Do contractors read AI chats? The Scale AI complaints, and who reviews yours

The Morse teamPublished October 29, 20269 min read

Yes, some AI chats are read by contractors. As of October 29, 2026, OpenAI, Google, and Meta each say on their own pages that outside service providers or vendors may review what you type, Microsoft says humans may review Copilot conversations, and some of the rating and labeling is done by freelancers hired through data platforms such as Outlier, which Scale AI runs. On October 27, the privacy group noyb filed five GDPR complaints accusing Scale of tracking those freelancers while they work. Scale calls the claims "unfounded and inaccurate."

On October 29, 2026, we read noyb's release and all five complaints, Outlier's privacy policy and terms, Scale's response as The Register reported it, and the privacy and help pages of OpenAI, Google, Meta, Anthropic, xAI, and Microsoft. Everything in the complaints is an allegation no regulator has ruled on. Google is the most open about outside review. It names service providers as reviewers, says it disconnects chats from your account before they get them, and says how long it keeps what they saw.

What did noyb file against Scale AI?

Five complaints, all dated October 27, 2026: one each with the data protection authorities of France (CNIL), Italy (Garante), and the Netherlands (Autoriteit Persoonsgegevens), and two with Austria's Datenschutzbehörde (noyb). noyb filed them with Intérêt à Agir, DiPLab, and Reversing.Works, under case numbers C109-01 to C109-05. The complainants are Outlier workers, and noyb says Outlier uses "more than 900,000 'freelance' workers."

The complaints make three charges:

• Trackers. noyb says that while workers are on Outlier, trackers from TikTok, LinkedIn, Facebook, the survey tool Sprig, and Datadog send their IP addresses, contact details, identifiers, and the page they're on to third parties "every few seconds." The French complaint lists 18 third parties found in its complainant's browser logs. The complaints describe Fullstory as a session replay tool that "records mouse movements, clicks, keystrokes," and the Dutch complaint says Sprig's tracker collected the worker's "full name, phone number, email address, unique identifiers, device information, pay information and page visited."
• Biometric checks. Workers must verify their identity through Persona again and again, with an ID upload and a video of their face. noyb says about 70,000 ID photos stored with Persona were exposed by hackers in October 2025.
• Ignored access requests. The Dutch complainant asked Outlier for their data and followed up on January 28, 2026. The complaint says no answer came.

noyb asks each authority for "a full investigation" and suggests a fine. Scale told The Register that "the report's claims are unfounded and inaccurate," that "Outlier's practices are disclosed to contributors and require their consent," and that it uses "standard tools to verify identities, prevent fraud, protect confidential project information, and ensure accurate payment."

What do Outlier workers do?

They write prompts, rate chatbot answers, and correct them. The Dutch complaint describes the chain: "A client sends data to the defendant, which (using algorithms) assigns a task related to that data to an individual," and the platform then assigns "the task of correcting them and confirming the quality of the work to another individual." That complainant's tasks involved "comparing conversational AI models by rating different responses provided by the LLM to a given 'prompt.'"

noyb names Outlier's best-known customers as Meta, OpenAI, xAI, Google, Anthropic, and the US government, and writes that "ChatGPT, Gemini and Grok are also trained by actual people who manually review or annotate their output." The customer list is noyb's. None of the AI companies' pages we read names its labeling firms. Outlier's privacy policy says it uses worker data to "develop, research, and test improvements to our and our Customers' products and services," and keeps identity verification data "for up to two years after your last interaction." Its terms, effective May 11, 2026, bind workers to hold task material "in the strictest confidence."

Are the complaints about your chats?

No. They're about the workers' own data: their devices, faces, and IDs. None of the five says a chatbot user's conversation reached Fullstory, LinkedIn, or any other tracker, and none says what was on the workers' screens when the logs were captured. We found nothing showing that it happened.

The complaints do show how many hands a training task can pass through. A chat an AI company decides to use can go from its servers to a labeling platform, then to a freelancer who is ID-checked and bound to confidentiality, then to a second freelancer who checks the first, inside software that, by noyb's account, carries analytics and advertising code. The route depends on the project. Scale told Business Insider in August 2025 that "at no point does the user generated data leave the customer's platform," and that many projects run on "the customer's labeling platform" (Business Insider). The news reports on the complaints that we read cover only the workers' side. For trackers inside the chat apps themselves, see do AI chatbots share your data with trackers.

Who reviews AI chats? What six companies say

The table covers personal accounts. Each cell quotes or sums up the company's own page, read October 29, 2026.

Google is the only one of the six whose page says who the outside reviewers are, what it strips first, and how long it keeps what they saw. Two don't mention contractors as reviewers. Anthropic's privacy center describes access only by its own Trust and Safety staff, though 404 Media reported in September 2026 that "Anthropic confirmed to 404 Media it is also using human review to improve its models." xAI's FAQ speaks only of "authorized personnel." The company behind Grok calls itself SpaceXSI on these pages.

What have contractors reported seeing?

Personal details, in reporting about Meta. On August 6, 2025, Business Insider reported that four contractors hired through Outlier and Alignerr "routinely encountered" Meta AI chats containing "full names, phone numbers, email addresses, gender, hobbies, and other personal details." One, who could finish as many as 5,000 tasks a week, estimated personal data turned up in "60% to 70%" of them. On an Outlier project called PQPE, each chat came with a list of facts about the user drawn from earlier chats and "social profile activity," and contractors couldn't reject chats containing personal information. Business Insider matched one chat's user details to a Facebook profile in under five minutes.

Meta told Business Insider it "intentionally limit[s] what personal information they see." All of this is reporting about projects in mid-2025, not a finding by any court or regulator. For OpenAI's reported contractor program, see our Project Lily explainer.

Does de-identification keep reviewers from knowing it's you?

Partly. Google disconnects chats from your account before service providers see them, and Anthropic separates feedback from your user ID before training. Both strip the account, but a chat that names your town, your employer, and your children can point to you without a username, which is how Business Insider found that Facebook profile.

De-identification also cuts against deletion. OpenAI and xAI both say a chat that has been de-identified and disassociated from your account falls outside their 30-day deletion, and Google keeps reviewed chats for up to three years after you delete your activity.

How do you keep your chats away from reviewers?

Turn off training before a sensitive chat, and know which kinds of review stay on:

• ChatGPT: turning off Improve the model for everyone in Settings, Data controls stops review to improve the model, not review for abuse, support, or legal matters. If you rate a reply, "the entire conversation" may be used for training anyway. We cover what each ChatGPT setting stops in the Project Lily explainer.
• Gemini: turning off Keep Activity, or using a temporary chat, stops review to improve Google services, but Google "still uses your chats to respond to you and help protect Google, our users, and the public, including with help from human reviewers." The rest of Gemini's review rules are in is Gemini private.
• Claude: don't send feedback on a chat you'd rather no one read. Anthropic keeps feedback five years.
• Copilot (older app): there's no opt-out of human review.

Every app's opt-out path is in our table of which AI companies train on your data, and retention and court orders are in our guide to whether AI chats are private.

How Morse keeps your chats out of the review chain

Every reviewer on this page, whether staff, a vendor, or a freelancer on a labeling platform, works from chats a company has kept on its servers and chosen to sample. Morse doesn't keep that pool. We don't train AI on your chats, our servers don't keep your messages or log what you write, and your history is stored only in your browser, encrypted. There's no archive on our servers to pull a review task from.

What a model provider sees depends on the mode you choose for each chat:

• E2EE, for anything no one in any chain should read. Your browser checks the enclave's hardware attestation, then encrypts the message so only that verified enclave can open it, and our servers relay ciphertext they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee depends on the enclave hardware and attestation working as designed. E2EE runs open-weight models, without web search or image generation.
• Private, the default, for chats you'd rather no host keep. Open-weight models like DeepSeek, Kimi, and GLM, plus xAI's Grok, run under the zero data retention agreements Morse has with its inference, search, and other providers. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
• Anonymous, for GPT, Claude, or Gemini without the provider knowing it's you. Our proxy strips your account, email, and IP address and sends the message under Morse's own account. The provider still sees the text and may keep it for a limited time to check for abuse, under its own policy, so leave out the names and places that would let a story point to you.

Your history is locked when you sign out, and opening it takes two keys: one that never leaves your device, and one our server gives you at sign-in. Your device holds the only copy of your history, so deleting a chat or clearing your browser removes it for good, and Morse can't recover it, so keep anything you need. Because we don't store chats, we usually have very little to give authorities, as our privacy policy explains.

Send a first message from the Morse home page without an account, or sign up free. E2EE chats come with Pro, at $19 a month.