Terms of service change checker for old and new versions

How to use Terms Change Risk Diff

1. Paste both versions. Put the older text in Old version and the newer text in New version. Files (.txt, .md, .html) can be loaded into each box.
2. Press Compare. The tool aligns the clauses, classifies each change and shows the strongest changes first.
3. Filter and check. Filter by severity or category, read each explanation, then open the clause text and follow the what-to-check note.
4. Copy or export. Copy the Markdown digest, or download the CSV or JSON for your own notes.

What it measures

This tool compares two versions of a terms of service or privacy policy, clause by clause. It does not read a contract the way a lawyer does. It finds the clauses that changed, matches each changed clause against a library of 24 categories (20 risk, 3 rights and 1 neutral), and quotes the words and numbers that triggered each match. You decide what to read next.

The steps run in order. First, pasted HTML is turned into plain text with its scripts removed. Second, each text is split into clauses: numbered lines start a clause, blank lines end one, and a paragraph longer than 90 words is split by sentence. Third, the clauses of the two versions are aligned with the shared text engine's diff.align at a similarity threshold of 0.4. A numbered clause that the alignment reports as removed and added under the same number is paired as a reworded clause when their Jaccard similarity is at least 0.2. That score is the distinct stemmed terms the two clauses share, divided by all the distinct terms in both clauses together. For two ten-term clauses, sharing two terms (a score of about 0.11) is not enough, and sharing four terms (0.25) is. Fourth, each changed clause is searched for category patterns and numeric facts. Fifth, the results are ranked and written out as a summary, a category matrix, a filterable diff and a digest.

This is general information, not legal advice. It is a rule-based aid that points at clauses worth reading. It cannot know your contract, your jurisdiction or the facts of your case, and a flag is not a finding that a term is unlawful. For a serious question, speak to a qualified lawyer or a consumer body where you live.

Reading time uses a stated rate of 250 words per minute, the rate assumed in McDonald and Cranor's 2008 pre-press draft on the cost of reading privacy policies. That draft gives a range of about eight to twelve minutes for policies on the most popular sites and warns that its estimates may be slightly low due to the jargon and advanced vocabulary. Treat the figure as an assumption about a reader, not a measurement of you. The source is the authors' draft PDF.

Before you use the results, read the explanation and the quoted clause text for each flag. Then read the guides on this site: how to compare two versions of terms of service and the red flags to look for in plain English. If you also want a separate check on how a page looks to automated crawlers, use the AI crawler HTML inspector, which is a different tool.

How the scoring works

Each category has a base severity, a ranking weight and a direction. A risk category is one where more of the feature restricts you: a longer arbitration clause, a wider licence, a shorter notice. A right category is one where more of the feature protects you, such as a refund window or a deletion right. A neutral category, such as governing law, is reported as changed without being judged favourable or unfavourable, because the effect depends on where you live.

A changed clause is labelled by how its category changed. The table below gives the rules in plain words.

Negated sentences are checked. A pattern marked as negatable is ignored when the words before it in the same sentence negate it, so we do not sell your personal data does not trigger a sale flag. Sentence-level negation can be fooled by long mixed sentences, so read the quote.

The category list is the core of the tool. Each row shows what it looks for and the base severity.

Numbers it compares

Seven kinds of number are read from the sentence they appear in. The tool reads the sentence's own words to decide the kind, then compares the old and new value of the same kind. The result is written as Notice before changes: 30 days → 7 days (worse for you), with the quoted sentence next to it.

• Notice before changes (days). Shorter notice is worse for you.
• Time to bring a claim (days, weeks, months or years). A shorter window is worse for you.
• Retention period (days). A longer period is worse for you.
• Refund window (days). A shorter window is worse for you.
• Breach notice window (hours or days). A longer wait is worse for you.
• Liability cap (a currency amount such as USD 500, or a number of months of fees). A lower cap is worse for you, because it limits what the company must pay.
• Renewal period. Shown and compared, but not judged, because a longer or shorter renewal can suit different people.

Two numbers in different currencies are reported as not comparable. Words such as thirty (30) days and one (1) year are read as numbers, and a year is taken as 365 days when comparing.

Dated legal context

The facts below are checked against the sources linked. Each one carries its date and source. They describe what those sources say at the time of checking, which was 17 October 2026. They are not a statement of the law where you live, and they do not say what your contract means.

• EU unfair terms. The annex to Council Directive 93/13/EEC of 5 April 1993 is described in its own text as an indicative and non-exhaustive list of the terms which may be regarded as unfair. Its item (j) covers terms that let the seller alter the contract unilaterally without a valid reason which is specified in the contract. Its item (q) covers requiring the consumer to take disputes exclusively to arbitration not covered by legal provisions. Only the original 1993 text was checked, not later amendments or national laws. EUR-Lex text
• Click to cancel in the United States. The FTC's 2024 amendment to the Negative Option Rule was published on 15 November 2024. The notice said the rule now applies to all negative option programs in any media. The FTC's final rule of 12 February 2026 says that, after the Eighth Circuit found the Commission's failure to issue a preliminary regulatory analysis procedurally insufficient and vacated the 2024 Rule, the rule is revised to restore it in the form it existed before the 2024 Rule became effective. On 13 March 2026 the FTC published an advance notice that seeks public comment on the need for amendments. Its 14 August 2026 regulatory agenda lists the removal of the vacated 2024 amendments among the current rulemakings likely to have some impact on small entities. A common belief that click-to-cancel is a current federal requirement is not supported by these notices. This tool does not check what the restored text requires. FTC final rule, 12 February 2026
• Arbitration prevalence. The Consumer Financial Protection Bureau's 2016 notice describes its March 2015 study as reviewing over 850 agreements for certain consumer financial products and services and reports that 15.8 percent of credit card issuers included such agreements in their contracts. The study itself was not read; these are the Bureau's own 2016 descriptions. The notice is a 2016 proposed rule, and this tool does not check its later status. Federal Register notice, 24 May 2016
• New purposes for personal data. Article 13(3) of the GDPR, as reproduced by a third-party site, requires that a controller give the person information on that other purpose before further processing for a purpose other than the one for which the data were collected. The tool uses this as a reading prompt for data-use clauses, not as a rule that governs contract terms in general. GDPR Article 13
• A negated example. Zoom's blog post, published 7 August 2023 and updated 7 February 2024, says that Zoom does not use any customer audio, video, chat, screen sharing, attachments, or other communications like customer content to train its own or third-party AI models. The post is dated, and Zoom's current terms were not checked. It is a useful example of why a negated sentence should not be flagged as training. Zoom blog post
• Grades are a different thing. ToS;DR publishes a points formula, number of good points − number of bad points − (3 × number of blockers), with a classification version as of Sept 2025. This tool does not produce a grade for any service. ToS;DR classification method

Limits and what it cannot tell you

The tool matches written patterns. It is not a language model, and it does not understand a clause the way a reader does. It will miss wording it has no pattern for, for example an arbitration clause written without the word arbitration. It can flag a phrase that appears in an unrelated sense, such as affiliates in a definition. It can misread a negated sentence when a long, mixed sentence places not before the risky clause. A clause split at the wrong place can give a number the wrong kind.

It compares two texts you provide. It does not fetch the current terms of any service, so it cannot tell you whether the version you saw is the one in force. It does not know dates unless the texts state them, and it does not check the law of any country. Clause pairing depends on numbering: when clauses are renumbered and reworded at once, a pair may be missed and shown as removed and added.

Very large unmatched sections are listed as removed and added without pairing, and the result says so. The input limit is 250 KB per box. In Node tests, the sample pair and a random 250 KB input each took about 0.05 to 0.15 seconds; browser timing was not measured. Cancelling stops the run before its next step.

Privacy: what happens to your input

The tool's code reads, splits and compares both texts in this tab, and it never sends your texts anywhere. It never stores them either: the tool writes nothing to browser storage. The tool's code records nothing about which clauses you read, and it uses no analytics.

A download creates a file in your browser from the results on screen, and the object URL is released after the download starts. The settings link carries only the filter choices (severity, category and whether to show only changes). It never carries the text.

When you load a file, the tool reads it in this tab. HTML files are parsed as inert documents: their scripts and styles are dropped, and nothing in them runs or is inserted into this page.

Every page on this site also loads a promo banner. The site-wide promo banner is separate and never receives your text.