
• Weekly overview: "The news items are summarised by an AI." (built, not yet in operation).
• News: always labeled as AI-generated (planned, not yet in operation).

9. Your choices and your rights

• You start the AI: Evaluations and project help only run when you trigger them. Matching your job only starts if the app cannot find your profession itself. Examples for professions are created in advance, without details about you (section 1).
• Without AI: The app checks questions with fixed answer options without AI, in reviews too. Your learning path, the rule-based feedback, the selection of examples and the AI finder are also created without AI. The AI only evaluates instructions that you write yourself.
• Disagree? Revise your instruction, change your profession in the settings or write to us at [TODO Betreiber: E-Mail]. Please also tell us if an AI answer is wrong, inappropriate or concerning.
• Your data protection rights: You can request access, rectification, erasure, restriction of processing and data portability. Where we process data on the basis of legitimate interests, you can object. You can complain to a data protection supervisory authority. Details are in the privacy policy. [Prüfen: Rechtsgrundlage je KI-Funktion (Vertrag oder berechtigtes Interesse) – davon hängt ab, ob ein Widerspruch nach Art. 21 DSGVO greift]
• Deleting your account: in the settings under "Data and account" (built, not yet in operation) or by email to [TODO Betreiber: E-Mail]. (Entwurfsänderung 2026-11-02: The delete function has been built since M10; previously this said "still to come".)

Changes

We update this page when functions, providers or models change.

As of: [TODO Betreiber: Datum der Veröffentlichung]