What Advice Does Claude Have for AI Privacy? (2026)

Claude 3 Opus, the model this post was originally about, was retired in January 2026. The current Claude family — Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 — works differently, and every major AI engine has since added privacy controls that did not exist in 2024: training opt-outs, incognito and temporary chats, and memory settings. This refresh replaces our October 2024 post with what is true now, checked against each vendor’s own documentation on 25 October 2026.

Updated 25 October 2026. The first version of this post (October 2024) was a generic roundup written around Claude 3 Opus. That model no longer exists, some of its claims were too broad even then, and the legal landscape has moved on. This version names the current controls in ChatGPT, Claude, Perplexity, Gemini and Grok, and what has changed in AI-privacy law. It is general information, not legal advice.

The short version

Your chats generally train models only if you allow it. Claude, ChatGPT and Perplexity each have a training setting you can turn off, and business, Enterprise and API tiers generally do not train on your data by default. Exceptions exist: a chat flagged for safety review, or one you submit as feedback, can still be used even after opting out.
Incognito exists now. Claude has incognito chats, ChatGPT has temporary chats, Perplexity has incognito search. They stay out of history and training, with a short safety copy (ordinarily up to 30 days on Claude and ChatGPT; longer retention can apply for Enterprise exports, legal holds and safety reviews).
Memory is a separate setting. Turning off training does not necessarily turn off memory, and deleting a chat does not necessarily delete what the engine remembered from it. Check both. Note that a temporary chat can still draw on existing memories without creating new ones.
A practical risk businesses can reduce is their own people pasting company data into a consumer chatbot. That has not changed since 2024.
The law has moved. The EU AI Act’s transparency duties apply from August 2026, Texas’s AI law is in force, and California and Colorado have automated-decision rules landing 1 January 2027.

Claude’s privacy controls in 2026

Training. On consumer plans (Free, Pro, Max), Anthropic trains on your chats only if you turn on “Help Improve Claude” under Settings, then Privacy — or if a chat is flagged for a safety review, you send feedback on it, or you join an explicit program. Incognito chats are excluded even when that setting is on. Claude for Work, the API and Claude Gov do not train on your data by default, with narrow exceptions such as an explicit opt-in.

Deletion. Deleting a chat removes it from your history immediately and from Anthropic’s back end within 30 days. The fine print matters: turning training off stops future use, but it does not pull your data out of a training run already underway or a model already trained. If training was on, de-identified copies can persist in training pipelines for up to 5 years.

Memory. Claude can remember things across chats (“Memory from chats”) and reference past chats. Both can be switched off, and memory entries can be deleted separately — deleting the chat that created them is not enough.

Who can read your chats. By default, Anthropic staff cannot read your conversations unless you send feedback or a usage-policy review is triggered. Human access varies by vendor: Google, for example, routinely has trained reviewers sample Gemini conversations to evaluate quality and improve services when activity tracking is on — review there is not limited to feedback or abuse cases.

The other four engines, in brief

ChatGPT. Free, Plus and Pro may train on your content unless you turn off “Improve the model for everyone” (Settings, then Data controls) or choose “Do not train on my content” in the Privacy Portal — either one is enough. Temporary chats stay out of history, memory and training, with a safety copy up to 30 days. Memory is a separate control. Business, Enterprise, Edu and API tiers do not train by default.
Perplexity. Free, Pro and Max have “AI data retention” on by default (Account, then Preferences); turning it off stops future training use but does not delete what was already used. Enterprise data is never used for training, and Perplexity says its contracts bar third-party model providers from training on Perplexity data. Incognito mode disables saved search history and memory, with sessions expiring within 24 hours.
Gemini. The Privacy Hub (updated 23 October 2026) auto-deletes activity after 18 months by default, adjustable to 3 or 36 months. With activity tracking off, future chats are not used for training unless you send feedback, but they are still stored for 72 hours. Human-reviewed chats can be kept up to 3 years. Google Workspace prompts are customer data and are not used to train models without permission.
Grok. X may share public posts plus Grok inputs, outputs and voice transcripts with xAI for training and personalization. Opt out under Privacy and safety, then Data sharing and personalization; a protected account keeps posts out of training. Deleting everything removes conversations within 30 days, aside from security or legal holds.

What the old post got wrong

Worth stating plainly, since the 2024 version is what ranked until now:

“Your input may be retained and accessible to other users.” Retention is real; other users reading your chats is not how any of these services work. Staff access is limited to feedback, safety reviews and abuse cases.
“Your data may be used for targeted advertising.” Anthropic says training data is not used for marketing and is not sold. OpenAI says it does not share or sell chat content to advertisers — but conversation context and intent can inform which ads ChatGPT shows you, and saved memories may too if personalization is on. Selling your conversations and using conversation signals for ad selection are different things; the 2024 sentence blurred them.
The old legal advice was generic. The 2024 post told companies to “understand applicable AI laws” without naming any jurisdiction’s live duties. Since then the picture has sharpened: the Colorado law everyone summarized was repealed before it applied and replaced with a narrower automated-decision law (see below), and the EU, California and Texas each have dated duties now.

The law since 2024

Four things a small business should know. This is not legal advice.

EU AI Act transparency (applies from 2 August 2026). Providers must disclose AI interaction and machine-mark certain synthetic outputs; deployers (that includes you, if you run the chatbot) have duties for deepfakes and some public-interest text, with exceptions — for example where AI involvement is obvious or a human editor reviewed the text. So a chatbot on your marketing site should tell visitors they are talking to AI, but the Act does not demand a visible label on every synthetic output. Systems already operating before August 2026 have until 2 December 2026 for the machine-marking duty. Broader high-risk duties were pushed to late 2027 and 2028. The model-provider duties fall on OpenAI and Anthropic, not on a firm that just uses their tools.
Texas (in force since 1 January 2026). Intent-based bans plus disclosure duties for government and health-care AI use. Enforcement is led by the Attorney General, with possible follow-on licensing sanctions.
California (automated-decision rules by 1 January 2027). CCPA-covered businesses using automated decision-making for significant decisions — lending, housing, employment, education, health care — must meet notice, opt-out and access rules, subject to statutory exceptions.
Colorado (replacement law, consequential decisions from 1 January 2027). Notice, explanation after an adverse outcome, correction and human review for automated decisions in areas like employment, housing and finance. No new private right of action.

Practical read: a support chatbot on your site is EU transparency territory. A hiring or lending tool is US state automated-decision territory. Most small marketing sites are the former.

Safeguards worth keeping from the old post

The 2024 version covered some ground this refresh should not lose:

Training data can contain personal information. Large models are trained on vast public datasets that can incidentally include personal data. That is a property of the models, not something your settings change — another reason not to paste sensitive data into prompts.
Models can be pushed into leaking. Generative systems can be manipulated into producing misleading content or surfacing data they were given. Treat model output as untrusted until reviewed, especially anything republished to customers.
Minimize and govern. The durable company advice: decide what data may go into AI tools, keep retention short, run a privacy impact assessment before a consequential deployment, and teach employees the rules. None of that has changed.
Kids use these tools too. If your household or classroom uses AI chatbots, the same rules apply — talk about what not to share — since children’s data gets the same retention treatment as anyone’s.

What to do this week

Turn off training in every consumer AI account your team uses — Claude, ChatGPT, Perplexity, Gemini — or move work to a business tier that does not train by default.
Check memory separately. Disabling training is not the same as disabling memory. Delete stored memory entries you do not need.
Use incognito or temporary chats for anything sensitive, and treat them as safer, not safe: short safety copies still exist.
Write the one-line rule for your team: no customer data, credentials, or unreleased business information in a consumer chatbot. That reduces accidental disclosure and training exposure more than any other single step.
If your site has an AI chatbot, add a visible note that visitors are talking to AI — the EU’s transparency duties apply as of August 2026.
Keep passwords, MFA and basic account hygiene. Unchanged since 2024, still the cheapest protection there is.

AI privacy FAQ

Is my data used to train AI models? On consumer plans, possibly — but on Claude, ChatGPT and Perplexity it is a setting you can turn off, and business, Enterprise and API tiers generally do not train on your data by default. Turning training off stops future ordinary use; it does not remove data from models already trained, and a chat flagged for safety review or submitted as feedback can still be used.

Can other users see my AI chats? No. None of the five major engines shares your chats with other users. On Claude, staff access is limited to feedback you send plus safety, abuse, support and legal review. Other vendors differ: Google has trained reviewers routinely sample Gemini conversations when activity tracking is on.

Does deleting a chat delete my data? Ordinarily it removes the chat from your history immediately and from the vendor’s back end within about 30 days (Claude, ChatGPT). But deleting a chat does not always delete memory entries made from it — delete those separately — it never un-trains a finished model, and longer retention can apply for Enterprise exports, legal holds and safety reviews.

What is the difference between incognito chats and turning off training? Turning off training stops future chats from being used to improve models, but chats are still saved to history and memory. Incognito (Claude), temporary chats (ChatGPT) and incognito search (Perplexity) additionally stay out of history and training, keeping only a short safety copy. Note that a temporary chat can still draw on existing memories without creating new ones.

Are AI companies selling my data to advertisers? Anthropic says it does not use personal data in training for marketing and does not sell it. OpenAI says it does not share or sell ChatGPT conversations to advertisers — but conversation context and saved memories can inform which ads you see. X says it does not sell Grok data.

Do small businesses need to worry about the EU AI Act? If your site has an AI chatbot, yes at a basic level: since 2 August 2026 the Act’s transparency duties mean telling visitors they are interacting with AI, with deployer duties for deepfakes and some public-interest text (subject to exceptions). It does not require labeling every synthetic output. The heavier high-risk duties mostly concern model providers and specific high-risk uses, not ordinary marketing sites.

What is the biggest AI privacy risk for a company? A practical, reducible one: employees pasting customer data, credentials or unreleased plans into a consumer chatbot. A written rule plus business-tier accounts reduces accidental disclosure and training exposure more than any other single step.

Should we run a private AI model instead? For most small businesses, no. Zero-retention deals exist for eligible API and Enterprise use, but they are contracts for specific setups, not something a consumer account offers. Turning off training and memory on a business tier covers the realistic risk at far lower cost.

Bibliography

Show sources

Anthropic, “Model deprecations.” https://platform.claude.com/docs/en/about-claude/model-deprecations
Anthropic, “Models overview.” https://platform.claude.com/docs/en/about-claude/models/overview
Anthropic, “I want to opt out of my prompts and results being used for training models.” https://privacy.anthropic.com/en/articles/10023580-i-want-to-opt-out-of-my-prompts-and-results-being-used-for-training-models
Anthropic, “How long do you store personal data.” https://privacy.anthropic.com/en/articles/10023548-how-long-do-you-store-personal-data
Anthropic, “How does Anthropic protect the personal data of Claude AI users.” https://privacy.anthropic.com/en/articles/10458704-how-does-anthropic-protect-the-personal-data-of-claude-ai-users
Anthropic, “Is my data used for model training.” https://privacy.anthropic.com/en/articles/7996868-is-my-data-used-for-model-training
Anthropic, “Data retention practices for covered models.” https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-models
OpenAI, “How your data is used to improve model performance.” https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance
OpenAI, “Data usage for consumer services FAQ.” https://help.openai.com/en/articles/7039943-data-usage-for-consumer-services-faq
Perplexity, “Data collection at Perplexity.” https://www.perplexity.ai/help-center/en/articles/11564572-data-collection-at-perplexity
Perplexity, “Are third-party model providers training on my data.” https://www.perplexity.ai/help-center/en/articles/10354963-are-third-party-model-providers-training-on-my-data
Google, “Manage your Gemini Apps activity.” https://support.google.com/gemini/answer/13594961?hl=en
X, “About Grok.” https://help.x.com/en/using-x/about-grok
EU, Regulation 2024/1689 (AI Act). https://eur-lex.europa.eu/eli/reg/2024/1689/oj
EU, Regulation 2026/1744 (Digital Omnibus on AI). https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=OJ:L_202601744
EU AI Act Service Desk, “Article 50: Transparency.” https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
EU AI Act Service Desk, “Implementation timeline.” https://ai-act-service-desk.ec.europa.eu/en/ai-act/eu-ai-act-implementation-timeline
Texas HB 149 (TRAIGA), enrolled text. https://capitol.texas.gov/tlodocs/89R/billtext/html/HB00149F.htm
CPPA, CCPA updates and regulations. https://www.cppa.ca.gov/regulations/ccpa_updates.html
Colorado SB 26-189. https://leg.colorado.gov/bills/sb26-189
Stack Overflow, “Privacy in the age of generative AI.” https://stackoverflow.blog/2023/10/23/privacy-in-the-age-of-generative-ai/
Forbes Tech Council, “How companies can use generative AI and maintain data privacy.” https://www.forbes.com/councils/forbestechcouncil/2023/06/22/how-companies-can-use-generative-ai-and-maintain-data-privacy/
Live documentation checks by WebbROI, 25 October 2026.