Nobody reads the terms.
So we built something that does.
Studies consistently show that fewer than 1% of users read terms of service agreements, and the ones who try would need weeks per year to get through them all. Companies know this. Some of the most consequential clauses on the internet, such as giving up your right to sue, licensing away your creations, or agreeing to future rules you haven't seen, hide in documents designed not to be read.
Fineprinted reads them for you.We continuously monitor the Terms of Service, Privacy Policies, and related legal documents of tracked services. Every document is split into individual clauses, and an AI classifies each one against a strict taxonomy of known user-hostile patterns, and results publish automatically, with the original clause text attached so you can always verify a finding yourself. AI can make mistakes: that's exactly why the receipts are part of the product.
And because terms change quietly, saving a service means you'll see what changed, when, and what it means, in plain English, not legalese.
How grading works
Every service starts with 100 points. Each kind of hostile clause deducts points, and genuinely pro-user clauses earn some back. Piling on more of the same kind of problem has diminishing effect, and a few of the most serious practices, like forced arbitration or selling your data, cap the best grade a service can earn no matter what else it does. The final score maps to a letter grade:
- A90–100Respectful terms. Rare.
- B75–89Minor concerns, nothing alarming.
- C50–74Several hostile clauses worth knowing about.
- D25–49The fine print works against you.
- F< 25Read nothing, agree to everything? Not here.
What we look for
The AI doesn't free-style its judgments. Every clause is classified against a fixed taxonomy with fixed point values, so grades are consistent and comparable across services. These are some of the most consequential patterns; the full taxonomy covers dozens more, from biometrics to dark-pattern cancellations:
- −15Data saleYour personal data is sold outright or handed to data brokers for their own use.
- −12Forced arbitrationYou waive your right to sue in court or join a class action; disputes go to private arbitration.
- −10Biometric collectionThey collect biometric identifiers like your face, voice, or fingerprints.
- −10Content ownership grabThey claim ownership, or a broad perpetual license, over everything you create and upload.
- −8AI training on your dataYour content and personal data are used to train their machine-learning models.
- −8Hard to cancelCancelling is deliberately difficult: phone-only, buried, or guarded by retention hoops.
- −6Third-party trackingYou're tracked across the web and profiled for targeted advertising.
- +6Right to delete your dataA pro-user right: you can ask the service to delete your personal data.
Direction matters: deductions only apply when a clause imposesa practice. A clause that explicitly rules one out, like “we do notsell your personal data”, counts in your favor instead. Each pattern counts once per service, no matter how many clauses repeat it.
How we keep grades honest
Fully automated, and upfront about it. Analysis publishes without human editing, so what you see is exactly what the AI found. AI can make mistakes, which is why everything below exists.
Confidence thresholds. When the AI isn't sure a clause fits a category, the finding is automatically excluded from the grade.
Semantic change detection. We diff documents by meaning, not formatting: a reworded sentence is a change; a reshuffled page layout is not.
Receipts included. Every flag links to the original clause text, so you can read exactly what the document says and judge for yourself.