WoskWosk — AI Disclosure
Version 1.0 · Effective 26 July 2026
Published at https://wosk.web.app/ai. Linked from the Privacy Policy and from the disclaimer in every result footer.
This document exists because of Regulation (EU) 2024/1689 (the EU AI Act), whose transparency obligations in Article 50 apply from 2 August 2026 — the month Wosk ships. It is written to be read by a user, and to be defensible to a regulator.
1. Everything Wosk tells you about a conversation is generated by AI
There is no editorial team, no psychologist, and no human reading your messages. When you upload screenshots, two AI model calls run on Google Cloud Vertex AI in the Netherlands:
- Extraction. A vision model reads the screenshots and transcribes the messages — who said what, in what order, at what time. It also redacts phone numbers, emails, addresses and account numbers.
- Analysis. A text model reads that transcript and writes the interpretation.
Both models are Google Gemini models. We supply the instructions and the schema; the words are the model's.
2. What is machine-generated, exactly
Generated by AI — every time, without exception:
| In the app | What it is |
|---|---|
| The transcript of your conversation | A machine transcription of your screenshots. It can misread text and can attribute a message to the wrong person. |
| The verdict label and its confidence percentage | A classification into one of our fixed categories, plus a number the model produced. The percentage is the model's own confidence, not a measured probability of anything happening in your life. |
| The one-line headline and the summary | Written by the model. |
| The interest score, 0–100 | Produced by the model. It is not calibrated against any dataset, any research instrument or any outcome. |
| The effort balance ("him 28% / you 72%") | The model's read of the conversation, not a count. |
| Every "what he meant" line in the decoder | Interpretation, message by message. |
| Red flags and green flags, and their severity | Selected by the model from our fixed taxonomy, with the model's own explanation. |
| Attachment styles, his and yours | The model's guess, from a small fixed list. This is not a psychological assessment — see §5. |
| Next moves, and the copyable script | Written by the model. |
| Reply drafts in the reply generator | Written by the model. |
| Answers in the follow-up chat | Written by the model. The follow-up chat is a conversation with an AI, not with a person. |
| The safety classification that switches a result into serious mode | Produced by the model, from a fixed list of signals. |
Not generated by AI:
- Category names and their descriptions — "Situationship", "Breadcrumbing", "Anxious", "He's avoidant" and the rest. Those are a fixed taxonomy we wrote and translated; the model only picks from it.
- All interface text, onboarding copy, paywall copy and error messages.
- The crisis helpline numbers, names, hours and descriptions in serious mode. Every one of those was checked by hand against the operating organisation's own page, with the source and the date recorded in the source file, and none of it passes through a model. This is deliberate: a hallucinated crisis number is the worst thing this app could ship.
3. How it is labelled inside the app
The disclosure is not buried in a settings screen. It appears at each of these points:
| Where | What it says | Who sees it |
|---|---|---|
| Consent screen, before your first upload | that the conversation is processed by AI, that the image is deleted immediately, that nothing trains a model | every user, once, before any upload |
| Analysing screen | the progress copy names the model stages | every user, every analysis |
| Footer of every result | "This is an AI interpretation, not a diagnosis or psychological advice." | every user, every result, free and paid — the disclaimer sits outside the paywall blur, so a user who never subscribes still sees it |
| Next to the disclaimer, on every result | a Report this analysis link, always available, never gated | every user, every result |
| Follow-up chat | presented as asking Wosk, the AI, not a person | Pro users |
| Store listing, in the description footer | "Wosk is an entertainment app. Analyses are generated by AI and are not psychological, therapeutic or legal advice." | before install |
| This page, linked from the policy | the full detail | anyone |
Exported share cards. A card you export carries the verdict, the number, and the wosk.app wordmark. That wordmark is the visible provenance mark: anyone who sees the card can find this page. Name-blurring is on by default and you have to turn it off deliberately.
4. Our position under the EU AI Act
This is our own assessment. It is written down so it can be checked, and it is the thing to hand a reviewer or a regulator who asks.
Our role. Google is the provider of the underlying general-purpose AI models. We build a system on top of them and put it on the market under our own name, which makes us a provider of the Wosk AI system as well as a deployer.
Not a prohibited practice (Art. 5). Article 5(1)(f) prohibits inferring emotions in the workplace or in education. Wosk is neither. Article 5(1)(a)–(b) prohibit manipulative or exploitative techniques that cause significant harm; Wosk's persuasion is ordinary marketing of a paid feature, and the safety gate exists precisely to stop the entertainment framing being applied to a conversation where someone is at risk.
Not an emotion recognition system as the Act defines it. Article 3(39) ties that definition to inferring emotions from biometric data. Wosk reads typed text that the user supplies. It processes no biometric data at all — no faces, no voices, no physiological signals. The transparency duty in Article 50(3) therefore does not attach. We state this precisely because the surface reading of what the app does invites the opposite conclusion, and we would rather answer the question before it is asked.
Not high-risk (Annex III). Wosk is not used for biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, or the administration of justice. It is not a safety component of anything.
Article 50(1) — interacting with a natural person. The follow-up chat is a direct interaction with an AI system, and it is labelled as such. In practice the whole app is labelled as such, from the consent screen onward.
Article 50(2) — marking synthetic output. Wosk's output is text, displayed inside the app. We label it visibly and prominently at every point listed in §3. We do not currently embed a machine-readable watermark in exported share cards. Article 50(2) requires marking that is machine-readable and detectable as artificially generated, "to the extent technically feasible", taking account of the state of the art. Robust, interoperable text watermarking is not something we can supply today; for images, provenance marking is a provider-level capability (Google's SynthID) that we do not control from the client, because the share card is rendered on the device from text, not generated by a model. This is an open item, and it is recorded as one in the compliance checklist rather than papered over. The likely fix is C2PA content credentials on the exported PNG, and it should be reviewed with counsel before the August 2026 deadline.
Article 50(4) — AI-generated text published to inform the public. Does not apply. Wosk produces private output for one user about her own conversation. It does not publish text on matters of public interest.
Article 50(5) — clear, distinguishable, timely. The disclosure is given at the first interaction (the consent screen) and repeated on every result, in plain language, in the same type as the surrounding copy rather than in fine print.
5. The limits of the output — read this part
This is the section that matters more than the legal analysis.
It is one model's reading of a fragment. Wosk sees the screenshots you chose to upload. It does not see the rest of the conversation, the calls, the years of history, the tone of voice, the joke you two have been running since March, or anything he has ever done. A person reading the same fragment with all that context would frequently reach a different conclusion — and would usually be right.
The numbers are not measurements. "Interest 34/100" and "confidence 87%" are generated tokens shaped to look like statistics. They are not derived from data about you, they are not benchmarked against outcomes, and they do not predict anything. They are there because a number is a sharper way to say "not much" than "not much" is. Treat them as writing, not as evidence.
Attachment styles are pop psychology, not a diagnosis. Attachment theory is real psychological research. What Wosk does with it is not: it labels a person neither the model nor we have ever met, from a handful of text messages, with a term from a four-item list. A clinical assessment involves a trained human, an interview, and validated instruments. This has none of those things and does not claim to.
Speaker attribution can be wrong, and when it is, everything is wrong. If the model puts his message on your side of the screen, the entire analysis inverts. This is why the app asks you which side of the screen you are on, and why it warns you when it is unsure. Check the decoded pairs against what you remember before you believe any of it.
It will not always be the same. The analysis model runs at a non-zero temperature, which means the same conversation can produce a different verdict on a different day. That is a property of the technology, not a bug we have not fixed yet.
It is not about him. It is about a piece of text. Nothing Wosk produces is a finding of fact about any real person's intentions, feelings, honesty, mental health or conduct, and it must never be presented to anyone — to him, to friends, to a court — as though it were.
6. If the output is wrong, offensive or harmful
Every result has a Report this analysis link in its footer, always visible, never behind the paywall. It takes four taps and an optional note. Reports are stored, read, and used to change the prompts.
You can also write to favamvv@gmail.com.
7. Human oversight
The pipeline is fully automated by design — the whole privacy story in §5 of the Privacy Policy rests on no human reading your conversation. Human oversight therefore happens at the level of the system, not the individual output: we read reports, we tune prompts, and we run a safety classifier between the two model calls that can block an analysis outright or strip the entertainment framing from it. If you want a human to look at a specific analysis, reporting it is how you ask.