Tinystudy

How Tinystudy uses AI

This page is available in English and in Lithuanian (lietuvių kalba). Both versions are equally binding.

Transparency notice (EU AI Act, Article 50). Version 1.2 · in force from 4 January 2027 (version 1.0 applied from 5 October 2026; 1.1, in force from 4 November 2026, removes the retired "Ask my material" feature; 1.2 adds, for Plus and the trial, the reading of handwriting, drawings and graphs in pictures by AI in the cloud, which starts on that day: until then version 1.1 applies and a picture leaves your device only when you press "Send the picture itself" in Help with a task). Until 3 January 2027 version 1.1 is the text that applies: read version 1.1. If the English and Lithuanian versions differ, the version more favourable to you applies.

What is automated

Without AI (always): reading your courses, deadlines and notifications; noticing new material; extracting text from documents and slides; taking still frames from video lessons; packing study units into tasks for your next test. These are rules, not a model.

Small AI models in your browser (when you press the button)

These read the printed text in pictures, scans and video frames, and printed formulas (Tesseract text recognition and the Pix2Text formula model). They only convert what is already in your material into text; they send nothing to us, and the results can contain recognition mistakes. They cannot read handwriting or say what a drawing, graph or diagram shows; they only notice that a picture seems to hold one. Such a picture stays unread unless the AI in the cloud reads it, as the next section and "Help with a task" describe.

Handwriting, drawings and graphs read in the cloud (when you press the button)

This is part of Plus and of the trial; on the free plan no picture is sent, and pictures are read by the reader in your browser only. When you press the button to make a study page for one test, the reader in your browser goes through that test's own files. A picture, a still frame of a video lesson or a scanned PDF page in which it finds handwriting, or a drawing, graph or diagram (a figure), is sent, one at a time, as a reduced JPEG copy (longest side at most 1,600 pixels) to our service and read by Cloudflare Workers AI (Google Gemma 4 26B); with the picture go only what to read (everything, or only the figures) and the language in which to describe a figure. The model is told to copy the words it can see and to say in words what a figure shows, from what is visible only: not to add facts from outside the picture, not to solve anything and not to translate. The text comes back to your browser and is kept there with your material; the description of a figure is one line that starts with a tag such as "[figure: graph]". A picture is sent only when the reader in your browser judges that it holds handwriting or a figure; it judges by simple rules and can be wrong either way, so it can leave such a picture unread or send one that holds printed text only. This runs only while you are signed in and only for the test you chose, never during the background reading of all your courses, and a daily limit applies: once it is reached, no more pictures are sent that day. The AI's reading can contain mistakes, for example a misread word or a wrong value taken from a graph, so check anything that matters against the original. The picture is not stored and not logged on our side, and the picture reading keeps no text either; only counts are kept (privacy policy, §2). The text read from a picture becomes part of your material's text, so when you make study content it is sent with that text, as the section on study content below describes. We do not use the pictures to identify anyone, and they are not used to train models.

Video lessons turned into text, in the cloud (when you press the button)

Your browser sends the audio of a lesson without captions in 60-second pieces, and Cloudflare Workers AI (the Whisper large-v3-turbo speech model) returns text that can contain mistakes, especially with names, numbers and noisy recordings. Audio is not stored or logged.

Study content from your material, in the cloud (when you press the button)

The material for one test is turned into study content: the ranked "likely questions", explanations, flashcards, quizzes and worked examples. The model is chosen per language, by our own test of the page text, from the open-weight models Cloudflare Workers AI runs: Google Gemma 4 26B for Lithuanian and Spanish, Mistral Small 3.1 24B for German and Portuguese, and Meta Llama 3.3 70B for every other language. The trial (7 days of Plus) and the paid plans may use it for any test; the free plan, which gives 1 study page a day, for the nearest test only. Help with tasks, essay feedback and lesson audio are part of Plus and of the trial, not of the free plan. So is the reading of handwriting, drawings and graphs by AI (see "Handwriting, drawings and graphs read in the cloud" above). Without this, you get a baseline built from the material's own headings and definitions, with no AI.

Writing feedback on your essay, in the cloud (when you press the button)

The text you typed (at most 6,000 characters) and the task you set go to Cloudflare Workers AI, an open-weight model run by Cloudflare as our processor (chosen per language by our own test: Mistral Small 3.1 24B for English, Spanish, French and Italian, Meta Llama 3.3 70B for the other languages), which returns scores and advice. The model is told not to rewrite your text, and its advice can be wrong or miss things. Your essay is not stored and not logged, and nothing else from your material is sent with it.

Help with a task, in the cloud (when you press the button)

Add a photo, a file (PDF, Word, PowerPoint and similar) or type a task from any subject. The photo or file is first read in your browser, and nothing is sent until you press a button. Only the text of the task or study text, as you confirmed or corrected it, together with the short line saying what you need and the name of your course, goes to Cloudflare Workers AI, which explains or solves it step by step; for a maths task it solves it a second time to check the answer. A picture itself is sent only in two cases, each on your own press. When the reader in your browser finds handwriting, the task box stays empty and you can press "Read the handwriting": then that one picture goes to Cloudflare Workers AI (Google Gemma 4 26B), which copies the words it can see and says in words what a figure shows, and the text is put in the task box for you to check and correct before you press "Get help". For a picture that cannot be read as text, for example a diagram, you can press "Send the picture itself": then that one picture goes to Cloudflare Workers AI (Google Gemma 4 26B), only for that one answer. Without one of these two presses, a picture is never sent. The answers and the text read from handwriting can still be wrong, so check them against your notes. The task, the text, the picture, your line and the answer are not stored and not logged; only daily counts are kept.

How AI results are marked

Every result that one of the AI features above produces for you (study content, help with a task, writing feedback and transcribed text) carries a machine-readable mark. In the data we return and store, the result has a field "ai" with generated: true, the system name Tinystudy, the model that wrote it and the time. The mark is kept in your stored study pages and in your data export. The containers in the app that show an AI result are marked data-ai-generated="true", and every study page carries the tag <meta name="ai-generated" content="true">. The visible labels (for example "AI-made, not verified") stay. Text that the AI reads from a picture (handwriting it copies, a figure it describes) is marked as well: the answer of the picture reading carries the same field "ai", and the description of a figure always starts with the tag "[figure: ...]", which is part of the text itself. Such text can contain reading mistakes, and the description of a figure is the model's own wording. For a course file it is kept with your material in your browser, where the record of the file notes that the AI read it, which model, and when. In Help with a task it is put in the task box for you to check and correct before you press "Get help"; until you change it, the task box carries the same machine-readable mark as the results. Study content that is later made from that material carries the full mark. We are the provider of these Tinystudy AI features and the models are run for us by Cloudflare; the way such marks should be written is still being settled for text, so we may change the format and will keep this section current.

What AI-generated content means

Ranked probabilities are the model's estimates from your material — not a measured frequency and not your teacher's plan. Explanations can be wrong, incomplete or worded differently from your teacher's key; the optional fact-checker (off by default) may replace a statement it judges certainly wrong or outdated and lists every such change with its basis. Check anything that matters against your teacher's material. Tinystudy does not grade you, does not evaluate your learning outcomes for your school, and does not decide anything about your education; it is a personal planning aid, not a decision system.

Your data and the model

Your material is kept in your own browser. The cloud options send only what you chose: the text of the material for the test you pick and, on Plus and during the trial, those of its pictures in which the reader in your browser finds handwriting or a figure (study content and picture reading), the audio pieces of a lesson (transcription), the essay you typed (writing feedback), or the text of a task or study text (Help with a task) and, only when you press "Read the handwriting" or "Send the picture itself", that one picture, to Cloudflare Workers AI, our processor, only to produce your result. The material text is deleted from our systems when the job ends and the finished content within seven days; audio, essay text, questions, tasks and pictures are not stored or logged (see the privacy policy, §3). Printed text in pictures and scans and printed formulas are first read by the models in your browser, and that reading sends nothing; the text it yields is part of your material or task and is sent only when you ask for study content or for help with a task, as described above. A picture itself leaves your browser only in the cases named above. Nothing you type into the app or send from your material is used to train models.

Related: Terms · Privacy policy

Vote on what comes next