We publish factual claims about New Zealand law and about what AI vendors do with data. Some of them will be wrong.

When that happens, here is what we do.

We correct it and say so. The correction appears on this page with the date, what was wrong, and what it should have said. We do not quietly edit a page and move on.

We say where it appeared. If a wrong claim went out in a newsletter or a course module, the correction names that, so anyone who read the original can tell whether it reached them.

We date everything. Every rating and every reference page carries the date it was last verified. A claim that was accurate in March may not be accurate now — particularly for vendor terms, which change without announcement.

Ratings reflect what was published when we checked. If a vendor has changed its terms since, that is a change rather than an error, and it appears in the alerts rather than here. If we read something wrong, that is an error and it appears here.

When a rating gets old, we say so on the rating. Every profile is fully re-checked at least every six months. If one passes six months without a full re-check, a notice appears on it automatically, without anyone deciding to put it there. You should expect to see those notices from time to time. They are the system working, not the system failing — a publication that never showed one would either be re-checking nothing or hiding it.

If we have a privacy breach, we will tell you. We will notify anyone affected as soon as we can, notify the Privacy Commissioner where the law requires it, and publish what happened on this page with the date — whether or not the breach meets the legal threshold for notification. We grade other organisations on how they handle data. It would be indefensible to hold ourselves to a lower standard than we hold them.

If you think something is wrong, tell us. quentin@aileakage.com. Point us at the specific wording if you can, it is faster than us guessing which part. We would rather know.

Corrections to date

This page said “none yet” until 3 September 2026. That was itself wrong. Corrections had been made on this site since 5 August 2026 — at least fourteen of them, across ten products, two of which moved a rating. Each was written onto the profile it belonged to, with a date and an explanation. None of them reached this page. By the standard set out above, that made every one of them a quiet edit, which is the exact thing this page exists to prevent. The list below is the backfill, assembled from the correction notices on the profiles themselves.

25 October 2026

• Alert of 8 September 2026, “Zoom has renamed AI Companion to ZoomMate”: wrong on the main point. We said Zoom had renamed AI Companion to ZoomMate, that it was the same product, that nothing on Zoom’s site said so, and that we did not know when it changed. Zoom retired the AI Companion name: the AI features included with paid plans continue under plain names as Zoom AI features, and ZoomMate is a separate agentic product, a paid add-on or standalone subscription, which Zoom says launched on 1 June 2026. Zoom explained this in a blog post of 24 June 2026 and a community guide of 22 June 2026. The alert has been rewritten, and its advice now says not to treat ZoomMate as already approved. Our Zoom profile carried the same “renamed” wording and is corrected too. No rating changed.
• GitHub Copilot: a wrong vulnerability number. We identified the CamoLeak vulnerability as CVE-2025-59145. MITRE’s record for that number describes a different vulnerability, a compromised npm package; the researchers’ disclosure and The Register’s report give no CVE, and we found none. The 9.6 severity rating comes from Legit Security, which found it. No rating changed.
• Cursor: plans and prices out of date, and one date wrong. Our plan table and recommendations still named Pro at US$20 a month and Teams at US$40 a user a month. Cursor’s plans are now Free, Plus (US$8), Premium (US$30), Super (US$100) and Ultra (US$200), with Teams from US$30 a user a month. Our rating labels still use the old plan names until the post-launch re-score, and the profile says so. We also dated the MCP installation vulnerability (CVE-2025-64106) to December 2025; it was published on 4 November 2025. No rating changed.
• Asana: an incident misdated, and prices mislabelled. We dated the Asana MCP server data exposure to June 2024 and called it a forerunner of the 2025 MCP vulnerabilities. Asana launched the MCP server on 1 May 2025 and found the flaw on 4 June 2025. We also gave the monthly-billing prices (US$13.49 and US$30.49) as annual prices; billed annually, Starter is US$10.99 and Advanced US$24.99. And we listed AI Studio on Advanced only; Asana’s pricing page includes it on Starter, Advanced and Enterprise. No rating changed.
• Linear: two plan details wrong. We described the Free plan as for small teams of up to 10 users; Linear’s Free plan has unlimited members, with a limit of 2 teams and 250 issues. We also said Business cost US$4 a user a month more than a “Standard” plan; Linear has no Standard plan, and Business (US$16) is US$6 more than Basic (US$10), billed yearly. No rating changed.
• Notion: price and plan table out of date. We gave the Business plan as US$15 a user a month; it is US$20 a member a month. Our table said AI provider retention was “Not published” on Free, Plus and Business, after our own 23 October correction had found it was; Notion’s pricing page gives 30 days. And we implied Notion AI was included on every paid plan; it is included from Business up, with a limited trial on Free and Plus. No rating changed.
• Zendesk: a stale plan table, a wrong line about the base plan, and an unsourced claim. Our table still listed Suite Growth, which Zendesk no longer sells, and gave Suite Team and Professional at their monthly-billing prices without saying so; it now shows both billing bases from Zendesk’s pricing page. We said the Support Team plan includes no AI features at all; it includes Zendesk AI agents with 5 automated resolutions per agent a month. We also said Zendesk was “among the first companies to use OpenAI’s European endpoints”, which we could not source, and have removed it. No rating changed.
• Salesforce: model providers named without a source. We said Salesforce’s third-party AI was Anthropic and OpenAI under zero-retention contracts. Salesforce’s Trusted AI page names no providers; it says prompts and responses are never stored or used to train the third-party models. The profile now says that. We also replaced a plan table we had already flagged as out of date with the editions read on Salesforce’s pricing page on 3 September. No rating changed.
• HubSpot: two leftovers from before our own corrections, and an unsourced detail. The profile’s introduction still said HubSpot had “invested heavily in privacy-positive defaults”, and its summary called it an “enterprise-grade privacy posture”, after we had found (and corrected elsewhere on 20 September and 1 October) that HubSpot trains its own AI on customer data by default. Both now say so. We also said which model generation HubSpot’s agents run on, which we could not source, and have removed it. No rating changed.
• Adobe Firefly: we said individual plans carried IP indemnification. They do not. Our profile said paid Creative Cloud and Firefly plans carried “full” indemnification and recommended them to sole traders on that basis. Adobe says IP indemnification is for eligible enterprise customers, for select outputs, through enterprise-specific plans. We also called Adobe “the only AI image vendor offering contractual IP indemnification”, which we could not support, gave a “$50K+ per claim” cap that Adobe’s product description does not state, and listed retired plans and prices (a $4.99 Firefly Premium, All Apps at $59.99). All are corrected, along with the summary on our vendor list and the search description. No rating changed: the Enterprise rating rests on the indemnification, which Enterprise customers do get.
• ElevenLabs: three unsupported claims, one overstated legal point, and out-of-date plans. We called ElevenLabs “the dominant infrastructure for AI voice scam attacks”, an exposed ElevenLabs key “the most actively-exploited credential class in the AI security landscape”, and its output “the highest-quality” available; we had no source for any of them and have removed them. We said voice is “legally classified as sensitive personal information” under the GDPR, CPRA and BIPA; those laws treat voice as sensitive when it is used to identify someone, which is how ElevenLabs’ own policy puts it. We also updated the controller (ElevenLabs Poland sp. z o.o. for the EEA, UK and Switzerland), the policy date (9 October 2026) and the plans (Scale is US$299, not US$330, and there is now a US$990 Business plan). No rating changed.
• Slack: a model provider named without a source, and an opt-out route Slack does not offer. We said Slack AI runs primarily on Anthropic’s Claude via AWS Bedrock under zero-retention arrangements. Slack’s privacy principles name no model provider; they say Slack hosts its LLMs on its own cloud infrastructure and the providers have no access to customer data. Our plan table also said Enterprise+ customers opt out of Slack’s general machine-learning training through admin controls; Slack documents only the email request from a workspace owner, on every plan. We also still called Slack AI a paid add-on, which Slack no longer sells separately. No rating changed.
• Granola: who can switch training off, stated two wrong ways. Our plan table and recommendations said Business users could not opt out of training; any user can, in Settings. Our 3 October correction then said a Business admin could opt out the whole workspace; Granola’s help centre and its 16 October 2026 blog say the switch is per user on Basic and Business, and only Enterprise admins set it for the workspace. No rating changed.
• Fathom: encryption overstated, and plans out of date. We said Fathom applies end-to-end encryption to recordings and transcripts. We have not confirmed that, and it cannot describe content Fathom’s own servers transcribe and summarise; the profile now says the content is encrypted, per Fathom’s documentation. We also listed a “Team Edition” at US$15 to US$29; Fathom now sells Team (US$19, or US$15 annual) and Business (US$34, or US$25 annual), and Premium is US$20 a month, not US$19. No rating changed.
• Cursor (second pass). Our profile said Cursor was “the most widely-adopted AI code editor in 2026”, that its cloud infrastructure was “primarily AWS”, and that in Privacy Mode code “is not stored on Cursor’s servers”. We had no source for the first two; Cursor’s subprocessor list names AWS, Azure, Google Cloud, Vercel and Cloudflare, and its data-use page (3 September 2026) says cached file contents are temporary and never permanently stored, with an exception for prompts that trigger abuse detectors. We also called five code-execution flaws “critical” (most are rated high), dated public disclosure of CVE-2026-26268 to 28 April 2026 (the CVE record was published on 13 February 2026), and said Privacy Mode is “enforced organisationally” on Teams, which repeats the error we corrected on 25 September: it is on by default, and an admin has to switch enforcement on. No rating changed.
• ElevenLabs (second pass). Our first pass on 25 October left five lines we could not support. We cited a 2023 Nature Communications finding on voice embeddings that we could not find, and have removed it. We said raw recordings are deleted after 3 years of inactivity; the current Terms and Privacy Policy (9 October 2026) do not say that, and the 3-year ceiling covers data generated about your voice. We called ElevenLabs “the technology of choice” for voice deepfakes and its API keys “the most actively-targeted AI credential class”, and said a stolen key allows “unlimited” generation; none of these was sourced, and usage is capped by the account’s limits. We also said paid plans carry “stronger contractual protections”; the Privacy Policy states one opt-out for every plan, and a business customer’s own contract governs where there is one. No rating changed.
• Slack AI (second pass). We said Slack runs “AWS primarily” without a source; Salesforce’s sub-processor list (19 December 2025) names AWS as the hosting provider for all Slack functionality, Slack AI included, and names no separate AI model provider for Slack. Our recommendation still told small teams to “pay for Slack AI separately”, though Slack no longer sells it separately. We have also added what Slack’s contract says: the Supplemental Terms (27 February 2026) instruct Salesforce to process Customer Data “to train models for use by the services”, and the generative-AI opt-in promise sits in the privacy principles rather than the terms. No rating changed.
• Notion AI (second pass). Three lines still said Notion does not publish how long its AI providers keep data; since 23 October the profile has said up to 30 days on Free, Plus and Business, and those lines now say so too. We have also added that Notion’s no-training commitment, including the bar on others training, is in its AI and Notion Credit Supplementary Terms (27 July 2026), which form part of its customer agreements. We had not read those terms before. No rating changed.
• Asana AI (second pass). Asana’s subprocessor list, last updated 11 September 2026, now also names Google (Gemini Enterprise Agent Platform) for AI features and Google Cloud for hosting; our profile named only Anthropic, OpenAI and AWS Bedrock, and said Asana runs on “AWS primarily” without a source. The profile now lists all four AI providers with their processing locations, and the hosting providers as Asana lists them. No rating changed.
• Granola (second pass). Three lines still said the training opt-out was available only on Enterprise, contradicting our own 3 October correction; every user can opt out, and on Enterprise training is off unless the customer opts in. We said Anthropic is Granola’s “primary” AI provider; Granola’s terms name providers only as examples. And two sentences said Granola’s bot-free design “structurally avoids” the consent question behind the Otter.ai and Fireflies cases, one of them adding that no pending outcome against Granola existed. Granola has been a defendant in a proposed class action on that question since 30 July 2026, which this profile reports further down. No rating changed.
• Salesforce (second pass): our own correction was wrong. Earlier today we removed Anthropic and OpenAI from the profile because we had found no Salesforce source naming them. There is one: Salesforce’s Agentforce Privacy FAQ (June 2025) gives “Anthropic which is hosted on AWS and OpenAI which is hosted on Microsoft Azure” as examples, and its sub-processor document (19 December 2025) lists OpenAI, Microsoft Azure, Google and AWS as generative AI providers. The profile now names them, with both sources. No rating changed.
• Adobe Firefly (second pass). The profile still called Firefly “the most legally-defensible AI image generation product for commercial use”, a superlative we cannot source and which sits badly with this morning’s correction that individual plans carry no IP indemnity. It now compares Firefly only with generators trained on scraped web content. No rating changed.
• How This Site Uses AI, the homepage and four page footers. The How This Site Uses AI page said every claim in every profile is checked by a person, every cited URL is visited by a human, and AI does not assign ratings; and it described a commercial Claude tier with no-training contract terms. None of that matched how the site works or the promise we have made since 31 August, and we could not stand behind the tier description. The page now carries that promise and says where a person decides. It, the homepage and four page footers also still described a “deliberate counter-correction” on the Claude profile, or a “no-training tier”, which we withdrew on 28 September in favour of scoring Claude by the same method as every other vendor.
• Otter.ai and Canva. The Otter profile cited the same untraceable 2023 Nature Communications statistic on re-identifying voices that we removed from ElevenLabs today; it is gone from Otter too. The Canva profile still called its no-training default for Teams “contractual” in one line, after our 23 October correction; it now says it is a Trust Center position, not a contract term. No rating changed.
• Our own pages. The methodology page said we sell industry packs and audit kits, which we do not, and the methodology and home pages gave “licensing to a managed service provider” as an example of a commercial relationship we do not have. The What is AI leakage? page described the 29 profiled tools as the “most-used”, which we have not measured, said every consumer AI tool in our database trains on what you paste by default (most do; some do not), and its search description still carried our old, data-only definition. The How we built this list page now matches the emails we send: why we write, roughly how many people, and how to stop. No rating changed.
• How This Site Uses AI, other tools. The page named the plans we used ChatGPT and Perplexity on, and said no other AI tools were used in our editorial work. We could not stand behind either statement and have removed them. The page now says we occasionally cross-check a fact with another AI tool, and that any claim we publish cites the vendor’s own document, not the tool.

24 October 2026

• ChatGPT: four facts we had wrong, re-checked against OpenAI’s own pages. We named the company responsible for New Zealand users’ data as “OpenCorp Inc., 3180 18th Street”; OpenAI’s Privacy Policy (6 February 2026) names OpenAI OpCo, LLC, 1455 Third Street, San Francisco. We said ChatGPT Business keeps chats for a 30-day abuse-monitoring period by default; that is the rule for the API, and Business chats are kept until deleted, with retention set by the workspace admin. We said a Mixpanel breach compromised data across about 8,000 Mixpanel customers; 8,000 is Mixpanel’s whole customer base, and it has not said how many were affected. We said several investigations over 2024 and 2025 found extensions stealing conversations from 900,000 users; that figure comes from one December 2025 report on two extensions. Two lines were out of date rather than wrong: Enterprise and Edu workspaces can now choose to store data in Australia, among ten regions, and the price table, which was in US dollars and left out Go and Pro, is now in NZ$ with all six plans. No rating changed.
• Claude: retention and data residency, re-checked against Anthropic’s own pages. This check was done by an instance of Claude, the product the profile rates; the corrections run both ways. Against Anthropic: we said Enterprise customers could choose data residency; Anthropic’s regional compliance page offers residency only through AWS Bedrock, Google Vertex AI and Microsoft Foundry. We said some API logs were kept for only 7 days from September 2025; Anthropic’s current article gives 30 days. In Anthropic’s favour: we said consumer chats with training off are kept 30 days (they stay in your history until you delete them, then leave Anthropic’s systems within 30 days), and that users who “did nothing” had all their conversations trained on and that a Pro user’s client information “has been included”; training applies to new and resumed chats while the setting is on, so we now say “may have been used”. We also removed an unsourced claim that Amazon Web Services is Anthropic’s main cloud provider, and updated the plan table (Team for 2 to 150 people, Enterprise 20-seat minimum). No rating changed.
• Gemini: a security incident overstated, and three out-of-date lines, re-checked against Google’s own pages. We described a January 2026 Calendar data leak as “observed in the wild”; it was a research demonstration by Miggo Security, and BleepingComputer reported no sign of real attacks. Our summary of Gemini’s agent risks said the same. We said Workspace data regions were an Enterprise-only option with a choice of specific regions; Business Standard and Plus have them too, and the only regions are the United States and Europe. The consumer training setting is now called Keep Activity, and the paid consumer plan is Google AI Pro, not Gemini Advanced. Workspace prices are now shown in NZ$. No rating changed.
• Microsoft Copilot: a withdrawn plan, a misattributed quote and a reversed comparison. We listed Copilot Pro at US$20 a month as a current plan; Microsoft stopped selling it to new subscribers in October 2025, and Copilot now comes with Microsoft 365 Personal, Family and Premium. We attributed the phrase “generative AI’s greatest security flaw” to NIST; it is a description used by other writers, not NIST’s wording. We wrote that Microsoft 365 Copilot’s rating of 2 was “lower risk” than Workspace Gemini’s 1; on that scale 2 is one band riskier. Prices are now shown in NZ$ from Microsoft’s New Zealand pages. The site’s vendor data feed was also still carrying the consumer no-training sentence we withdrew on 23 October; it now matches the page. No rating changed.
• Grok: a missing incident, a missing plan and two dates. We had not recorded that in August 2025 more than 370,000 shared Grok conversations became searchable on Google, as Forbes reported; it is now in the profile’s incident history. We said there was no Grok offering for small teams on the privacy axis; SpaceXAI sells Grok Business at US$30 a user a month and lists it as not trained on. We dated the default-on training of public X posts to November 2024; the setting was found switched on in July 2024, and November 2024 is when X’s terms added it. We also listed Microsoft Copilot among consumer tools that train by default, which our Microsoft profile no longer says. No rating changed.

23 October 2026

• Vendor list summaries: the short descriptions on our vendor list, and in the site’s public data feed, still carried claims we had already withdrawn from the profiles. Xero “operates regional data residency for AU/NZ customers” (withdrawn 26 September); Fathom named Fireflies.ai as “currently defending” a class action (amended later on 23 October 2026: Fireflies’ December 2025 case was voluntarily dismissed on 11 March 2026, but a new class action against it was reported filed in March 2026, so that summary may have been right; the new summary does not mention Fireflies); Intercom customer data “not used for model training” (Intercom’s Terms of 8 September 2026 allow it); Jasper’s no-training commitment “across all plan tiers” (Business plan only); Perplexity’s “pending” class action and a 4 of 5 rating (the case was dismissed on 1 May 2026 and the rating moved to 3 on 3 September). Canva’s summary said its AI Product Terms made the tier rule “crisp” (they do not state it), and HubSpot’s that it had “invested heavily in privacy-positive defaults” (HubSpot trains on customer data by default). All 29 summaries are now written from each profile’s corrected opening line. ZoomMate’s search description said it is included with paid plans; it is a paid add-on, licensed per user. Summaries and search descriptions are now on our correction checklist.
• AI Tool Risk Directory: nine rows had not been updated since 3 September. Perplexity’s consumer tier showed 4 of 5 and Microsoft’s consumer Copilot 3 of 5, each a band worse than its own profile; Midjourney listed a Free plan it does not have; Cursor and ElevenLabs were missing a band; ChatGPT, GitHub Copilot, Granola and Replit showed out-of-date plan labels. Every row is now copied from the rating on the tool’s own profile. Cursor’s lowest-risk band was labelled Business/Enterprise, a plan our 25 September correction says does not exist; it now reads Teams/Enterprise. No rating moved.
• Perplexity: eight lines on the profile still treated a dismissed case as pending. The opening line said a 2026 lawsuit concerned how Perplexity’s crawler behaves (it concerned ad trackers, and it was withdrawn by the plaintiff on 1 May 2026, before any ruling); the recommendations called the consumer tiers higher-risk than ChatGPT, Claude and Gemini and “the worst in this database”, and the plan table and jurisdiction list referred to a “pending lawsuit”. All are rewritten to match the dismissal the same profile records. No rating moved.
• Shopify Magic: the training section opened with a claim our own re-check had withdrawn, and the litigation paragraph described a superseded position. It began “No training on customer data by default. Shopify’s subprocessor agreements prohibit training on merchant or shopper data. This is contractual”, directly above our 3 September note that Shopify’s commitment is narrower; the jurisdiction list repeated it. Shopify does not name its AI providers or publish their terms. The litigation paragraph still said we had applied a 1-point penalty because we could not read the docket; we read the docket record on 28 September 2026, which shows the motion to dismiss denied on 30 July 2026, and apply 3 points in our 1 to 10 scores. The comparison table that quotes the training line now shows the corrected wording.
• Canva: we called the no-training position for Teams, Business, Enterprise and Education “contractual” and “permanent”. Canva’s Trust Center says the training control is disabled on those accounts and that it does not use business, team or enterprise content to improve AI-powered features, but also that an admin opt-in is “coming soon”, and the AI Product Terms do not state the tier rule. The opening line, the training section, the plan table and four other lines now describe it as a stated, locked-off default. No rating moved.
• Zendesk AI: two lines kept wording our 3 October correction had withdrawn. The retention list said provider retention was zero “under Zendesk’s contractual arrangements”, and the jurisdiction list named OpenAI and Anthropic as sub-processors “under no-training contractual terms”. Zendesk’s own pages name OpenAI, Google Gemini and ElevenLabs on zero-data-retention endpoints, and models hosted on AWS Bedrock, Microsoft Azure, Groq and Fireworks AI; they do not name Anthropic.
• Fathom: twice described Otter’s Enterprise plan as contractually excluding training. Our 26 September Otter correction found it is a stated default, not a contract term. Both lines now say so.
• Midjourney and Adobe Firefly: both said the Disney/Universal case against Midjourney goes to trial in late 2026. No trial date has been set, as the Midjourney profile’s own 26 September correction records. Midjourney’s profile also dated the Bartz v. Anthropic settlement to March 2026 and gave it as about $3,000 per work for 7 million books, which would be $21 billion. It is $1.5 billion for 482,460 works, about $3,000 each, with final court approval on 20 July 2026.
• Grammarly: the profile still opened with a 13 August “under re-verification” notice, and listed Business at $15 per user per month with a 3-user minimum. The profile was re-checked against Grammarly’s terms on 3 October 2026, and the team plan is Grammarly Pro (1 to 149 seats) at US$12 per member per month billed annually, US$30 monthly. Amended later on 23 October 2026: this entry first gave that price for Business; Grammarly no longer sells a plan called Business.
• Otter.ai: the opening line called the class action a 2026 case, and the profile counted “11 other states” with all-party consent laws. The consolidated cases were filed in August and September 2025. Other profiles on this site counted the states differently, so the line no longer gives a number.
• AI notetakers alert (published 9 July 2026): said Fireflies.ai “was hit with its own class action in Illinois in December 2025” as evidence the category was being tested. That case was voluntarily dismissed by the plaintiff, without prejudice, on 11 March 2026, before the alert was published, and a new class action against Fireflies.ai, Fricker v. Fireflies.AI Corp. (N.D. Ill.), was reported filed in March 2026. The alert’s point stands, and it now carries a dated note giving both. Amended later on 23 October 2026: this entry first said only that the case had been dismissed.
• Free course, Basic module 2 (all nine editions): said business or paid tiers “usually” offer stronger protections, “like not training on your data and shorter retention”. Our own corrections since September show several paid business tiers still train on customer data by default, and we withdrew the shorter-retention claims for several tools. Since 22 October it reads that paid tiers often offer stronger protections, such as not training on your data, but not always, so check the plan you are actually on. In the course since 30 July 2026; four Basic certificates were issued while it read the old way.
• Intermediate and Advanced final tests: one question treated sending personal information to an overseas AI vendor as engaging the Privacy Act’s cross-border rule (principle 12) in every case. A provider that only stores or processes information for you is not receiving a disclosure (section 11); principle 12 applies when the vendor uses the information for its own purposes, such as training its models. Since 22 October the question says so. It was in the Intermediate test by 12 August 2026 and the Advanced test from September; two Intermediate certificates were issued while it was in the test. No Advanced certificate has been issued.
• Microsoft Copilot: we said Microsoft states that the current consumer app does not use your prompts and files to train its foundation models. The Microsoft page we relied on now applies only to Copilot in Microsoft 365 apps for home, and Microsoft’s pages for the consumer Copilot app do not say either way; its Privacy Statement says Microsoft may use your data to train its AI models. The profile now says so. The consumer rating, moved from 3 to 2 of 5 on 22 August on the old wording, is under review.
• Fathom: said Fireflies.ai is not defending a class action. Fireflies’ first case was voluntarily dismissed in March 2026, and a new class action against it was reported filed the same month. The profile now gives both.
• ZoomMate: we said it requires a paid Zoom base plan. Zoom’s pricing pages, read 23 October 2026, do not say so, and Zoom’s community guide describes ZoomMate as an add-on to Zoom Workplace or a standalone subscription, from US$20 per user per month.
• Granola: our summary said having no meeting bot removes the notice other participants would get. Granola offers an optional chat message and an on-screen watermark that users or admins can switch on, and leaves notice and consent to the customer.
• Grok: we said the opt-out from training on your X posts is desktop-only. X’s help centre now gives the setting’s path without limiting it to the desktop website, and the Grok app has its own data control.
• Claude: our summary said Pro and Max carry a training default you have to know to look for. Anthropic’s announcement says users choose at sign-up and can change it later. (Claude is the model that helps write this site; see How This Site Uses AI.)
• Cursor, ElevenLabs and Zendesk: three summaries said more than the vendor does. Cursor states that Privacy Mode is on by default for Teams and Enterprise, not that it is off on individual plans; ElevenLabs’ privacy policy says it may be handling biometric data under some laws, not that voice is biometric data; Zendesk says it trains label models on sanitised customer data, without mentioning a default. The summaries now follow each vendor’s wording.

9 October 2026

• Atlassian alert: we said opting out does not unwind data already absorbed into a trained model, and that some financial-services customers were excluded. Atlassian’s own data-contribution pages (read 9 October 2026) say it retrains any models previously trained on data from an organisation that opts out, and its exclusions name organisations associated with a government entity, not financial services. The post also still told readers to act “before 17 August”; the change went into effect that day, so the post now says so and keeps the opt-out steps.

6 October 2026

• Otter.ai: we said Business and Enterprise customers could set or negotiate their own data retention. Otter’s own pricing comparison (read 4 October 2026) lists custom data retention, audio removal and a recording disclaimer email as Enterprise-only features; Business has no retention control, and no default retention period is published for any plan. The retention and zero-retention sections of the profile now say so.

5 October 2026

• Home page — we listed Perplexity among vendors defending class actions about how their products handle your data. That case had been dropped. Noel v. Perplexity AI (N.D. Cal., 3:26-cv-02803) was voluntarily dismissed on 1 May 2026, as our own Perplexity profile records. The home page now names Otter.ai and Granola only.

4 October 2026

• Canva — we said Enterprise customers could negotiate custom retention terms through a Canva Shield contract. We could not find that in any Canva document. Canva’s Master Services Agreement has no retention clause, its Data Processing Addendum says personal data is destroyed on termination without a period, and its Statement of Technical and Organisational Measures publishes no retention period or customer-set retention control. Canva Shield is an IP indemnification programme. The profile now says so. No ratings are published yet.

3 October 2026

• Salesforce — we said Salesforce’s own use of your data for product improvement was opt-in and off by default. It is on by default. Salesforce’s Spring ’26 release notes say access to Customer Data is on by default (except Government Cloud) and lets Salesforce train global predictive AI models, improve services and do research with it. Admins can now switch it off in Setup. The profile now says so. No ratings are published yet.
• Adobe Firefly — our profile did not say that partner models may keep your generated content under their own terms. Adobe’s Partner Models security fact sheet says partner models “may store generated content before returning it to Adobe” and points to each partner’s terms of use. The profile now says so and lists the partners Adobe currently names, including ByteDance, Kuaishou and HUMAIN. No ratings are published yet.
• Gemini — we said the Business and Standard/Plus editions of Gemini Enterprise app were equivalent on data controls. We had applied Google’s documentation for its developer platform to the Business edition. Business edition’s own help centre says it has no administrator settings for these protections and cannot restrict where data is stored; encryption-key, residency and compliance controls are Standard and Plus features. The profile now says so and no longer states a score. No ratings are published yet.
• Notion — we said Notion does not publish how long its AI providers keep your data on Free, Plus and Business, and that Notion’s own retention is the same on every plan. Notion’s pricing page states 30-day retention by its AI providers on those plans and zero retention by them on Enterprise; Enterprise also adds custom retention settings for Notion’s own storage. The profile now says so, and that Enterprise’s zero retention is at the providers, not at Notion. No ratings are published yet. (Notion names Anthropic as an AI provider; Anthropic also makes the AI model that helps produce this research.)
• Zendesk — we described zero data retention for AI features as contractual and did not separate it from Zendesk’s own storage. Zero retention applies at the AI model providers and is described by Zendesk as practice; its AI Services Addendum contractually limits providers to using your data only to provide and secure the AI features. Zendesk’s own copy is kept under its normal retention, and AI-agent conversations not handed to a human stay in your account until you delete them. The profile now says so. No ratings are published yet.
• HubSpot — we said HubSpot’s ban on its AI providers training on your data was contractual, and we listed AI providers it no longer names. HubSpot states the ban in its help centre; your contract (the Product Specific Terms) says only that AI inputs and outputs are shared with its providers to run HubSpot AI, for content moderation and other business purposes. Its sub-processor list names Amazon Web Services, Google, OpenAI and Hyperdoc, adds Baseten Labs, Bright Data and Exa Labs from 16 October 2026, and does not name Stability AI. The profile now says so. No ratings are published yet. (HubSpot’s AI Model Cards list Anthropic models; Anthropic also makes the AI model that helps produce this research.)
• Claude — our profile described the consumer training opt-out without its exceptions. Anthropic’s Privacy Policy (effective 10 September 2026) says that even after you opt out, conversations flagged for safety review, and materials you explicitly report (for example through the thumbs up/down feedback buttons), are still used for model improvement. The profile now says so and tells readers not to use the feedback buttons on client work. No ratings are published yet. (Anthropic makes Claude and also makes the AI model that helps produce this research.)
• ElevenLabs — we said zero data retention was not offered, and our profile did not say that outside AI model providers handle your content. ElevenLabs’ Trust Center describes a Zero Retention Mode, and its sub-processor list names Google (including Gemini and Anthropic models), Anthropic, OpenAI and, if enabled, xAI as AI model providers — some of which may use the data for training “to the extent permitted under the applicable agreement with the customer”. Both are now on the profile. No ratings are published yet. (Anthropic also makes the AI model that helps produce this research.)
• Microsoft Copilot — we said Microsoft 365 Copilot processes your data entirely within Microsoft’s service boundary. That is no longer true by default. Microsoft’s own documentation says Anthropic is now a Microsoft subprocessor and its Claude models are switched on by default for most commercial customers outside the EU, EFTA and UK, including in New Zealand. Microsoft’s Data Protection Addendum still applies, and for standard models Anthropic keeps no prompts or responses. Admins can switch it off. The profile now says so. No ratings are published yet. (Anthropic also makes the AI model that helps produce this research.)
• Asana — we said zero data retention was available at Asana’s Enterprise tier through its arrangements with AI providers. Asana’s AI FAQ says something different and broader: on every plan, its AI partners must delete customer data once each AI request is completed. That statement is on a help page; we have not found it in Asana’s published contracts. Corrected. No ratings are published yet.
• Grok (SpaceXAI) — we said zero data retention was not offered on the Grok API in the same form as other providers. SpaceXAI’s Enterprise Terms offer it, and on stricter terms than most. The Enterprise Terms (14 August 2026) offer a zero-data-retention API option under which content is deleted within an hour and no logs or backups are kept, even for legal compliance. They also say business content is deleted automatically within 30 days by default. Corrected. No ratings are published yet.
• Granola — we said only Enterprise customers could opt out of Granola training on their data, and that Granola had zero-retention contracts with its AI providers. Both were wrong. Granola’s Platform Terms (19 December 2025) let every customer opt out, and its pricing and security pages say any user can do it in Settings with one click. On Enterprise, training happens only if the customer opts in. We found no zero-retention commitment from its AI providers in Granola’s security page or terms; what Granola does say is that third parties may not train on your data. Both lines are corrected. No ratings are published yet.
• Fathom — we said zero data retention was available at Team Edition and Enterprise through Fathom’s arrangements with its AI providers. We could not find that in any Fathom document. Fathom’s Terms of Service (4 March 2026), Data Processing Agreement (5 May 2026) and help centre say its AI providers may not train on your content; none says they keep nothing. We also said provider retention was “typically zero” on paid plans. Both lines are corrected. No ratings are published yet.
• Intercom — on 21 September we said we could not find Intercom’s position on using your data to train its own AI. Its Terms of Service state it. Intercom’s Terms of Service, effective 8 September 2026, give it the right to use customer data “to provide and improve our Services, including for Intercom AI model training”; the words about AI model training were added in that version. We read the wrong documents on 21 September and missed it. The profile now says what the Terms say. Intercom’s separate contractual bar on its third-party AI providers training on your data is unchanged. No ratings are published yet.
• Replit — we said paid Replit plans are covered by a contract that rules out training, and that no training opt-out exists. Neither is right for Pro. Replit’s self-serve paid plan, Pro, is governed by the Replit Pro Agreement, which lets Replit use your content to develop, train and improve its models and lets you opt out in settings (Settings, Personalization, Model Improvement; new content only). The contractual no-training term we described is in Replit’s Enterprise Agreement. Starter and Core have no documented training switch, as we said. No ratings are published yet.
• Jasper — our 2 October correction went too far: it applied Jasper’s Terms of Service to every plan. They cover the Pro plan, not the Business plan. Jasper’s legal centre says its Terms of Service apply to Creator and Pro customers, and that Business-plan customers are governed by its Software as a Service Agreement (version 5, effective 12 June 2025). That agreement says Jasper “shall not use, nor permit a third party to use, Customer Property to train artificial intelligence models utilized by the Service”, and it contains no licence to display or promote your content. On 2 October we wrote that Jasper’s terms let it use your content to improve its AI models without saying that holds on Pro only. Earlier text on the same page, which said Jasper had a contractual no-training commitment across all tiers, was also wrong in the other direction: it holds on Business only. The page now says which applies to which plan. No ratings are published yet.
• Grammarly — we said Grammarly’s Business and Enterprise plans have zero data retention with its AI providers. We could not find that in any Grammarly document. We checked its Terms of Service, Customer Business Agreement, Data Privacy Addendum, sub-processor list and privacy FAQ on 3 October 2026. What they do say: organisations can delete their content at any time, and Grammarly does not allow its AI providers to train on user content. We also listed only two AI providers; Grammarly’s list names four (Anthropic, Azure, Baseten, OpenAI). And we said Business and Enterprise both had a contractual no-training commitment: that is true of accounts bought through Grammarly’s sales team (Customer Business Agreement s5.4), not of team accounts bought on the website, where training is on until an admin switches it off. All corrected. No ratings are published yet.

2 October 2026

• Jasper — we said Jasper had not stated whether it uses your content for its own AI. Its terms do state it. Jasper’s Terms of Service (effective 30 May 2025) say Jasper uses your input and output “for the purpose of developing and enhancing Jasper’s artificial intelligence models”, under a perpetual, irrevocable licence; only EU and UK users can object. We read this on 28 September and changed our internal rating then, but the profile’s training section, plan table, zero-data-retention section and search description still said Jasper’s own use was “not stated” or described a “no-training default”. All are corrected. Our internal rating does not change today; it already reflected the terms. No ratings are published yet.
• Canva — we said Canva’s paid plans have zero data retention with its AI providers by default. Canva’s own terms say otherwise. Canva’s AI Product Terms (effective 26 June 2026) say your input may be shared with its Technology Partners to provide the feature, and that your Privacy Settings control whether Canva and those partners “may use your data to improve our AI services for everyone”. They do not mention zero retention, and neither do the Trust Center or Canva Shield pages. The retention and zero-data-retention sections are corrected, along with a quote from an older version of the AI Product Terms. Our internal rating for Canva Pro, and for the AI add-on scored at Pro, falls from 4.0 to 3.5, because we had credited those plans with a restriction on AI partners that Canva’s terms do not give. No ratings are published yet.
• Shopify — we said Shopify keeps zero retention with its AI model providers by default. We cannot find Shopify saying so. Shopify’s Terms of Service (1 August 2026), Privacy Policy (7 July 2026), Data Processing Addendum (7 July 2026) and Shopify Magic help pages do not mention zero retention or how long its large language model providers keep merchant data. The retention and zero-data-retention sections are corrected. Our internal rating does not change: we had scored Shopify on its published documents, not on that claim. No ratings are published yet.
• Midjourney — we said images are retained indefinitely. That overstated it. Midjourney publishes no retention period for content you keep, but its Data Deletion and Privacy FAQ (updated 3 September 2026) says deletion completes within 30 days, with a 7-day window to reverse it. Content stays in the public archive until you delete it; it is not kept indefinitely regardless. The retention section is corrected. Our internal rating does not change: we had already scored retention as unpublished rather than permanent. No ratings are published yet.
• Xero — we said Xero had not published whether it uses your data to improve its own AI. Its contract does say. Xero’s Terms of Use (last updated 3 August 2026), section 19, give Xero a licence to use all data you submit “to improve, develop and protect our services” and to “create new services”, with no exception for AI and no opt-out. Our page said Xero was silent on its own use, and its training section opened with “No training on customer data by default”, a line that is true only of JAX’s outside AI providers. Both are corrected. Our internal rating for Xero does not change: we had already scored it as if Xero could use your data with no way to stop it, and the contract confirms that. No ratings are published yet.

1 October 2026

• HubSpot — we said HubSpot does not say whether it trains its own AI models on your data. It does say, and it does train, by default. HubSpot’s Customer Terms of Service (effective 16 September 2026), section 5.3.1: “We may also use Customer Data to train our AI models in compliance with our obligations under the Agreement.” Its help article on AI model training says the setting is on by default and that a Super Admin can switch it off without losing any AI features. Our page said HubSpot’s own position was not stated, and that the only way to opt out also turned off HubSpot’s AI features. Both were wrong: we had described a different setting. The opening line, the training section and the page’s index field are corrected. Our internal rating for HubSpot rises, because a training opt-out that keeps every feature working is a real control, and we had scored HubSpot as if there were none. No ratings are published yet.

30 September 2026

• Grammarly: we understated the protection on business accounts, and two lines still said individuals could not opt out. Our profile described Business and Enterprise as having an “admin-configurable opt-out”. Grammarly’s support article on its Product Improvement and Training control says accounts bought through its sales team, including Enterprise and Education, have training off by default, while multi-user accounts bought on the website start with it on and only an admin can switch it off. The retention, training, plan table and recommendation lines now say so. Two older sentences that still said Free and Premium users had no opt-out, left over from our 13 August correction, are also fixed. No rating changed.

29 September 2026

• Otter.ai: we said the Business plan offered stronger contractual training exclusions. Otter states no Business exclusion. Our profile said Business and Enterprise both offered “stronger contractual training exclusions” to be verified per contract, and the plan table, an FAQ answer and a summary paragraph repeated it. Otter’s privacy and security page says it de-identifies user data before training its models, with no plan distinction, and its Enterprise help article describes an opt-out for Enterprise workspaces only. The Business lines now say that Otter states no separate exclusion for Business, and the plan table shows Business as trained by default. The Enterprise wording is unchanged.

26 September 2026

• Xero — we said Xero is headquartered in Melbourne. Xero says its headquarters are in New Zealand. Xero’s own GDPR page names New Zealand as “where our headquarters are located”. Our profile said Xero was “now headquartered in Melbourne” and listed the primary processor as Xero Limited, Melbourne, Australia, with a Wellington subsidiary as the origin entity. Both lines are corrected. Xero is listed on the Australian Securities Exchange, which is probably where the error came from, but a stock-exchange listing is not a headquarters.
• Otter.ai — our plan table said Enterprise training was “contractually excluded”. We had no contract to support the word. Otter’s own help centre (Enterprise Admin Controls Overview, updated 17 July 2026) says Enterprise workspaces are opted out of AI model training by default, and that an organisation can reverse this through its account manager. That is a stated default, not a contractual exclusion, and the table now says so. The Business row is unchanged: we could find no Business-plan training position from Otter, and its privacy and security page says it de-identifies user data before training its models without distinguishing between plans.
• Fathom and HubSpot — the first line of each profile still carried a claim we had already withdrawn further down the same page. Every profile opens with a one-line answer to its main question. Fathom’s read “No training on any tier, a meaningfully better default than the main alternative in this category.” We corrected that on 3 August 2026: Fathom trains its own models on de-identified meeting content by default, on every tier, which is much the same position as the main alternative rather than a better one. The body of the page was corrected, and so was the description search engines show — but the opening line was not, and it has said the opposite of our own correction for eight weeks. HubSpot’s read “No training by default”, when HubSpot’s commitment covers the AI providers it uses, not HubSpot itself; we corrected that section on 20 September and missed the opening line. Both now say what the vendors say. The Fathom case is the one we hold up to ourselves as a warning — a withdrawn claim surviving in a place nobody re-reads — and it had happened again, on Fathom, in a third place. We found both by checking the opening line of every profile against what each vendor actually commits to; two of the four that claimed no training were wrong.
• Jasper — we told readers Jasper does not train on their content on any tier. Jasper does not say that, and we had already noticed. The one-line summary at the top of the Jasper profile read “No training on any tier, which is a stronger default than the general-purpose assistants offer”, the training section said training is “contractually excluded by default across all tiers”, and the plan table said “No — contractually excluded” against every plan. Read at source on 26 September 2026, Jasper’s trust page says only that “Your company data and outputs are never used to train third-party LLMs” — a commitment about the AI providers Jasper uses, not about Jasper itself. The embarrassing part is that we knew. On 3 September we recognised the claim was broader than Jasper’s words, changed the rating, and corrected the description that appears in search results — and left the summary, the headline and the table on the page itself saying the opposite. For three weeks the profile carried the caveat and the claim it contradicted, a few lines apart. All of it now says what Jasper says. This is the fourth time this month we have found a vendor’s promise about its suppliers presented as a promise about the vendor itself, after Zendesk, HubSpot and Intercom.
• Intercom — we said Australian data residency was “on roadmap”. For the AI features Intercom says the opposite: it is not planned. Two claims are corrected. Our jurisdiction section said “AU Data Residency on roadmap”, and our zero-retention section said Australian workspace data processing was “migrating to AU per Intercom roadmap”. Read at source on 26 September 2026: Australian hosting does exist, but Intercom states that it is “only available to new workspaces on the Advanced and Expert plans”, is “not available through the Early Stage program”, requires a contract, that “Historical conversation data cannot be migrated to the new workspace”, and that “Emails (inbound and outbound) will go through the US due to infrastructure limitations”. And on the AI features specifically, Intercom says Fin AI Agent, Copilot and inbox AI are available to Australian-hosted workspaces but “the hosting will remain in the US”, and that “Regional data hosting for these workspaces is currently not planned.” European workspaces are treated differently: EU customers get those features “with data processed within the European Union”. Why this one matters here: a New Zealand business picking Australian hosting for regional proximity would reasonably assume the AI processing followed. It does not, and our page told readers the move was coming when Intercom says it is not planned. Both the page and the index behind it are corrected.
• Shopify and Midjourney — we scored both as having no litigation while our own profiles described live cases against them, and we named neither case. Our home page says we record the case number, the court and what is actually pleaded. On these two vendors we were not doing that, and the scoring fields behind them read “No penalty. No litigation.” Both are corrected and both ratings have moved down. Shopify: Briskin v. Shopify Inc., N.D. Cal. No. 21-cv-06269-PJH, Judge Phyllis J. Hamilton. It is pleaded that Shopify’s software collects “name, address, email, credit card number, IP address, what you bought, and your geolocation” and installs cookies tracking transactions across the whole Shopify merchant network. On 21 January 2026 most claims were dismissed with leave to amend; Briskin amended, and on 30 July 2026 the motion to dismiss the amended complaint was reportedly rejected. Midjourney: Disney Enterprises Inc. et al v. Midjourney Inc., C.D. Cal. No. 2:25-cv-05275, Judge John A. Kronstadt, filed 11 June 2025. We said a trial date was set for late 2026. No trial date has been set; no motion-to-dismiss ruling and no merits ruling have issued, and the case is in discovery, with all motions due by 23 November 2026. The Warner Bros. action was separate and was consolidated into it on 4 November 2025. We also credited “Court filings” as a source for Midjourney, which implied we had read the docket. We had not, and the false trial date is what that produced.
• And a note on how far we went on each, because the two came out differently. Our rubric moves a rating by three points once a court has let a case proceed past a motion to dismiss, and by one point where a live case exists below that line. Midjourney is plainly the one-point case: no such ruling exists. Shopify probably is not — the 30 July 2026 order suggests the higher threshold is met — but that order rests on a single secondary report and we could not read the docket ourselves to confirm it. So Shopify carries one point rather than three, which almost certainly understates it, and we would rather publish that gap than score a court order we have not verified.
• Replit — we stated New Zealand law more strongly than it is, on our own page, in our own words. The profile said that sending data to Replit “is a cross-border disclosure under IPP 12 of the NZ Privacy Act 2020 and APP 8 in Australia”. Often it is not a disclosure at all. Under section 11 of the Privacy Act 2020, information held by an agent is treated as held by the agency it acts for, and the Privacy Commissioner says that “in most circumstances, you’ll be responsible for the personal information that you put in the cloud – not the cloud service provider”. A provider that only hosts or processes your data on your behalf is not the recipient of a disclosure. It can become one where the provider uses your inputs for its own purposes — and training is exactly that, which is why the answer differs between Replit’s free tier, where our own training section records that training is enabled, and a paid or Enterprise contract, where the Commercial Agreement limits use to providing the platform. And where Principle 12 does apply it is not a prohibition: it is met by any one of several grounds, including reasonable grounds to believe the recipient must protect the information with comparable safeguards, which can be achieved contractually, and there is an exception where compliance is not reasonably practicable. The reference to Australia’s APP 8 is withdrawn rather than replaced, because APP 8 has its own tests and we have not read them at source for this profile. The corrected passage quotes the Commissioner, cites the section, scopes the answer to the tier, and carries the general-information note that every legal statement on this site is supposed to carry. It did not, which is the part that should not have happened — the rule requiring it is our own.
• Xero — we said Xero offers AU/NZ data residency. Xero’s own pages say it hosts in the United States. The profile stated that Xero offers AU/NZ data residency and processes AU/NZ data in AU/NZ regions, and recommended that regulated-industry readers use “Xero with appropriate AU/NZ data residency”. All of that is withdrawn. Read at source on 26 September 2026: Xero’s own GDPR page says it uses “Amazon Web Services (AWS), a top-tier, third-party data hosting provider with servers located in the US to host our online and mobile services”, and names no Australian or New Zealand data centre and no regional hosting option. Xero’s New Zealand subprocessor list gives United States for Amazon Web Services, Google Cloud Platform, Microsoft Azure, Snowflake, Anthropic and OpenAI; the only Australian entry is Superchoice, a payroll service. One detail cuts the other way and is worth knowing: the same Xero page says “some of our EU customers’ data is processed in New Zealand”, so New Zealand appears as a place Xero processes other customers’ data rather than a place it keeps yours. This is the most New Zealand–specific error we have published, on the tool most New Zealand small businesses use, and it ran in the vendor’s favour — it told readers their financial data stayed onshore when Xero does not say that. If your business needs its financial data held in New Zealand or Australia, Xero as published does not give you that. How this one was found, because the process matters: it was raised internally on 24 September, and for two days this page carried it as “under review” rather than as a correction, because we had not yet read Xero’s pages ourselves and would not publish a second-hand finding as our own. The flag went up immediately; the correction followed the source read.

25 September 2026

• Cursor — we named a plan that does not exist, and we described a default as if it were locked. The profile said “Business and Enterprise — Privacy Mode is forcibly enabled by default. Organisation admins enforce it; individual users cannot disable it.” Two things were wrong. Cursor has no plan called Business — its plans are Hobby, Individual, Teams and Enterprise, and the tier at $40 per user per month we were describing is Teams. And we collapsed a default into an enforcement: Cursor says Privacy Mode “is enabled by default for all team members”, and says separately that admins “can enforce it organization-wide” so members cannot disable it. Those are different promises. Under our wording a reader would believe they were protected by the plan; in fact a team member can switch Privacy Mode off until an admin turns enforcement on. That is the kind of error that matters most on this site, because it tells someone a job is already done. Twelve references to the wrong plan name and four to the wrong framing are corrected, on the page and in the index behind it. The honest history of this one is worth recording. Twice before, we searched for this claim, failed to find it, and wrote in our own notes that it could not be sourced — while leaving it published. It turned out to be substantially real and badly worded, and the page that states it is the one our own profile had cited all along. Searching for a claim, failing, and leaving it up is not the same as checking it.

21 September 2026

• Intercom Fin — we claimed a no-training position that Intercom does not state about itself. The profile read “NO TRAINING ON CUSTOMER DATA BY DEFAULT across all Fin tiers”, and the evidence offered beneath it was a quote from Intercom’s Additional Product Terms: “Intercom contractually restricts Third Party AI Providers from using Customer Data for training or otherwise improving Third Party AI Provider’s services.” That term binds Intercom’s suppliers, not Intercom. We read three Intercom documents at source — the Additional Product Terms, the Security at Fin help page, and the privacy blog post this profile cites — and found no equivalent commitment that Intercom itself will not train on Customer Data. To be clear about what we have and have not established: we have shown that our claim was unsourced, not that Intercom trains on your conversations, and we are not saying that it does. The supplier commitments are genuinely strong on their own terms. The headline now says what the evidence supports. This is the third time in a month we have found a first-party no-training claim on this site resting on a third-party instrument, after Zendesk AI and HubSpot AI.
• Asana AI — our correction of 20 September was itself wrong, and went further than the facts. Yesterday we changed the Asana profile to say “ASANA’S OWN POSITION IS NOT STATED”, on the basis that we could find no published statement in which Asana committed that it does not train its own models. Asana does state it, on its published AI FAQ: “Asana does not use customer data for training and has contractual agreements with its AI partners that prohibit the use of customer data to train or improve their AI models and services.” So for one day this site denied a commitment the vendor had made. We cannot establish whether that sentence was present on 20 September and we missed it, or was published since, so we are recording it here as an error rather than in the alerts as a vendor change — where we cannot tell, the entry belongs on the page that counts against us. The profile now carries Asana’s statement in full, and notes that the word contractual in it attaches to Asana’s AI partners: Asana’s own position is a help-page statement it can revise, not a term of your contract. A second fault surfaced with it. Yesterday’s correction was written to the page a reader sees but not to the field index behind it, so for a day the two said opposite things — the exact reverse of the 20 September entry above, where the fix reached the index and not the page. Both now match. We are not going to claim this is solved. Twice in two days a correction reached one of the two places it needed to reach, in opposite directions, and both times it was found by accident rather than by a check. A reconciliation of every profile’s index against its page has been requested and is not yet built, and until it is, this class of fault can recur without anyone noticing.

20 September 2026

• Our corrections of 13 September did not reach readers until today, and this page said otherwise. On 13 September we recorded corrections to the Zendesk AI and ElevenLabs profiles here. Those corrections were written to the field index the site uses for listings and schema, but not to the page content a reader actually sees. For seven days this page stated that the Zendesk claim was corrected while the Zendesk profile still carried it, word for word. Both corrections are now live on both profiles. The mistake was ours and it is exactly the kind this page exists to surface: the change was checked by re-reading the field we had just written, rather than by opening the page. We now verify corrections by reading the rendered page.
• ElevenLabs — the withdrawn framing survived in three more places on the same profile. The 13 September correction fixed one bullet in the retention section. The profile introduction, the closing line of the training section, and the third of the three plain-English takeaways all continued to say, in various words, that ElevenLabs retains rights to the models trained from your voice indefinitely and that this puts them beyond your reach. The perpetual licence is real and remains described; what was wrong was presenting it as an inability to have a voice model deleted, when the Terms provide for deletion on request. All three now match the corrected position.
• Asana AI — we claimed a blanket no-training position that Asana does not state. The profile read “NO TRAINING ON CUSTOMER DATA BY DEFAULT across plan tiers”. The only evidence beneath it concerned Asana’s sub-processor agreements, which is a commitment about the third parties Asana sends content to. We have found no published statement in which Asana commits that it does not train its own models on customer content. The headline now says what the evidence supports, and the profile records that Fireworks AI was added to Asana’s sub-processor list on 31 August 2026 with no statement as to whether the no-training default extends to it.
• HubSpot AI — same error, same shape. We read “NO TRAINING ON CUSTOMER DATA BY DEFAULT across plan tiers”, while the sentence we quote from HubSpot’s own FAQ commits only that HubSpot does not permit the AI service providers it engages to use customer data for model training. Corrected. We have also added Hyperdoc, Inc. to the named providers: it appears on HubSpot’s published sub-processor page for meeting recordings and transcripts, and we had named only OpenAI, AWS and Google.
• Salesforce Einstein / Agentforce — same error again. The profile opened “NO TRAINING ON CUSTOMER DATA BY DEFAULT”. The sentence we quote from Salesforce scopes its guarantee to the underlying third-party large language models, and the profile separately records an optional first-party opt-in for product-improvement data sharing. That programme is off by default, which is the right default, but its existence means the honest description is no training by default with a clearly labelled opt-in, not an absence of training.
• Linear AI — a sourcing note, not a correction. Linear’s profile does state that Linear itself does not train on customer data, so it is unlike the three above. But that statement is not tied on our page to a quoted sentence from a named Linear document, and we have marked it as such. Our internal scores for Linear are recorded as reported rather than verified for the same reason.

13 September 2026

• Zendesk AI — we said Zendesk does not train on customer data. It does. Our profile’s headline read “NO TRAINING ON CUSTOMER DATA BY DEFAULT across plan tiers”. Zendesk’s own published AI Data Use Information says it takes measures to sanitize customer data “before using it to train any of our models that could be used by other customers”. That is training on customer content, by default, and it is not limited to the higher tiers as a note we added on 3 September suggested. The claim was wrong from the day the profile was written and it stood for over a year. The distinction we missed is real and worth stating in Zendesk’s favour: Zendesk’s generative features genuinely are not trained on customer content, and we were reporting that part accurately. It is Zendesk’s own non-generative models — the ones that output labels such as triage, intent and sentiment rather than text — that are trained, on data that is first stripped of identifier fields, passed through an NLP scrub and tokenized. In the same field we also said Zendesk did not name its AI providers individually. It names seven. Both statements are corrected, and the safety score for all four Zendesk tier-units has moved from 8.5 to 6.5 — a fall driven by the training finding and partly offset by the provider naming, which is better than we had credited.
• ElevenLabs — our retention field contradicted our own training field. The retention section said that once trained, the voice model derived from your recordings is retained under the February 2025 perpetual licence clause, and that “the derived models persist”. The training section on the same page had already been corrected to say that this overstated the position and that User Voice Models can be deleted on request. A reader who reached the retention field first got the superseded answer. The retention field now matches: you can request deletion of your voice models, and what genuinely cannot be undone is the contribution your data made to ElevenLabs’ foundation models before you opted out, because the opt-out is prospective only. No rating changed — the score rested on the published three-year ceiling, which is unaffected.

6 September 2026

• Zoom — we over-corrected yesterday, so here is the correction to our correction. On 5 September we told you ZoomMate is “no longer included in the Zoom plan you already pay for”. Reading Zoom’s Workplace pricing page the next day — the one we had admitted was still carrying an 11 August reading — showed that too blunt. A Workplace Pro subscriber at NZ$24.99 per user per month does get substantial AI without buying anything else: unlimited meeting summaries, unlimited in-meeting questions, unlimited AI note-taking, agentic search inside Zoom. Even free Basic carries a metered amount. What the paid ZoomMate licence adds is the agent layer — credits, custom agents, deep research, reach into third-party sources. So the accurate statement is narrower than the one we published: the AI is not all behind a paywall, but the agent product is. The profile now carries both corrections rather than only the tidy final answer, and the plan table has been rebuilt with the Workplace prices we had not re-read. The rating has still not been moved.

5 September 2026

• Zoom — we have been rating a product under a name Zoom no longer uses, and our scope note about it was wrong. Our profile is titled “Zoom AI Companion” and uses that name throughout. The page we cite as the product’s home now reads “Meet ZoomMate, your AI teammate”, and the words “AI Companion” do not appear anywhere on it. We read that page in a browser on 5 September 2026. We have also withdrawn our scope note of 3 September, which said the rated product does not exist on Zoom’s free plan: Zoom now lists ZoomMate Basic as included free with Workplace Basic, metered rather than absent — three hosted meetings of summaries a month, twenty AI queries, three uses of AI note-taking, ten workflow runs. We cannot tell you whether Zoom changed this or whether we read it wrong, and we are not going to guess. The rating has not been moved and is being re-checked. Renamed the same day. The profile is now titled “ZoomMate (Zoom)” and the new name is used throughout it and in the six other pages that referred to the product. Two things were deliberately left alone: the titles of the source documents we cite, because a citation has to say what the document actually says, and this page and the dated alert of August 2026, because a correction record that rewrites itself is not a record. The profile also carries a caution that its plan-inclusion detail has not yet been re-verified.
• Slack — our scope note of 3 September was wrong and is withdrawn. It said this rating covered a product that does not exist on Slack’s free plan. Slack’s pricing page on 5 September lists “Basic AI” on the Free plan: conversation summaries, Slackbot, workflow generation, AI search, daily recaps and file summaries. Again we cannot say whether the page moved in two days or whether we misread it, and we would rather tell you that than pick the version that flatters us. The rating has not been moved and is being re-checked.
• Zoom — a second, larger correction the same day: ZoomMate is not included in your Zoom plan, and we had been saying it was. Having flagged the rename, we went to Zoom’s own ZoomMate pricing page in a browser rather than trusting its marketing page, and found the profile wrong in a way that costs money. We had described the product as bundled into eligible paid plans “rather than sold as a per-seat add-on”, and drew a favourable contrast with Microsoft 365 Copilot on exactly that basis. Zoom now sells it per user: NZ$27.67 per user per month, billed annually as a single upfront charge, carrying 2,200 AI credits per user per month. Below it sits a metered free tier, ZoomMate Basic, included with Workplace Basic. For a ten-person firm that believed AI came free with its Zoom subscription, that is roughly NZ$3,300 a year it did not know it was looking at. We have corrected the summary, the deck, the plan table, the recommendations and the metadata, each with a dated note saying what the line used to claim. The rating has not been moved: price is not a privacy fact and the training and retention positions are unchanged. But the “it is already included, you may as well use it” framing ran through the whole profile, and that framing was doing work.
• What we checked at the same time, and what held. Two of the four scope notes we added on 3 September survive the re-check. Asana stands: its pricing page still shows no AI features on the free Personal tier, with AI Studio credits beginning at Starter. HubSpot substantially stands: the free tier carries no credit allowance while Starter, Professional and Enterprise each do, and the AI agents run on those credits — though we could not find the term “Agent Hub” on the pricing page we read, so treat that particular wording as ours rather than HubSpot’s until we confirm it.

3 September 2026

• Jasper and Shopify Magic — two ratings rested on a claim broader than the vendor’s own, and both moved. We rated both 2 of 5 partly on the basis that training on customer data was excluded across all tiers. Neither vendor publishes that. Jasper commits that data is “never used to train third-party LLMs” — a commitment about third-party model providers, not about Jasper’s own systems. Shopify commits that it “doesn’t use any merchant’s store-level data to power Shopify Magic for other merchants” — a commitment about cross-merchant use, and its Magic documentation does not address shopper data. Both ratings moved to 3 of 5. The error was ours in both cases: we wrote a broader claim than the vendor makes. Neither vendor’s conduct has changed and nothing new is alleged about either.
• Search-result descriptions — Fathom, Perplexity and Jasper. Three profiles carried a claim in their search-result description that the page itself had already corrected. Fathom’s said its training posture was “structurally better than Otter’s” — retracted on 3 August and swept from the page body on 2 September, but left standing in the description for a month. Perplexity’s said the April 2026 lawsuit was “unresolved”, four months after it was dismissed. Jasper’s asserted a “no-training-across-tiers default” that the profile had, that same morning, declined to assert — because Jasper’s published commitment covers third-party providers rather than Jasper itself. These are the sentences shown under our pages in search results. For anyone who found us that way, they were the claim, whatever the page went on to say. They do not appear on the page, which is why every proof-read missed them. All three corrected 3 September 2026. The other twenty-six vendor descriptions were read the same day and none carried a claim its page had corrected.
• New Zealand consent law — Granola, Fathom, Zoom AI Companion. We described New Zealand as a jurisdiction requiring the consent of all parties to record a conversation, and on the Granola profile attributed that requirement to the Privacy Act 2020. Both were wrong. New Zealand is a one-party consent jurisdiction: under section 216B(2) of the Crimes Act 1961 a person who is party to a private communication may lawfully record it. The Fathom and Zoom profiles additionally listed the EU under GDPR as an all-party consent regime; the GDPR is a lawful-basis regime, and consent is only one of its six bases. Our advice — disclose before recording — has not changed, but the reason we gave for it was wrong, and a reader who checked would have been entitled to discard the advice along with the reason. The obligations that do apply in New Zealand are information privacy principle 3 of the Privacy Act 2020 and the employment duty of good faith. Wording corrected 20 September 2026. This previously paraphrased principle 3 as a flat requirement to tell people what you are collecting and why. It is not. The Office of the Privacy Commissioner states that an organisation collecting personal information directly "must take reasonable steps to make sure that the person knows" why it is being collected, who will receive it, whether giving it is compulsory or voluntary, and what will happen if the information is not provided, and the principle carries exceptions. Read at privacy.org.nz on 20 September 2026. General information, not legal advice.
• Cruz v. Fireflies.AI Corp. — Otter.ai and Fathom profiles. We presented this case as a live class action and placed it in the Northern District of Illinois. It was filed in the Central District, and it was voluntarily dismissed on 11 March 2026 before any responsive pleading, so nothing was decided. We described a company as currently defending a case that had ended nearly six months earlier.
• ChatGPT retention figures. Retention detail corrected against two contemporaneous reports, both read that day.
• GitHub Copilot — an unsourced breach note. The profile carried a one-line note asserting a confirmed May 2026 breach affecting credential stores, with no source, and an instruction to itself to re-verify the scope later. It was vaguer than the facts and vague in a direction unfavourable to GitHub. Removed.

2 September 2026

• Otter.ai — a court ruled and we did not know. This profile said the motion to dismiss was fully briefed and pending and that no ruling had issued. That was accurate when last checked on 3 August 2026. The order landed ten days later, on 13 August 2026, granting the motion in part and denying it in part, with federal wiretap, California Invasion of Privacy Act and Illinois biometric claims surviving. We went on publishing the old position for three weeks, in four separate places, because our monitoring watched vendors’ privacy pages and not court dockets.
• Perplexity — a dropped lawsuit published as live, and the rating moved. We described the case as unresolved and at an early procedural stage, named it Doe v. Perplexity, and said the plaintiff was proceeding anonymously. The court had denied pseudonymity, the plaintiff was named, and the case was voluntarily dismissed on 1 May 2026 and terminated on 6 May — sixteen days before this profile’s own last-verified date. The consumer rating moved from 4 to 3 as a result.
• Zoom AI Companion — the page contradicted itself for three weeks. One line said “standard 30-day retention applies”, contradicting a correction published in the retention section of the same page on 11 August 2026. Both claims sat on the page at once.

August 2026

• 24 August — Fathom. An earlier correction had been written into one of the two places the claim lived and not the other, so readers of the page never saw it. Completed.
• 22 August — Microsoft Copilot, and the correction moved the rating. We rated the consumer product 3 of 5 on the basis that training was on by default. Microsoft’s documentation states otherwise.
• 15 August — Replit. We continued to publish a commitment to delete data within 30 days of an account-deletion request, and a clause about de-identified data, after both had been removed in the policy revision dated 3 August 2026. The vendor changing its terms is a change; our still publishing the old ones afterwards is an error, and it belongs here.
• 13 August — Grammarly. Corrected on the availability of the training opt-out for Free, Premium and single-user Pro accounts.
• 5 August — Replit. Source list corrected.
• 3 August — Fathom. The profile stated that training was contractually excluded across all tiers. That was wrong: it conflated the sub-processor exclusion with Fathom’s own first-party training on meeting content. The error was ours, not a change by the vendor.

Recommendations withdrawn because a vendor was sued, or because its terms changed, are not listed here. Those are changes rather than errors, and they appear on the profile and in the alerts. This page is for the times we got something wrong.