{"id":113867,"date":"2026-11-18T12:00:00","date_gmt":"2026-11-18T12:00:00","guid":{"rendered":"https:\/\/www.red-gate.com\/simple-talk\/?p=113867"},"modified":"2026-10-19T13:16:50","modified_gmt":"2026-10-19T13:16:50","slug":"delete-data-from-trained-ai-model","status":"publish","type":"post","link":"https:\/\/www.red-gate.com\/simple-talk\/ai\/delete-data-from-trained-ai-model\/","title":{"rendered":"How to delete data from a trained AI model (when DELETE doesn&#8217;t work)"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\"><strong>You can\u2019t delete a single record from a trained AI model the way you <code>DELETE<\/code> a row from a database. A model stores what it learned as numerical weights, not records, so there\u2019s no row to drop. The only guaranteed way to remove someone\u2019s data is to retrain the model without it. <\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Machine unlearning, output filters and guardrails only lower the chance that data resurfaces, which is no use when you have to delete data and prove you did. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This matters for GDPR and CCPA erasure requests, which don\u2019t stop at your databases. As my <a href=\"https:\/\/www.red-gate.com\/simple-talk\/data-security-privacy-compliance\/how-to-delete-personal-data-gdpr-ccpa-compliance\/\" target=\"_blank\" rel=\"noreferrer noopener\">GDPR and CCPA data deletion guide<\/a> flagged, derived data &#8211; such as machine learning (ML) models and embeddings &#8211; are a blind spot. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This article covers that harder half: why AI models can\u2019t forget, what regulators say, your options after training, how <a href=\"https:\/\/www.red-gate.com\/simple-talk\/data-security-privacy-compliance\/how-to-anonymize-pii-in-llm-pipelines-5-key-techniques-explained\/\" target=\"_blank\" rel=\"noreferrer noopener\">large language model (LLM)<\/a> providers handle your data, and how to choose what goes into a model in the first place.<\/p>\n\n\n\n<h2 id=\"h-why-can-t-you-delete-a-record-from-a-trained-ai-model\" class=\"wp-block-heading\">Why can\u2019t you delete a record from a trained AI model?<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">You can\u2019t delete one record from a trained AI model because each training example nudges the model\u2019s weights slightly, spreading a person\u2019s data across millions (or billions) of parameters. There\u2019s no row, key or index pointing back to a single record, so there\u2019s nothing to remove.<\/p>\n<\/blockquote>\n\n\n\n<h2 id=\"h-do-trained-ai-models-memorize-personal-data\" class=\"wp-block-heading\">Do trained AI models memorize personal data?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>The data isn&#8217;t stored as a row, yet can still come back out.<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/arxiv.org\/abs\/2012.07805\" target=\"_blank\" rel=\"noreferrer noopener\"><em>In Extracting Training Data from Large Language Models<\/em><\/a>, Carlini et al. pulled <strong>verbatim names, phone numbers and email addresses<\/strong> out of <a href=\"https:\/\/en.wikipedia.org\/wiki\/GPT-2\" target=\"_blank\" rel=\"noreferrer noopener\">GPT-2<\/a> &#8211; including some from sequences that only appeared in one training document. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Some of their later work, meanwhile, <a href=\"https:\/\/proceedings.iclr.cc\/paper_files\/paper\/2025\/hash\/cce0e917b050208170151f77b497fc71-Abstract-Conference.html\" target=\"_blank\" rel=\"noreferrer noopener\">extracted training data from aligned production models such as ChatGPT<\/a>.<\/p>\n\n\n\n<h2 id=\"h-is-a-trained-ai-model-personal-data-under-gdpr\" class=\"wp-block-heading\">Is a trained AI model personal data under GDPR?<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\"><strong>A trained AI model counts as personal data under the <a href=\"https:\/\/www.red-gate.com\/simple-talk\/databases\/its-2026-why-are-databases-still-failing-gdpr-compliance-audits\/\" target=\"_blank\" rel=\"noreferrer noopener\">General Data Protection Regulation (GDPR)<\/a> if personal data can be extracted from it. <\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In that case, the AI model <em>does<\/em> fall under GDPR, including the<a href=\"https:\/\/www.red-gate.com\/simple-talk\/data-security-privacy-compliance\/how-to-delete-personal-data-gdpr-ccpa-compliance\/#:~:text=unreadable%20at%20once.-,GDPR%20%26%20CCPA%3A%20what%20the%20law%20actually%20requires,-Under%20the%20European\" target=\"_blank\" rel=\"noreferrer noopener\"> right to erasure<\/a> (Article 17).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The European Data Protection Board&#8217;s <a href=\"https:\/\/www.hunton.com\/privacy-and-cybersecurity-law-blog\/edpb-publishes-opinion-on-processing-of-personal-data-in-the-context-of-ai-models\" target=\"_blank\" rel=\"noreferrer noopener\">Opinion 28\/2024<\/a> says anonymity is assessed case by case, and a model only counts as anonymous if the chance of <a href=\"https:\/\/cms.law\/en\/int\/legal-updates\/edpb-opinion-28-2024-key-takeaways-on-processing-personal-data-in-the-context-of-ai-models\" target=\"_blank\" rel=\"noreferrer noopener\">extracting personal data, directly or through queries, is negligible<\/a>. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">It also warns that unlawful training can affect whether you can lawfully deploy the model. <strong>If you can&#8217;t show your model is anonymous, assume it&#8217;s in scope.<\/strong><\/p>\n<\/blockquote>\n\n\n\n<h2 id=\"h-what-is-algorithmic-disgorgement-in-the-us\" class=\"wp-block-heading\">What is algorithmic disgorgement in the US?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">In the US, the <a href=\"https:\/\/www.ftc.gov\/\" target=\"_blank\" rel=\"noreferrer noopener\">Federal Trade Commission (FTC)<\/a> uses <strong>algorithmic disgorgement<\/strong>: <a href=\"https:\/\/www.mintz.com\/insights-center\/viewpoints\/54731\/2024-01-23-algorithmic-disgorgement-increasingly-important-part\" target=\"_blank\" rel=\"noreferrer noopener\">forced deletion of models built on improperly collected data<\/a>.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Everalbum, Kurbo, and Rite Aid were all ordered to destroy AI models derived from such data. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Simply put: <strong>if you can&#8217;t delete a record from an AI model, a regulator may order you to just delete the model.<\/strong><\/p>\n\n\n\n<h2 id=\"h-does-ccpa-cover-ai-models\" class=\"wp-block-heading\">Does CCPA cover AI models?<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">Under the <a href=\"https:\/\/oag.ca.gov\/privacy\/ccpa\" target=\"_blank\" rel=\"noreferrer noopener\">California Consumer Privacy Act (CCPA)<\/a>, consumers can <a href=\"https:\/\/www.red-gate.com\/simple-talk\/data-security-privacy-compliance\/how-to-delete-personal-data-gdpr-ccpa-compliance\/#:~:text=In%20California%2C%20meanwhile%2C%20a%20resident%20can%20request%20deletion%20of%20their%20data%20under%20Section%201798.105%20of%20the%20California%20Consumer%20Privacy%20Act%20(CCPA).\" target=\"_blank\" rel=\"noreferrer noopener\">ask a business to delete personal information it holds about them<\/a>. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This can include personal information an AI model is able to output, so the same extraction test is the safe working assumption.<\/p>\n<\/blockquote>\n\n\n\n<h2 id=\"h-how-do-you-delete-data-from-a-trained-ai-model\" class=\"wp-block-heading\">How do you delete data from a trained AI model?<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\"><strong>The only way to guarantee a record is deleted from a trained AI model is to retrain it from scratch without that record. <\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Cheaper options exist: sharded training (SISA) and checkpoint rollback remove the data for the affected part of the model, while approximate unlearning and output suppression can\u2019t provably remove it.<\/p>\n<\/blockquote>\n\n\n\n<h3 id=\"h-how-do-the-deletion-methods-compare\" class=\"wp-block-heading\">How do the deletion methods compare?<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><td><strong>Method<\/strong><\/td><td><strong>Removes the data?<\/strong><\/td><td><strong>Relative cost<\/strong><\/td><\/tr><\/thead><tbody><tr><td>Retrain from scratch<\/td><td>Yes, probably<\/td><td>Highest<\/td><\/tr><tr><td>Sharded training (SISA) \/ checkpoint rollback<\/td><td>Yes, for the affected part<\/td><td>Medium<\/td><\/tr><tr><td>Approximate unlearning<\/td><td>Not likely<\/td><td>Low<\/td><\/tr><tr><td>Output suppression<\/td><td>No<\/td><td>Lowest<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n<div class=\"block-core-list\">\n<ul class=\"wp-block-list\">\n<li><strong>Retrain from scratch.<\/strong> The gold standard &#8211; and the only way to <em>guarantee<\/em> the data is deleted in its entirety. Affordable for your own smaller models if you batch erasure requests on a schedule.<br><br><\/li>\n\n\n\n<li><strong>Sharded training (SISA).<\/strong> SISA (Sharded, Isolated, Sliced, Aggregated) training, from <a href=\"https:\/\/ar5iv.labs.arxiv.org\/html\/1912.03817\" target=\"_blank\" rel=\"noreferrer noopener\">Bourtoule et al<\/a>., splits the data into shards, each training its own sub-model, so a deletion only retrains one shard. Exact, cheaper, and best suited to classic ML such as fraud or churn models.<br><br><\/li>\n\n\n\n<li><strong>Checkpoint rollback.<\/strong> If you know which fine-tuning batch held the record, reload the checkpoint before it and retrain without it.<br><br><\/li>\n\n\n\n<li><strong>Approximate unlearning.<\/strong> This method edits the weights directly, and can\u2019t provably show the data is gone.<br><br><\/li>\n\n\n\n<li><strong>Output suppression.<\/strong> This method involves filters that block a person\u2019s identifiers in responses, but the data stays in the weights.<\/li>\n<\/ul>\n<\/div>\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<section id=\"my-first-block-block_02b53065e6266387ea22d45812363086\" class=\"my-first-block alignwide\">\n    <div class=\"bg-brand-600 text-base-white py-5xl px-4xl rounded-sm bg-gradient-to-r from-brand-600 to-brand-500 red\">\n        <div class=\"gap-4xl items-start md:items-center flex flex-col md:flex-row justify-between\">\n            <div class=\"flex-1 col-span-10 lg:col-span-7\">\n                <h3 class=\"mt-0 font-display mb-2 text-display-sm\">Subscribe to the Simple Talk newsletter<\/h3>\n                <div class=\"child:last-of-type:mb-0\">\n                                            All the latest articles, guides, podcasts, and more &#8211; delivered straight to your inbox.                                    <\/div>\n            <\/div>\n                                            <a href=\"https:\/\/www.red-gate.com\/simple-talk\/subscribe\/\" class=\"btn btn--secondary btn--lg\" aria-label=\"Subscribe now: Subscribe to the Simple Talk newsletter\">Subscribe now<\/a>\n                    <\/div>\n    <\/div>\n<\/section>\n\n\n<h2 id=\"h-do-openai-anthropic-and-google-train-on-your-data\" class=\"wp-block-heading\">Do OpenAI, Anthropic and Google train on your data?<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">When an AI provider such as OpenAI, Anthropic or Google deletes your data, it removes stored copies such as logs and chat history. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">It doesn\u2019t remove what a model has already learned. Whether your data is used for training at all depends on your account tier:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n<\/blockquote>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><td><strong>Provider<\/strong><\/td><td><strong>API \/ business data used for training?<\/strong><\/td><td><strong>Consumer \/ free tier<\/strong><\/td><\/tr><\/thead><tbody><tr><td>OpenAI (<a href=\"https:\/\/help.openai.com\/en\/articles\/5722486-how-your-data-is-used-to-improve-model-performance\" target=\"_blank\" rel=\"noreferrer noopener\">full policy<\/a>)<\/td><td>No by default; zero data retention for qualifying API customers<\/td><td>Different rules for consumer ChatGPT<\/td><\/tr><tr><td>Anthropic (<a href=\"https:\/\/www.anthropic.com\/policy\" target=\"_blank\" rel=\"noreferrer noopener\">full policy<\/a>)<\/td><td>No; commercial and API use is excluded from the <a href=\"https:\/\/privacy.claude.com\/en\/articles\/10023580-is-my-data-used-for-model-training\" target=\"_blank\" rel=\"noreferrer noopener\">consumer training policy changes<\/a><\/td><td>Users choose whether chats train models<\/td><\/tr><tr><td>Google Gemini API (<a href=\"https:\/\/ai.google.dev\/gemini-api\/terms\" target=\"_blank\" rel=\"noreferrer noopener\">full policy<\/a>)<\/td><td>Not on paid services<\/td><td>Unpaid use may improve products; human reviewers may read inputs<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><em>Note: <a href=\"https:\/\/trust.anthropic.com\/faq\" target=\"_blank\" rel=\"noreferrer noopener\">Anthropic&#8217;s FAQ<\/a> states that t<strong>urning off training later leaves your data in models that have already been trained<\/strong>.<\/em><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Never use free or consumer tiers for customer data<\/strong> &#8211; and you must sign a Data Processing Addendum (DPA), use zero data retention for regulated workloads, and <a href=\"https:\/\/www.red-gate.com\/simple-talk\/data-security-privacy-compliance\/how-to-anonymize-pii-in-llm-pipelines-5-key-techniques-explained\/\" target=\"_blank\" rel=\"noreferrer noopener\">anonymize personally identifiable information (PII) before it leaves your network<\/a>.<\/p>\n\n\n\n<h2 id=\"h-which-data-is-safe-to-train-an-ai-model-on\" class=\"wp-block-heading\">Which data is safe to train an AI model on?<\/h2>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\"><strong>Any record you put into an AI model can come back out<\/strong>, so prevention is the only deletion strategy that scales. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Use this information to decide what\u2019s safe to train on:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n<\/blockquote>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><td><strong>Data type<\/strong><\/td><td><strong>Train on it?<\/strong><\/td><td><strong>Safer alternative<\/strong><\/td><\/tr><\/thead><tbody><tr><td>Direct identifiers (name, email, phone)<\/td><td>No<\/td><td>Drop or tokenize<\/td><\/tr><tr><td>Quasi-identifiers (ZIP code\/postcode, birth date)<\/td><td>Rarely<\/td><td>Generalize (region, age band)<\/td><\/tr><tr><td>Free text (tickets, notes, transcripts)<\/td><td>Only after redaction<\/td><td>PII redaction (plus spot-checks)<\/td><\/tr><tr><td>Special-category and children&#8217;s data<\/td><td>No<\/td><td>Aggregates or differentially private training<\/td><\/tr><tr><td>Records with a pending erasure or objection<\/td><td>No<\/td><td>Exclude at source<\/td><\/tr><tr><td>Product docs, aggregates, masked records<\/td><td>Yes<\/td><td>None needed<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h2 id=\"h-how-to-safely-prepare-ai-training-data\" class=\"wp-block-heading\">How to safely prepare AI training data<\/h2>\n\n\n<div class=\"block-core-list\">\n<ol class=\"wp-block-list\">\n<li><strong>Classify<\/strong>. Tag columns by sensitivity using your data map. <a href=\"https:\/\/www.red-gate.com\/products\/sql-data-catalog\/\" target=\"_blank\" rel=\"noreferrer noopener\">Redgate&#8217;s SQL Data Catalog<\/a> discovers and classifies sensitive columns across SQL Server.<br><br><\/li>\n\n\n\n<li><strong>Minimize and pseudonymize.<\/strong> Only retain columns the model actually needs, and swap IDs for keyed hashes (keys held in a vault). Pseudonymized data is still personal data.<br><br><\/li>\n\n\n\n<li><strong>Mask what remains.<\/strong> Replace real values with realistic fake ones using static data masking, such as Redgate Data Masker, so the training set contains no real identifiers.<br><br><\/li>\n\n\n\n<li><strong>Scrub free text<\/strong> with an entity recognizer such as <a href=\"https:\/\/microsoft.github.io\/presidio\/\" target=\"_blank\" rel=\"noreferrer noopener\">Presidio<\/a>.<br><br><\/li>\n\n\n\n<li><strong>Deduplicate and test.<\/strong> Repeated text is memorized more easily. Plant fake &#8216;canary&#8217; records and try to prompt them back out after training.<br><br><\/li>\n\n\n\n<li><strong>Consider differential privacy for sensitive data.<\/strong> Differential privacy adds calibrated noise during training so no single record measurably changes the model. <a href=\"https:\/\/research.google\/blog\/vaultgemma-the-worlds-most-capable-differentially-private-llm\" target=\"_blank\" rel=\"noreferrer noopener\">Google\u2019s VaultGemma<\/a> shows how it works at scale (albeit at some cost to quality).<\/li>\n<\/ol>\n<\/div>\n\n\n<h2 id=\"h-how-do-you-handle-an-erasure-request-involving-an-ai-model\" class=\"wp-block-heading\">How do you handle an erasure request involving an AI model?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Add one more step to your AI model data deletion runbook: query the <strong>training manifest<\/strong> for every active model (and its derivatives) trained on this person&#8217;s data. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Then, take the cheapest action that truly removes it. Delete a tenant adapter, retrain affected shards, roll back a recent fine-tune, or queue the record for the next scheduled retrain with an output filter in the meantime.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"901\" height=\"464\" src=\"https:\/\/www.red-gate.com\/simple-talk\/wp-content\/uploads\/2026\/10\/image.gif\" alt=\"Image showing the process of handling an erasure request involving an AI model.\" class=\"wp-image-113868\" style=\"aspect-ratio:1.9449838187702266;width:601px;height:auto\" srcset=\"https:\/\/www.red-gate.com\/simple-talk\/wp-content\/uploads\/2026\/10\/image.gif 901w, https:\/\/www.red-gate.com\/simple-talk\/wp-content\/uploads\/2026\/10\/image-300x154.png 300w, https:\/\/www.red-gate.com\/simple-talk\/wp-content\/uploads\/2026\/10\/image-767x395.png 767w\" sizes=\"auto, (max-width: 901px) 100vw, 901px\" \/><figcaption class=\"wp-element-caption\"><em>The process of handling an erasure request involving an AI model.<\/em><\/figcaption><\/figure>\n\n\n\n<h2 id=\"h-conclusion-you-can-t-delete-what-you-never-trained-on\" class=\"wp-block-heading\">Conclusion: you can&#8217;t delete what you never trained on<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Only retraining an AI model guarantees that data will actually be deleted from it<\/strong>, which is why the decisions you make <em>before<\/em> training the model matter most. <\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Keep personal data where <code>DELETE<\/code> works, filter training data through an auditable view, scrub free text, record what went in, and never send customer data through a tier that trains on it.<\/p>\n\n\n\n<div id=\"callout-block_d13527e69a566aaf803cb05198bea9f9\" class=\"callout alignnone\">\n    <div class=\"child-last:mb-0 child-first:mt-0 bg-gray-50 dark:bg-gray-950 p-4xl my-3xl\">\n\n<p class=\"wp-block-paragraph\"><strong>You may also be interested in&#8230;<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">For the rest of the data deletion process, see my earlier <a href=\"https:\/\/www.red-gate.com\/simple-talk\/data-security-privacy-compliance\/how-to-delete-personal-data-gdpr-ccpa-compliance\/\" target=\"_blank\" rel=\"noreferrer noopener\"><strong>GDPR and CCPA data deletion guide<\/strong><\/a> here on Simple Talk.<\/p>\n\n<\/div>\n<\/div> \n\n\n<section id=\"my-first-block-block_aa40fe8ae4f1529808ee907634680c75\" class=\"my-first-block alignwide\">\n    <div class=\"bg-brand-600 text-base-white py-5xl px-4xl rounded-sm bg-gradient-to-r from-brand-600 to-brand-500 red\">\n        <div class=\"gap-4xl items-start md:items-center flex flex-col md:flex-row justify-between\">\n            <div class=\"flex-1 col-span-10 lg:col-span-7\">\n                <h3 class=\"mt-0 font-display mb-2 text-display-sm\">Simple Talk is brought to you by Redgate Software<\/h3>\n                <div class=\"child:last-of-type:mb-0\">\n                                            Take control of your databases with the trusted Database DevOps solutions provider. Automate with confidence, scale securely, and unlock growth through AI.                                    <\/div>\n            <\/div>\n                                            <a href=\"https:\/\/www.red-gate.com\/solutions\/overview\/\" class=\"btn btn--secondary btn--lg\" aria-label=\"Discover how Redgate can help you: Simple Talk is brought to you by Redgate Software\">Discover how Redgate can help you<\/a>\n                    <\/div>\n    <\/div>\n<\/section>\n\n\n<section id=\"faq\" class=\"faq-block my-5xl\">\n    <h2>FAQs: How to delete data from a trained AI model<\/h2>\n\n            <h3 class=\"mt-4xl\">1. Can you delete data from a trained AI model?<\/h3>\n        <div class=\"faq-answer\">\n            <p dir=\"ltr\">Not in the way you delete a database row. A trained model stores what it learned as weights, not records. The only guaranteed way to remove a person\u2019s data is to retrain the model from scratch without it. Sharded training and checkpoint rollback cut the cost by retraining only the affected part. Machine unlearning and output filters lower the risk of data resurfacing but can\u2019t prove it\u2019s gone.<\/p>\n        <\/div>\n            <h3 class=\"mt-4xl\">2. Does GDPR\u2019s right to erasure apply to AI models?<\/h3>\n        <div class=\"faq-answer\">\n            <p dir=\"ltr\">It can. The European Data Protection Board\u2019s (EDPB) Opinion 28\/2024 says a model is only anonymous if the chance of extracting personal data, directly or through queries, is negligible. If you can\u2019t show that, treat the model as personal data and within scope of the right to erasure. Unlawful training can also affect whether you may lawfully deploy the model.<\/p>\n        <\/div>\n            <h3 class=\"mt-4xl\">3. What is algorithmic disgorgement?<\/h3>\n        <div class=\"faq-answer\">\n            <p dir=\"ltr\">Algorithmic disgorgement is a remedy used by the US Federal Trade Commission (FTC) that forces a company to delete AI models built on improperly collected data. Everalbum, Kurbo and Rite Aid were all ordered to destroy models derived from such data. If you can\u2019t delete one record from a model, a regulator may order you to delete the whole model.<\/p>\n        <\/div>\n            <h3 class=\"mt-4xl\">4. Do OpenAI, Anthropic and Google train on your data?<\/h3>\n        <div class=\"faq-answer\">\n            <p dir=\"ltr\">By default, OpenAI and Anthropic don\u2019t train on API or business data, and Google doesn\u2019t on paid Gemini API services. Consumer and free tiers differ: Google\u2019s unpaid use may improve products and allow human review, while consumer chat products may let users choose. Deleting your data later removes stored copies such as logs and chat history, not anything a model has already learned.<\/p>\n        <\/div>\n            <h3 class=\"mt-4xl\">5. Which data should you never train an AI model on?<\/h3>\n        <div class=\"faq-answer\">\n            <p dir=\"ltr\">Never train on direct identifiers (names, emails, phone numbers), special-category or children\u2019s data, or records with a pending erasure request or objection. Quasi-identifiers such as ZIP code\/postcode and birth date should rarely be used, and free text only after redaction. Drop, tokenize, generalize or exclude these at source, because any record shared with a model can come back out.<\/p>\n        <\/div>\n    <\/section>\n","protected":false},"excerpt":{"rendered":"<p>You can\u2019t DELETE a record from a trained AI model. Compare retraining, SISA, unlearning and filters, and see how to meet GDPR and CCPA erasure requests.&hellip;<\/p>\n","protected":false},"author":346911,"featured_media":103099,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[159169,143523,53],"tags":[159075,4483,4168,4170],"coauthors":[159385],"class_list":["post-113867","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-databases","category-featured","tag-ai","tag-data","tag-database","tag-database-administration"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/posts\/113867","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/users\/346911"}],"replies":[{"embeddable":true,"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/comments?post=113867"}],"version-history":[{"count":17,"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/posts\/113867\/revisions"}],"predecessor-version":[{"id":113933,"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/posts\/113867\/revisions\/113933"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/media\/103099"}],"wp:attachment":[{"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/media?parent=113867"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/categories?post=113867"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/tags?post=113867"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/www.red-gate.com\/simple-talk\/wp-json\/wp\/v2\/coauthors?post=113867"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}