# MIT Technology Review’s The Download on scraped personal data and chatbot health advice [Canonical reading](https://escapemeridian.com/readings/mit-technology-review-s-the-download-on-scraped-personal-data-6a9adade) MIT Technology Review’s The Download, published July 21, 2025, reports two research findings about AI and personal risk. Researchers found thousands of images, including faces and identity documents, in a small sample of DataComp CommonPool, an open-source AI training set for image generation. Because they checked only a small fraction of the data, they estimate the full set may contain hundreds of millions of images with personally identifiable information. The newsletter’s warning is stark: material put online can be scraped and used in training data. The edition also reports research finding that AI companies have mostly stopped including medical disclaimers when chatbots answer health questions. It says many leading models now ask follow-up questions and attempt diagnoses. The researchers argue that disclaimers remind people not to treat chatbot responses as medical advice; without them, users may be more likely to trust unsafe advice. The evidence here is the newsletter’s account of the two studies, not their full methods or results. How often the scraped images identify real people, and how much removing disclaimers changes users’ trust or medical decisions, remain unknown from this account. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source MIT Technology Review · Article Original source date: Jul 21, 2025 (2025-07-21T12:10:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-13T02:33:24.730Z. First observed: 2026-11-02T21:47:52.320Z. Retrieved: 2026-11-02T21:49:19.986Z. - [The Download: how your data is being used to train AI, and why chatbots aren’t doctors](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Limits Text as retrieved on November 2, 2026; originally published July 21, 2025. The newsletter summarizes the research and links to full stories; their methods and underlying evidence are not included in the retrieved text. The source gives an estimate based on an audit of a small subset; it does not establish the actual number of images containing personally identifiable information. The account reports the researchers’ claim about users’ trust, but gives no measure of how disclaimer changes affect trust or medical decisions. ## Open question How often do the identified images expose real people, and how much does removing chatbot medical disclaimers affect users’ trust or decisions? ## Source-stated Researchers found thousands of images, including identifiable faces, in a small subset of DataComp CommonPool, and estimate the full dataset may contain hundreds of millions of images with personally identifiable information. > Thousands of images—including identifiable faces—were found in a small subset of DataComp CommonPool, a major AI training set for image generation scraped from the web. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source-stated The newsletter says researchers estimate that the full dataset contains hundreds of millions of images with personally identifiable information, including faces and identity documents. > Because the researchers audited just 0.1% of CommonPool’s data, they estimate that the real number of images containing personally identifiable information, including faces and identity documents, is in the hundreds of millions. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source-stated The newsletter reports research finding that AI companies have mostly abandoned medical disclaimers in chatbot responses to health questions. > AI companies have now mostly abandoned the once-standard practice of including medical disclaimers and warnings in response to health questions, new research has found. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source-stated The newsletter reports that many leading AI models answer health questions, ask follow-ups, and attempt diagnoses. > In fact, many leading AI models will now not only answer health questions but even ask follow-ups and attempt a diagnosis. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source-stated The researchers say medical disclaimers remind users not to rely uncritically on chatbot answers, and that their absence means users are more likely to trust unsafe medical advice. > Such disclaimers serve an important reminder to people asking AI about everything from eating disorders to cancer diagnoses, the authors say, and their absence means that users of AI are more likely to trust unsafe medical advice. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Editorial provenance Version 1. Publication ID: 532e8120-faa9-93e0-58da-40445bacc1fa. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T02:28:48.170Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: d62967f5da1ddfb8aa91523f6d12c30d620b684b5a97ef1fb09cd92f79323b41. Source text hash: c9a3e8e0a3192c28ed491a68e6352c30ccf3fc75225ad62f38a11ddd31a2554b. ## Revision and source-check history ### Version 1 — 2026-11-13T02:33:24.730Z MIT Technology Review’s The Download, published July 21, 2025, reports two research findings about AI and personal risk. Researchers found thousands of images, including faces and identity documents, in a small sample of DataComp CommonPool, an open-source AI training set for image generation. Because they checked only a small fraction of the data, they estimate the full set may contain hundreds of millions of images with personally identifiable information. The newsletter’s warning is stark: material put online can be scraped and used in training data. The edition also reports research finding that AI companies have mostly stopped including medical disclaimers when chatbots answer health questions. It says many leading models now ask follow-up questions and attempt diagnoses. The researchers argue that disclaimers remind people not to treat chatbot responses as medical advice; without them, users may be more likely to trust unsafe advice. The evidence here is the newsletter’s account of the two studies, not their full methods or results. How often the scraped images identify real people, and how much removing disclaimers changes users’ trust or medical decisions, remain unknown from this account. ## Original source MIT Technology Review · Article Original source date: Jul 21, 2025 (2025-07-21T12:10:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-13T02:33:24.730Z. First observed: 2026-11-02T21:47:52.320Z. Retrieved: 2026-11-02T21:49:19.986Z. - [The Download: how your data is being used to train AI, and why chatbots aren’t doctors](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Limits Text as retrieved on November 2, 2026; originally published July 21, 2025. The newsletter summarizes the research and links to full stories; their methods and underlying evidence are not included in the retrieved text. The source gives an estimate based on an audit of a small subset; it does not establish the actual number of images containing personally identifiable information. The account reports the researchers’ claim about users’ trust, but gives no measure of how disclaimer changes affect trust or medical decisions. ## Open question How often do the identified images expose real people, and how much does removing chatbot medical disclaimers affect users’ trust or decisions? ## Source-stated Researchers found thousands of images, including identifiable faces, in a small subset of DataComp CommonPool, and estimate the full dataset may contain hundreds of millions of images with personally identifiable information. > Thousands of images—including identifiable faces—were found in a small subset of DataComp CommonPool, a major AI training set for image generation scraped from the web. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source-stated The newsletter says researchers estimate that the full dataset contains hundreds of millions of images with personally identifiable information, including faces and identity documents. > Because the researchers audited just 0.1% of CommonPool’s data, they estimate that the real number of images containing personally identifiable information, including faces and identity documents, is in the hundreds of millions. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source-stated The newsletter reports research finding that AI companies have mostly abandoned medical disclaimers in chatbot responses to health questions. > AI companies have now mostly abandoned the once-standard practice of including medical disclaimers and warnings in response to health questions, new research has found. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source-stated The newsletter reports that many leading AI models answer health questions, ask follow-ups, and attempt diagnoses. > In fact, many leading AI models will now not only answer health questions but even ask follow-ups and attempt a diagnosis. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source-stated The researchers say medical disclaimers remind users not to rely uncritically on chatbot answers, and that their absence means users are more likely to trust unsafe medical advice. > Such disclaimers serve an important reminder to people asking AI about everything from eating disorders to cancer diagnoses, the authors say, and their absence means that users of AI are more likely to trust unsafe medical advice. - [Supporting source](https://www.technologyreview.com/2025/07/21/1120525/the-download-how-your-data-is-being-used-to-train-ai-and-why-chatbots-arent-doctors/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Editorial provenance Version 1. Publication ID: 532e8120-faa9-93e0-58da-40445bacc1fa. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T02:28:48.170Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: d62967f5da1ddfb8aa91523f6d12c30d620b684b5a97ef1fb09cd92f79323b41. Source text hash: c9a3e8e0a3192c28ed491a68e6352c30ccf3fc75225ad62f38a11ddd31a2554b. ### passed — 2026-11-13T02:28:48.170Z — anthropic/claude-opus-5.5 Publication: 532e8120-faa9-93e0-58da-40445bacc1fa attribution — pass: The reading attributes the content to MIT Technology Review's The Download. This newsletter issue comments on several items in its own words, so it counts as one piece. The study findings are attributed to researchers and the newsletter rather than presented as established fact. > This is today's edition of The Download, our weekday newsletter that provides a daily dose of what's going on in the world of technology. summary — pass: The reading clearly explains both items, with DataComp CommonPool described as an open-source image-generation training set, and a lay reader can follow it. It renders the newsletter's line as material that can be scraped and drops 'probably has been', which slightly softens its force without changing its meaning. The question follows from gaps the source leaves open. > The bottom line? Anything you put online can be and probably has been scraped. date — pass: July 21, 2025 matches the UTC calendar day of originalPublished. The page's dateModified falls before publication, so there is no later edit. The Deep Dive listings may be later additions, but the reading neither uses nor relies on them. The evidence is a publisher page, so no archive disclosure is needed, and there is no hindsight framing. support — pass: Each claim matches its excerpt. Describing the disclaimers as reminders not to rely on chatbots as doctors is a mild paraphrase consistent with the headline and the authors' stated reminder. The stated unknowns and limitations stay within the evidence and invent no stakeholder impacts. > Such disclaimers serve an important reminder to people asking AI about everything from eating disorders to cancer diagnoses, the authors say media — pass: No media was selected, and no verified candidates were supplied. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/6a9adadeb1a1c2f833f71dac6f725ce08615ac4edf5fa6c7090dc2a43d4a8c5b/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/mit-technology-review-s-the-download-on-scraped-personal-data-6a9adade.json)