Where do you find an older version of the terms?
Start with copies you already hold, then the company’s own record, then public archives. Record where each copy came from and the date printed on it.
- A copy you saved. Save the text or a PDF on the day you accept terms inside an account. A dated copy is the simplest evidence of what you were shown.
- The company’s change record. Some services print a “last updated” date or keep a change log. Check the footer and any history section.
- Web archives. Public archive services sometimes keep dated snapshots. This is general practice, not something our sources cover, and coverage is uneven.
Dated copies matter because terms change. Zoom’s blog post on how its terms apply to AI features is dated 7 August 2023, with an editor’s note saying it was updated on 7 February 2024. In it Zoom says, “We evaluate and update our terms of service from time to time.”
Why is a raw diff noisy, and how do you compare clause by clause?
A raw line diff is noisy because layout changes shift the lines below them. Re-wrapped lines, renumbered clauses and an inserted heading get reported as edits, even when the words are the same.
| Edit (invented) | Line diff | Clause comparison |
|---|---|---|
| Clause 7 renumbered to 8, same words | Old line removed, new line added | Same clause, label changed |
| Paragraph re-wrapped at a new width | Several lines changed | No change |
| New clause inserted above clause 9 | Every line below it shifts | One added clause |
| Notice period changed from 30 days to 7 | One changed line in a long block | One changed clause, numbers flagged |
Clause-by-clause comparison works in five steps. You can run them by hand or with a tool.
- Normalise. Collapse extra spaces and line breaks, strip markup, and keep clause numbers as labels.
- Split. Cut each version at numbered clauses and paragraphs, and cut very long paragraphs at sentence ends.
- Align. Pair each old clause with the new clause it most resembles. A clause with no close partner counts as added or removed.
- Diff the words. In each changed pair, mark the inserted and deleted words, so a “30” that became “7” stands out.
- Label moves. A clause that moved with no words changed is a layout change.
Which changes deserve attention first?
Start with changes that alter what you can do, what you pay, or what happens to your content and data. Four groups usually matter most:
- How the terms can change: new wording that lets the company change them, and any shorter notice.
- How disputes are handled: arbitration instead of court, class-action waivers, and which court or law applies.
- Content and data: how wide a content licence is, whether content may train AI or machine-learning models, and sharing with partners.
- Money: auto-renewal, cancellation steps, refunds and liability caps.
The tool’s severity rules follow similar logic. A new or stronger arbitration clause, class waiver, AI-training licence, data sale, shorter notice or wider licence is rated high. A weaker or removed term is reported as less restrictive, and it still deserves a read. The red-flags guide explains ten clause types in plain English.
How do you read numeric changes?
Read each changed number with its sentence, because a number alone does not say what it governs. Notice before a change, liability caps, retention periods, claim deadlines and renewal periods are the numbers the tool extracts and compares.
Invented example: the old version gives 30 days’ notice before changes, and the new version gives 7. That is 23 fewer days to decide whether to stay.
The annex note ties its carve-out from item (j) to notice and to the consumer’s right to exit. Its annex note says item (j), on unilateral changes, is “without hindrance to” terms that let a seller change a contract of indeterminate duration, provided the seller gives reasonable notice and the consumer is free to end the contract. Read a shorter notice together with your right to leave, and check the directive text on EUR-Lex.
What do the EU unfair-terms annex items cover?
The annex to the 1993 EU directive on unfair terms in consumer contracts lists terms that “may be regarded as unfair”. The list is indicative and non-exhaustive, so a listed item is a reason to read closer, not a finding that a term is unfair.
| Annex item | What it describes | Plain English |
|---|---|---|
| Item (j) | Lets the seller alter the contract unilaterally without a valid reason specified in it | The company can change the deal with no stated reason |
| Note 2(b) on item (j) | Item (j) is without hindrance to such changes if reasonable notice is given and the consumer can end the contract | A scope note, not a finding that such a change is fair |
| Item (q) | Hindering legal action, particularly by requiring disputes to go exclusively to arbitration | Forced arbitration in place of court |
| Item 1(a) | Excluding or limiting liability for death or personal injury | Caps on liability for serious harm |
| Item 1(b) | Inappropriately excluding or limiting the consumer’s legal rights | A catch-all for rights removed by wording |
Article 5 says that where terms offered to the consumer are in writing, they “must always be drafted in plain, intelligible language”. Where a term is in doubt, the interpretation most favourable to the consumer prevails. This guide did not check later amendments or national laws, so the table describes the directive as written.
What does reading the terms cost in time?
McDonald and Cranor’s 2008 draft estimates reading time for privacy policies at 250 words per minute. It is a pre-press draft, and the published version was not checked, so treat the figures as a 2008 model.
| Policy length in the draft | Words | Minutes at 250 words per minute |
|---|---|---|
| Short policy | 2,071 | 8 |
| Median policy | 2,514 | 10 |
| Long policy | 3,112 | 12 |
Do the same sum for any document: word count divided by a stated reading rate. A 3,000-word page at 250 words per minute takes 12 minutes (invented example). The authors say their estimates “may be slightly low due to the jargon and advanced vocabulary”.
The draft also gives two annual figures: “approximately 201 hours a year” in its conclusions and a “point estimate of 244 hours per year” in its Table 7, with a range of 181 to 304 hours. Show the range rather than one figure. Its dollar estimate, “on the order of $781 billion”, rests on assumed values of time, so it is a model, not a measured cost.
ToS;DR, a volunteer project, makes a qualitative point. Its About page says “almost no one really reads the terms of service we agree to all the time.” That is not a measured rate.
How does the diff tool automate the steps?
The Terms Change Risk Diff runs the same steps in your browser. You paste an old and a new version, and it returns a clause-level diff. Each changed clause is classified by risk category, with a plain-English explanation and what to check next.
Numbers get their own pass. Notice days, caps, retention periods, claim windows and renewal periods are extracted from each version and compared, so “30 days” and “7 days” sit beside the clause. With one text pasted, the tool becomes a risk scan of that document. The tool processes the text in your browser.
Use the tool to decide what to read first, not what a clause means. A category match is a pattern, so read the flagged clause in full. Paste both dated versions into the Terms Change Risk Diff and check its top items against your own reading.
What this cannot tell you
A diff shows what changed in the words. It does not show what a change would mean in a dispute, or whether to accept it. Common mistakes:
- Comparing the wrong versions. Regional, app or business versions can differ. Confirm both apply to you.
- Trusting an undated copy. Without dates you cannot tell which change came first.
- Reading one number alone. A 7-day notice means little until you read its clause.
- Skipping favourable changes. A removed restriction can sit beside a new one in the same clause.
The sources here date from 1993 to 2026 and do not cover every country’s rules. The 2016 CFPB notice describes a 2015 study as it stood in 2016, and this guide does not check what followed. Remove personal details before you ask a chatbot about a clause, as the guide on what not to tell a chatbot explains. This is general information, not legal advice.
Common questions
How do I compare two versions of a terms of service?
Get both versions and note the date of each. Compare clauses rather than lines, so renumbering and re-wrapping do not look like changes. Check changed clauses by risk: how the terms can change, disputes, content and data, then money. Check every changed number against its full sentence.
Where can I find an old version of a website’s terms?
Start with a copy you saved, then the company’s change log or last-updated date, then public web archives. Archives sometimes hold dated snapshots, but coverage is uneven, and this is general practice rather than something our sources verify. Record the date and source of every version.
How long does it take to read terms of service?
McDonald and Cranor’s 2008 draft estimates about eight to twelve minutes for popular-site privacy policies, with ten minutes as the point estimate for the median policy at 250 words per minute. The authors say the figures may be slightly low. Treat it as an old model, and divide your own word count by a stated reading rate.
Why does a diff show every line as changed?
Renumbering, re-wrapped lines and an inserted heading shift the text below them. A line-based diff then reports text as removed and added even when the words match. Normalising whitespace and numbering before comparing separates layout changes from real word changes.
Does a shorter notice period make a term unfair?
A shorter notice period is one fact to weigh. The annex note makes its carve-out from item (j) depend on reasonable notice and on the consumer being free to end the contract, so check the notice against both. Whether a particular term is unfair is a legal question this guide does not answer.
Does Zoom use my meetings to train AI?
Zoom’s blog post dated 7 August 2023, with an editor’s note of 7 February 2024, says Zoom does not use customer audio, video, chat, screen sharing, attachments or other communications to train its own or third-party AI models. It is a dated company statement, not a check of Zoom’s current terms, which this guide did not read.
Sources and further reading
- Council Directive 93/13/EEC on unfair terms, original 1993 text, EUR-Lex (OJ L 95, 21 April 1993)
- CFPB, “Arbitration Agreements” proposed rule, Federal Register web edition, 24 May 2016 (not the official legal edition)
- McDonald and Cranor, “The Cost of Reading Privacy Policies,” authors’ pre-press draft, 2008
- Zoom, “How Zoom’s terms of service and practices apply to AI features,” 7 August 2023, updated 7 February 2024
- ToS;DR, About page, volunteer terms-of-service project, read 17 October 2026