Q3 2026 Policy Watch Report: Two Chinese Labs Changed Who Holds Your Data
Five lab policy changes held up against the source text, one of our own July calls did not and has been reversed, and the Frontier Privacy Index closes the quarter at 3.56.
From July 1 through September 30, the Frontier Watch monitor flagged 42 changes across the ten labs we rate. Read against the source text, five were real changes to what a lab says it does with your data. The two with the largest consequences involve Chinese labs: Moonshot AI (Kimi) moved its privacy policy under a Beijing company in August, and Alibaba (Qwen) changed the company that acts as its data controller in September. Neither change is reflected in the published scores yet.
The Frontier Privacy Index, the mean of our ten consumer-product ratings, closes the quarter at 3.56 out of 10. One rating moved on a policy change this quarter (OpenAI, in July). We have since reversed that call, and the correction is below.
How to read the count
The monitor captured about 360 policy documents this quarter. It flagged 42 lab-policy changes (34 material, 5 critical, 3 minor) and 139 events across the app and host layer, of which the classifier called 10 substantive. Most of that volume is duplication. Anthropic's single July 8 update was flagged repeatedly through the summer. Several labs serve different versions of one web address depending on language or region, and our capture flipped between them from one day to the next. Every change below was confirmed by reading the policy text itself: the text our monitor captured, plus a live read where the site allowed one.
Moonshot AI (Kimi): the controller moves from Singapore to Beijing
Severity: critical
Through July, Kimi's privacy policy (last updated July 7, 2025) named Moonshot AI PTE. LTD. as provider and controller. The version our monitor captured on August 3, effective August 4, names Beijing Moonshot Technology Co., Ltd. and says it complies with China's Personal Information Protection Law and Data Security Law.
The new text is specific about where data lives: "We store your personal information within the territory of the People's Republic of China," and it will not be provided outside China without separate consent and compliance with national law. Model training on user content is disclosed, on content that has been de-identified. Opting out means contacting customer service with a request and passing identity verification. The old policy said data may be transferred to and stored on servers outside the user's country of residence. A further revision took effect August 31, adding a link to a children's privacy policy.
This is a policy read. We have not tested where Kimi's traffic actually goes.
Alibaba (Qwen): a different Singapore company becomes the controller
Severity: critical
The version our monitor captured on September 10 (last updated May 19, 2026) described services provided by Alibaba Cloud (Singapore) Private Limited. The version captured September 21, dated effective September 7, names NTH POWER GLOBAL TECH SINGAPORE PTE. LTD. as the data controller. It extends the policy to the Qwen mobile app and a new desktop client, adds a warning against uploading sensitive data such as biometric data, and describes advertising and analytics partners that receive device information.
Data handling stays in Singapore on paper. The policy says user input is sent to Alibaba Cloud (Singapore) for processing as a data processor, and that personal data is collected, processed and stored in Singapore. It allows de-identified content to be used for training and says raw input is not used to train models directly. The policy does not say how NTH POWER relates to the Alibaba corporate group, and we are not going to guess.
Anthropic: the July 8 update adds ID verification, study data and connected-service terms
Severity: material
We compared Anthropic's July 8 policy with the January 12 version, both from Anthropic's own archive. The July 8 version adds:
- Verification Data. If you choose to verify your age or identity, Anthropic collects an image of a government ID and the information on it, your image in photo or video form, and facial geometry templates, which the policy notes may count as biometric data in some jurisdictions.
- Study Participation Data. Responses to Anthropic research studies, surveys or interviews, combined with account data and possibly with other account data for aggregated analysis.
- Connected services. Claude may send your inputs, outputs and instructions to third-party services, and some outputs can act outside the product, such as sending messages or modifying files. The policy makes you responsible for having the rights to what you connect.
- Softer feedback language. Rating a response used to mean "we will store the entire related conversation." It now says "may."
The September 10 revision is much smaller. It adds a South Korea addendum: identity verification through Persona (data goes to the United States, retained three years by the recipient), fraud checks through the Korea Credit Bureau, and a new domestic representative. Nothing else in the text changed. Our alerts for September 16 and 22 described that revision as if it added biometrics and study data. Those were July items, so those alerts overstated it.
OpenAI: one policy became three
Severity: flagged critical; see the correction below
OpenAI now publishes three separately dated consumer privacy documents at the same address: a US policy (updated May 18, 2026, unchanged all quarter), a general policy with a Korea addendum (updated June 4, then July 30, then September 10), and a Europe policy (updated August 24). The documents point to one another.
The sentence about advertising is in the general and Europe documents: "For Free and Go users, we also collect information about your ads history and interests." The general document later softened "collect" to "may collect." The US document contains no such sentence. It does disclose that OpenAI receives data from advertisers to measure ads shown to Free and Go users, offers advertising controls in account settings, and describes sharing that counts as targeted advertising under state privacy laws. One capture on September 17, dated September 10, lacks the ads-history sentence, while later captures show the July 30 text again, so we cannot say which general version is current. OpenAI's site refuses automated reads, so all of this rests on our monitor's captures rather than a live reading.
Meta: agentic AI added
Severity: material, with a limit on what we verified
Meta's privacy policy, effective July 23, opens with a notice that it now includes "details about agentic AI capabilities and how we personalise your AI experiences." We confirmed the notice on the live page. We could not read the body sections behind it, so we can confirm that Meta made a change and cannot yet say what it says.
No substantive change: Google, DeepSeek, xAI, MiniMax and Zhipu
Google's policy carries a May 26, 2026 effective date, which predates the quarter. We confirmed that date on the live page and found no substantive change behind our five Google flags. DeepSeek's English policy has read "Last Update: Feb 10, 2026" all quarter. The four DeepSeek flags came from Japanese and Korean versions served from the same address. The Korean one, dated May 6, 2026, discloses more than the English one, including named transfer recipients such as Beijing DeepSeek Artificial Intelligence Co., and it is a regional document. The global policy did not change. The monitor flagged nothing at xAI, MiniMax or Zhipu.
Apps and hosts
Of 139 app and host events, these held up on reading:
| Layer | Change | Rating now |
|---|---|---|
| Fathom (app) | Policy updated August 16 adds a Voiceprints section: opt-in iOS users have voiceprints collected for speaker recognition. The policy says other users' voiceprints are not collected. | 4.8 |
| Manus (app) | Policy updated September 28 names Financial Data, Health Data and Third-Party Credentials as categories, limited to what users provide or connect. | 3.2 |
| Janitor AI (app) | Policy updated September 17 adds a section on Google sign-in through Supabase: account ID, email, name, photo, auth tokens. | 1.7 |
| Notion AI (app) | Removed its Data Privacy Framework section covering EU, UK and Swiss transfers (by mid-August). | 2.0 |
| Replit (app) | Privacy policy restructured by August 5, with a new category for collaboration and group data. | 3.4 |
| OpenRouter (host) | Added a Files API data category (uploaded files retained until deleted) and extended in-region routing to the US and to Business plans. | not rated |
Where the scores stand
| Lab (consumer product) | Privacy score |
|---|---|
| Anthropic (Claude) | 5.3 |
| OpenAI (ChatGPT) | 5.1 |
| Google (Gemini) | 4.9 |
| Alibaba (Qwen) | 4.1 |
| DeepSeek | 3.9 |
| xAI (Grok) | 3.8 |
| Meta (Meta AI) | 3.1 |
| Moonshot AI (Kimi) | 2.8 |
| Zhipu (GLM) | 1.4 |
| MiniMax | 1.2 |
The index moved from 4.6 in June to 3.56 on July 12. Those two readings are not comparable. June was a plain average of six providers on the old seven-factor rubric, and July was the first reading on the rebuilt six-dimension rubric across ten. The one policy-driven move this quarter came on July 13, when OpenAI's rating fell from 5.1 to 4.8 and the index went from 3.56 to 3.53. We reversed it on October 3, which restored 5.1 and 3.56.
Among the 14 rated apps, Microsoft Copilot leads at 5.7, then Cursor (4.9), Fathom (4.8), Grammarly (3.9), Replit (3.4), Manus (3.2), GitHub Copilot (2.3), Notion AI (2.0), Janitor AI and Perplexity (1.7 each), Poe, Comet and Jasper (1.6 each), and Character.AI (1.5). Across all 24 backend chains, the Stated and Effective scores match, meaning the app's own policy is the weakest link, with one exception. Comet's Kimi option scores 1.6 stated and 1.3 effective, because Moonshot's terms sit below Comet's.
What the scores do not reflect yet
The Moonshot (2.8) and Alibaba (4.1) ratings were set on July 10, before either controller change. They have not been revised. Part of the reason is visible in the public breakdown: the jurisdiction dimension sits at 5.0 for all ten consumer ratings, which the rubric defines as a standard commercial jurisdiction with lawful access through due process. With that dimension flat, a change of legal entity gives a score nothing to move. We will publish any revision here.
Corrections and limits
Our July 13 OpenAI report. We wrote that OpenAI's effective date moved from May 18 to June 4 and that no advertising category existed in the previous version. The captured text shows those dates belong to two different documents, the US policy and the general policy, and our monitor had compared one with the other. The sentence we quoted is real and appears in the general and Europe policies. It is not in the US policy. On October 3 we restored OpenAI's rating to 5.1, the score set from the US policy, and the index to 3.56. The general and Europe policies do disclose ads-history collection for Free and Go users, so if we rate the non-US documents separately, that finding will be scored there.
Board labels. The public status board currently shows Anthropic and DeepSeek as "Changed, Material." Anthropic's latest revision is a Korea-only addendum, and DeepSeek's flag came from a regional-language version. It shows Moonshot as "Stable" after a critical change in August. Those labels do not match what we found in the text.
Reach. Findings above come from policy text. We did not test what any company does with data in practice. Any rated company can dispute a score with evidence at info@q16pbc.com, and evidence-backed corrections are re-scored within ten business days.
Full ratings and the six-dimension breakdown for every lab and app are public at watch.q16pbc.com. Earlier coverage: OpenAI Adds Advertising Data Collection to ChatGPT's Free Tier (July 13, see the correction above).