What 13 AI vendors commit to on training, retention, human review, intellectual property (IP) indemnity, notice and model retirement, answered for each plan tier. Every stated value cites the document, section and a quote found word for word in our capture, with the capture date and a snapshot hash, so anyone on your team can check it. “Not stated” names the documents we searched, and values still in review are marked as such.
The matrix is $299 one time. The weekly edition is from $149 a month.
Two real cells from the 26 Oct 2026 matrix. Same vendor, same question, two plans, opposite answers.
OpenAI · ChatGPT Plus and Pro
Trains on your data by default
Opt-out Yes, unless you opt out
“When you use our services for individuals (Free, Plus, Pro), we may use your content to train our models.”
Help Center: how your data is used to improve model performance, § Services for individuals
Captured 26 Oct 2026 · snapshot 0e2127a7bcf5 · 4 of 4 models agreed
OpenAI · ChatGPT Business and Enterprise
Trains on your data by default
Opt-in No, unless you opt in
“By default, we do not use your business data for training our models.”
Enterprise privacy & API data usage, § Enterprise privacy at OpenAI > Does OpenAI train its models on my business data?
Captured 26 Oct 2026 · snapshot 3f9b541c600b · 4 of 4 models agreed
What it covers
12 of the first 15 vendor documents we tracked changed between 11 Jun 2026 and 2 Oct 2026. In the single week to 19 Oct 2026 we graded 9 changes Material, at AWS (Amazon Web Services), Google and OpenAI, and counted 12 models due to retire within 30 days.
A consumer plan, a team plan, an API (application programming interface) key and a cloud contract from the same vendor often carry different answers on training, retention and indemnity. The matrix keeps them apart, so you read the row for the plan you actually signed.
Terms Watch is the weekly edition: what changed for each plan tier, what it means, who should look, and which models retire soon. It is written to be acted on, not a raw diff of web pages.
The questions each plan row answers
Consumer paid plans are rated on the training questions. API, business, enterprise and cloud plans get all thirteen.
From the Terms Watch edition of 19 Oct 2026
Between 13 Oct 2026 and 15 Oct 2026, AWS added exclusion (viii) to Section 50.10.2 of its Service Terms. Section 50.10.1 is the promise to defend you against third-party IP claims over generative AI output. The new exclusion carves out output from a “Kiro Excluded Model” that you chose to use.
Material · AWS · Amazon Bedrock · IP and indemnity
Before: the list of exclusions ended at (vii)
“(vi) arising from Generative AI Output that you know or reasonably should know may infringe or misappropriate another party’s intellectual property rights; or (vii) alleging that your use of Generative AI Output infringes a third party’s trademark or related rights.”
Added
“(viii) arising from Generative AI Output generated by a Kiro Excluded Model” … “where you chose to use the Kiro Excluded Model.”
Who should look: counsel and vendor-risk teams that rely on the AWS IP indemnity for Bedrock output.
Source: AWS Service Terms, section 50 (machine learning and AI services, including Bedrock), § 50.10.2. Old text last seen 13 Oct 2026, new text first seen 15 Oct 2026, snapshot ffed6ce92155. Graded Material by our published rubric and confirmed by 3 of 3 models that answered.
Who it is for
Re-review a vendor’s data terms when they change, with the clause and date in hand.
Keep the internal AI use policy in line with what each vendor and plan actually allows.
Compare plan tiers before renewal, and keep the vendor register current with citations.
Run several AI subscriptions and APIs, and plan model migrations before retirement dates.
Sources
Public vendor documents
59 documents from 13 vendors: terms of service, privacy policies, data processing addenda, usage policies, subprocessor lists and model retirement pages. Captured every day, logged out, and stored as fetched with a SHA-256 (Secure Hash Algorithm) hash, a fingerprint that shows our copy has not changed.
Earlier versions back to January 2024, where the Internet Archive kept copies.
Vendors tracked: Anthropic, AWS, Cohere, Cursor, DeepSeek, GitHub Copilot, Google, Meta, Microsoft, Mistral AI, OpenAI, Perplexity, xAI.
The coverage list on the method page names every document, with its link and the date we first captured it.
Method
Document, section, a quote of 30 words or fewer found word for word in our capture, the capture date and the SHA-256 hash of the page we read. “Not stated” is an answer too. It has no quote; it lists the documents we searched.
Models from Anthropic, OpenAI, xAI and Google read each document independently. A value is accepted only when at least three agree and its quote is found word for word in our capture. In the 26 Oct 2026 matrix that held for 275 of 289 cells (95%). 6 were settled by a review call made by an AI agent, not a person, with its quote checked against our capture. 8 are marked “in review” and, for now, still show the previous edition’s value next to that mark.
Notice periods, retirement lead times, change counts and days since the last change are computed by code and footnoted. Day counts inside matrix cells, such as a retention period, and model retirement dates are read from the vendor’s text by the models, then checked by code. Each change is dated to the window between two captures, never to a guessed day.
Every edition says which documents we track, which ones missed a capture and which sources are degraded. When a value changes and the documents behind it did not, the next edition lists it as our correction, and so does our corrections page. If a cited fact does not reproduce against its source, tell us within 30 days and we refund the purchase.
The full method and the materiality rubric are on the method page. Mistakes are listed on the corrections page, and the refund policy says when we refund.
On sale now. A new Terms Watch edition comes out every Monday; the latest is the edition of 19 Oct 2026. The matrix is current as of 26 Oct 2026.
Checkout runs in test mode for now, so no card is charged.
Degraded sources: none.
Sample
The free sample is the current matrix for Anthropic, OpenAI and Google, with the same evidence as the full matrix: the quote and source for each stated value, and the documents searched for each “Not stated”. The 12 Oct 2026 Terms Watch edition is online in full, so you can read a whole weekly edition first.
Ways to buy
The current quarterly matrix, one time.
$299 one time
Every Monday: what changed for your plan tier and what it means.
$149/mo or $1,490/yr
Terms Watch for several readers, priced by team size.
Custom pricing
Matrix PDF, 4 pages: the matrix, 90 days of Material changes, the model retirement calendar, sources
Matrix CSV (comma-separated values): document, section, quote, capture date and snapshot hash for each stated value; documents searched for each “Not stated”
Changes CSV: every graded change in the last 90 days
Weekly web edition, made to read on a phone
How you get it
| Format | Matrix | Terms Watch | Team |
|---|---|---|---|
| Matrix PDF, 4 pages: the matrix, 90 days of Material changes, the model retirement calendar, sources | Yes | Refreshed weekly | Refreshed weekly |
| Matrix CSV (comma-separated values): document, section, quote, capture date and snapshot hash for each stated value; documents searched for each “Not stated” | Yes | Refreshed weekly | Refreshed weekly |
| Changes CSV: every graded change in the last 90 days | Yes | Refreshed weekly | Refreshed weekly |
| Weekly web edition, made to read on a phone | No | Every Monday | Every Monday |
| How you get it | Download after purchase | A link by email each Monday | Same, for the people you name |
Not offered yet: logins, an API, Slack or Teams alerts, and alerts between Monday editions.