How the AI works

Everything DatingAppBoost gives you back — the diagnosis, the photo-by-photo observations, the rewritten bios, the opening lines, the reply suggestions — is produced by an artificial intelligence system, without prior human review. This page states it in black and white, in line with article 50 of the European AI Act.

No content produced here is written by a person. If you publish a bio or a message generated by the service, you are publishing text written by a machine.

The service relies on a large language model able to read images. The provider currently chosen is Google (Gemini models). If Google is temporarily unavailable, the analysis is automatically retried with Anthropic (Claude), established in the United States: both may therefore receive your screenshots, depending on the availability of the first. The application can also run with OpenAI, and automatic text moderation, when enabled, goes through that third provider.

Your screenshots and your texts are sent to whichever provider handles your analysis, for as long as it takes to produce it. They are not used, by us, to train any model. The detail of the transfers is in the privacy policy.

The exact model name and the version of the instructions used are recorded with every analysis. If you contest a result, we can therefore know precisely what produced it.

This is an important distinction, and it is true in the code, not only in intent. The model never produces the score out of 100. It fills in a grid of criteria — framing, variety, technical quality, coherence of the text… — each rated from 0 to 5, and writes observations.

The final score is then computed by a deterministic formula, a weighted average of those criteria, written once and covered by automated tests. Given an identical grid, the score is identical. That is what makes the score explainable: it always breaks down into categories you can read.

When the model has not seen enough to reach a conclusion, the score is shown as preliminary rather than presented as a verdict.

The object of the analysis is the effectiveness of a profile: a photo, a framing, a text. Never a person. The instructions given to the model explicitly forbid it from assessing:

• anyone's beauty, physical attractiveness or desirability;
• personal worth, intelligence, social status or income;
• origin, age, gender, orientation, religion, weight, health condition or disability.

Age is never estimated from a face: it is only picked up if it is written in plain text. A low score describes a presentation problem, never a flaw in the person. The system performs no emotion recognition, no biometric categorisation and no identification of individuals.

The "My matches" feature reads the profile of a person who asked for nothing and who will never read what is written about them. A separate and stricter policy applies there. About them, the system produces:

• no score, no rating, no ranking, no tier;
• no judgement about their appearance, their "worth", their social standing or what they would be worth as a partner;
• no inference about their origin, health, religion, orientation, income or state of mind — even if a screenshot seems to suggest it;
• no hypothesis about what they think, feel or really want.

The very format of the analysis contains no numeric field about them: a score that exists, even internally, always ends up being displayed. The system also refuses persistence tactics, guilt-tripping, fake common ground and comments about the body, and it reports clear disinterest rather than suggesting another approach.

No screenshot of their profile or of your conversation is kept: only the text that was read remains, in your account.

You should know this before following any advice:

• It gets things wrong. It can misread an image, invent a detail, or treat a deliberate choice as a flaw.
• It judges what it sees. A blurry, cropped or low-resolution screenshot produces a weaker analysis — and the service tells you so rather than ruling anyway.
• It does not know your context. Not your city, not the audience you are aiming for, not what you are actually looking for.
• It reflects conventions. What it treats as "effective" comes from learned regularities, which carry their own cultural and aesthetic biases.
• It is not reproducible word for word. Two analyses of the same photo can differ in their wording.

A score is not a verdict on you. It is a fallible reading of a presentation you remain free to keep exactly as it is.

The score and the recommendations have no legal effect and produce no comparable effect: they condition access to no right, no service, no benefit. They are editorial suggestions. Article 22 GDPR, which targets automated decisions producing significant effects, therefore does not apply.

The only automated processing that restricts anything is the counting of usage quotas, which rests on a tally, not on profiling.

If an analysis strikes you as wrong, unfair or out of line, write to contact@datingappboost.com quoting the date of the analysis. A person reviews it, replies to you, and corrects the system's instructions if the problem is real. You can also ask for the analysis concerned to be deleted.

We want to be told about any discriminatory, insulting or manipulative result: those are precisely the behaviours the rules above are meant to prevent, and a report is the best way to check that they hold.