The Scale
Every clause gets a score from 1 to 10. 1 = amazing (the company isn't taking anything weird from you). 10 = crazy (a clause that would make most people stop and say wait, they can do THAT?). We only display clauses rated 5 or above. Anything lower is standard boilerplate that wouldn't surprise a privacy-conscious reader.
Genuinely unusual clauses that go well beyond industry norms. Things that would surprise even someone who reads TOS regularly. Examples: perpetual irrevocable content licenses, blanket AI training rights, unilateral termination with no appeal.
Clauses that are problematic but common across the industry. Still worth knowing about, but you'll find similar language in many services. Examples: broad data sharing with “partners,” forced arbitration, vague content moderation policies.
Genuinely user-respectful clauses. Things companies got right where most don't. We surface these in the “Doing it right” section on the homepage, even when the company's overall score is bad. Credit where credit is due, clause by clause. A 1 on this scale is a good thing, not a bad one.
What the AI Evaluates
We use Claude (by Anthropic) to read the full Terms of Service document and identify the most concerning clauses. The AI evaluates each clause on these specific criteria:
How much control do they claim over your content, data, or behavior? Perpetual worldwide licenses score higher than limited use permissions.
Is the consent meaningful? Burying major rights grabs in dense paragraphs scores higher than clear, upfront disclosures.
Do competing services require the same thing? A clause that's unique to one company scores higher than industry-standard language.
What can this actually do to you? A clause that enables selling your biometric data scores higher than one about email preferences.
Can you undo it? Clauses that survive account deletion or waive class action rights score higher because the damage is permanent.
Clause Selection
We include 5-10 clauses per analysis depending on how much is worth flagging. Every clause we show must pass this test:
"Would this make someone think: wait, they can do THAT?"
If a clause wouldn't surprise a privacy-conscious person, we leave it out. We prioritize: rights they take from you, things they can do without your consent, and ways they limit your recourse.
Exact Quotes Only
Every clause shown on CrazyTOS is a verbatim quote from the original Terms of Service document. We never paraphrase, summarize, or editorialize the legal language.
EXACT MATCH vs CLOSE MATCH: EXACT MATCH means the clause text appears character-for-character in the source. CLOSE MATCH means the AI paraphrased lightly (dropped a comma, smoothed a phrase) but the substance is clearly identifiable in the source. We still record what we showed on Arweave so anyone can audit that the displayed text matches our hash. Just be aware that for CLOSE MATCH clauses, the recorded hash is of our displayed wording, not byte-identical source text.
Community Voting
The AI rating is a starting point. As people vote on clauses, the community score gradually replaces the AI score. This means:
- New clauses show the AI rating until enough votes come in
- Your vote directly influences the displayed score
- The community can correct the AI when it over- or under-rates something
- Vote counts are shown so you can see how many people weighed in
The Translation
Each clause comes with a plain English translation written by the AI. These are intentionally casual - "like texting a friend" - so anyone can understand what the legal language actually means for them. Translations are not legal interpretations and should not be treated as legal advice.
The Crazy Graph
The Crazy Graph plots every company on a two-axis chart based on their flagged TOS clauses. Up and to the right is better.
Based on termination, arbitration, liability, and payment clauses. Can you leave? Can you sue? Can you get a refund? Higher severity in these clauses means fewer rights, pushing the company further left.
Based on content license, data collection, AI training, privacy, and third-party sharing clauses. What happens to your stuff? Higher severity means they take more, pushing the company further down.
The Algorithm
For each axis, we compute a weighted severity from the flagged clauses in that category:
weighted_severity = (average * 0.6) + (worst_clause * 0.4)
position = (10 - weighted_severity) * 1.25
The 60/40 weighting means one truly bad clause pulls the score down significantly - a single 9/10 clause can't be buried by a handful of moderate ones. The 1.25 scaling factor means a weighted severity of 6 lands at center (5), and anything above 6 puts the company below center.
If no clauses were flagged in a category, the company scores a 10 in that dimension - no red flags means things are probably fine. Community votes gradually replace AI ratings as people weigh in.
You keep your data, content, and rights
You keep your data but they limit your rights
You keep your rights but they take your data and content
They take your data and your rights
What We Don't Do
We don't provide legal advice. CrazyTOS highlights clauses for awareness. Consult a lawyer for legal decisions.
We don't editorialize ratings. The AI rates based on the criteria above, not opinions about the company.
We don't accept payment to alter scores. No company can pay to lower their rating or remove a clause.
We don't hide our methods. This page exists because we believe the same transparency we demand from others should apply to us.
Disclaimer: CrazyTOS is a consumer awareness tool. Initial ratings are generated by Anthropic's Claude AI and may be adjusted by community votes. This is not legal judgment. All clauses are direct quotes from publicly available Terms of Service documents. Always refer to the original document for authoritative terms.