AI Disclosure

1. What this page is for

1.1. books.dev is operated by NewGen Books LLC, a limited liability company of the State of Wyoming, United States, at 30 N Gould St Ste N, Sheridan, WY 82801, USA. Much of what this service does for you is done by machine: the adaptations, the dictionary entries, the explanations, the translations, the study hints and the voice that reads to you.

1.2. Article 50 of the EU AI Act (Regulation (EU) 2024/1689) has applied since 2 August 2026. It requires two things that matter here: that you be told when you are dealing with an AI system, and that machine-generated text and audio be marked so that both a person and a program can tell. This page is that disclosure.

1.3. We have written it to be read, not to satisfy a checkbox. It says where the machine is, what it gets wrong, which companies see your text, and what they may do with it. Some of it is uncomfortable to publish; publishing it is better than letting you find out later.

1.4. This page applies to every reader, wherever you are. The Regulation reaches a company outside the Union whose AI output is used inside it, and we saw no reason to write a quieter version for everyone else.

1.5. The /terms and the /privacy are the binding documents; this page explains them and never narrows them.

2. Where AI is used

2.1. Each feature below is produced by a language model or a speech synthesis service, not by a person.

2.2. Adaptations, their notes, book pages and catalogue covers are made in advance, and so is the narration of some catalogue books. Dictionary entries are partly prepared in advance and partly written the first time anyone taps a word. Everything else is produced at the moment you ask for it. Results are kept in a shared cache, so the next reader who asks the same thing gets the same answer.

2.3. Some things here are not AI: the catalogue is selected and checked by hand against the "life plus seventy years" rule in the /terms, the originals are published as their sources give them, word transcriptions come from published dictionaries, and reader reviews are written by readers.

2.4. Dictionary entries, explanations, study hints and sentence translations are machine-generated. An entry starts from open data sources credited at /credits, but much of what you see in it — translations, definitions, examples — is written by a language model and checked only by other models. No person proofreads them, and they contain mistakes, including false friends that read like facts.

3. You are told when the answer comes from a machine

3.1. Article 50(1) requires that a person interacting with an AI system be informed of it, and Article 50(5) requires that this be clear at the first interaction rather than buried afterwards.

3.2. For the adaptations, the notice is in two places: the /terms, and the end of each adapted book in the reader. Dictionary entries, explanations, study hints and sentence translations carry no separate label on screen; this page is where we say that they are machine-generated. Narration is synthetic speech, not a recording of a person, and the player does not label it separately.

3.3. Nothing here is dressed up as a conversation with a human being, and no support agent is secretly a model: write to support@books.dev and a person reads it and answers.

4. How machine-written text is marked

4.1. An adapted book ends with a notice in the reader: the text was written by a language model from the original, and no human has proofread it. The /terms say the same. The original is always available beside the adaptation.

4.2. Authorship is credited honestly. The author of the original work is credited as the author of the original, never as the author of the adapted text. NewGen Books LLC is named as the publisher of the adaptation, not as its author.

4.3. We do not claim copyright in the machine-generated text. The United States Copyright Office, in Copyright and Artificial Intelligence, Part 2 (January 2025), concluded that prompts do not make a person an author and that purely machine-generated output is not protected. What restricts redistribution of files you get from books.dev is the contract in the /terms, not a copyright claim over the adaptation text.

4.4. Article 50(2) also requires machine-readable marking. We do not yet add a dedicated machine-readable marker to the text of an adaptation or to a file you download; the marking is the notice in words described in 4.1.

4.5. For narration, the marking is this page. The audio is streamed rather than delivered as a file, and we do not watermark the signal itself. Better to say where the marking stops than to imply more.

5. Nobody proofreads the adaptations

5.1. An adaptation is written by a model and published without a human reading it first: no editor, no proofreader.

5.2. Adaptations can and do distort the original. Detail disappears, voice flattens, irony is a frequent casualty, a sentence comes out meaning something the author did not write, and names, dates and facts can be wrong.

5.3. So: do not use an adaptation where the text has to be right. Not for quotation, not for studying the work itself, not for translation, not for scholarship. Read the original, which sits beside every adaptation at the same price.

5.4. This is not a fault in one edition that we will eventually fix; it is a property of the method, and section 9 says more about it.

6. Which companies see the text, and what they may do with it

6.1. We run no models of our own. Every model and speech service named here belongs to an outside company, and the text it works on is sent to that company from our server. Our requests carry no name, email address or account identifier, and the vendor sees our server rather than your device. Words and sentences from public-domain books and from books you upload travel the same way.

6.2. Words and sentences. When you tap a word the dictionary does not have yet, the word and up to two words on each side go to a language model. An explanation or a phrase sends the sentence you selected, up to 1,500 characters; study hints and memory aids send the word and up to 300 characters around it. Requests go down a chain until one link answers: DeepSeek; then OpenAI's GPT-6 Luna, reached through OpenRouter; then Groq, on its free tier; then a GLM model on GonkaRouter, a decentralised network in which independent nodes serve the requests. The order of the chain can change.

6.3. Checking the answers. A freshly written dictionary entry, and the sentence (up to 600 characters) when the service works out which meaning of a word fits it, go to two verification models: Jev, made by TypeSafe and reached through OpenRouter, and Clef-flash, run by Cloudflare on Workers AI under its free plan.

6.4. Translation and narration. Sentence translation is done by DeepL through the DeepL API Free plan, which receives the selected sentence. Narration is generated by Microsoft Azure Speech in the Germany West Central region, which receives the passage being read or the word you tapped. Narration recorded in advance for some catalogue books is made on our own computers with open-source speech models and involves no outside company.

6.5. Smaller calls. A text built from your words sends the words and meanings you saved to DeepSeek. To make a cover for a book you upload, its title, author and first 600 or so characters go to DeepSeek, which describes an image; that description, not your text, goes to the image service Higgsfield. To find an icon for a word, the English word goes to the public Iconify icon search. Adaptations are written in advance, from public-domain text only, and contain nothing of yours.

6.6. What they may do with it. We do not promise that these companies leave the text alone. Groq's free tier, DeepL API Free and Cloudflare's free Workers AI plan are free tiers, and GonkaRouter is a decentralised network; their terms may allow them to keep the text they receive or use it, including to improve their models. The other vendors handle it under their own terms, which we have not negotiated. Words and sentences from public-domain books, and from books you upload, are sent to all of them as described above.

6.7. Your uploaded books leave the building too. There is no offline mode. When you tap a word, translate a sentence or listen to a book you uploaded, the same fragments go to the same companies as for a catalogue book, and listening sends the text of the chapter to Azure. Answers are kept in a shared cache that is not tied to your account, and a translated sentence is kept there together with its text. If you have a text that must never leave your own computer, do not upload it here.

6.8. The current named list of vendors, with their roles and locations, is kept up to date at /subprocessors; the legal bases and transfer mechanisms are in the /privacy.

7. What this service does not do

7.1. No emotion recognition. We do not infer your mood, engagement or frustration from anything — not from your reading speed, not from a camera or a microphone.

7.2. No biometric categorisation or identification. We collect no faces, voices or other biometric data, so there is nothing to categorise.

7.3. No deepfakes. Narration uses synthetic voices: the standard voices of Microsoft Azure Speech and, for narration recorded in advance, voices of open-source speech models. None is a clone of a real person, living or dead, and we generate no images, video or audio depicting one.

7.4. No automated decision-making with legal or similarly significant effects on you. Suspensions, refunds, takedowns and anything else that changes your standing with the service are decided by a person; no model decides what you may read or what you are charged.

7.5. We do not train models of our own, on your data or anyone else's.

7.6. Some of your reading does reach analytics: the word you tap or save, and which book you open, start and finish. Google Analytics receives this from your first page view, PostHog only if you allowed it. The text of a book never does. See the /cookies.

8. When the machine gets it wrong

8.1. Tell us. Write to support@books.dev with the book, the edition and level, and the passage — one sentence is enough to find it. Report a wrong word explanation the same way, with the word and the sentence it appeared in.

8.2. We read the passage against the original and answer within ten business days. If the report is right, we correct the passage, regenerate the edition, or withdraw it.

8.3. Withdrawal means the edition stops being available to read here. The original stays in the catalogue, and so does your progress on the work; a file you downloaded earlier stays on your device and we cannot recall it. If the same failure shows up across a level or a language, we withdraw the whole set rather than wait for reports book by book.

8.4. Claims that a book or an adaptation infringes someone's rights follow a different route: copyright@books.dev and the procedure at /dmca.

9. The limits of adaptation as a method

9.1. A simplified adaptation is a model rewriting an entire novel to a target level without understanding the book the way a reader does. It is very good at one thing: producing a large amount of text you can actually read in a language you are still learning. Volume of comprehensible reading is what moves a learner, and this is the cheapest way to produce it.

9.2. It is bad at knowing what matters: a model cannot tell which sentence carries the book and which is scenery, so it simplifies both at the same rate. Style is the first thing lost, because style is mostly the difficulty.

9.3. The most dangerous property is fluency: machine text reads smoothly whether or not it is faithful, so an error does not feel like an error, and there is no seam where the adaptation stopped following the author.

9.4. What we do check is limited, and machines do all of it. The level of every adaptation is measured from the text itself by a program; How we measure reading levels explains the method and how it compares with exam texts. Editions made since October 2026 are also compared with the original paragraph by paragraph by AI reviewers from more than one model family, and automatic checks confirm that no paragraph, line of dialogue or name has gone missing. These checks catch many errors and miss some. There is no single score of fidelity, and editions made before October 2026 were not checked against the original this way.

9.5. Read adaptations as reading practice. The original is the edition of record, and it is always beside the adaptation.

10. Changes and contact

10.1. We update this page whenever a feature or a vendor changes; changes that affect you are announced 30 days in advance by email, as the /terms require.

10.2. Questions about anything here: support@books.dev. About your personal data: privacy@books.dev. Legal notices: legal@books.dev, or NewGen Books LLC, 30 N Gould St Ste N, Sheridan, WY 82801, USA.

Effective 17 October 2026