What each AI assistant does with what you type, plan by plan: whether it trains its models on it, how long it keeps it, where it’s stored and what an admin can control. Every line comes from the maker’s own pages, with the date we read it.
Checked 17 October 2026 against 141 sources, each the maker’s own page. Makers change these settings, so check the tool’s own settings before you rely on a line. See every source
A personal account and a work account of the same assistant can treat what you type differently. That’s why the work account matters.
ChatGPTby OpenAI
Free, Go, Plus and ProPersonal accounts
Does it train on what you type?
Yes, unless you switch it off. Even after you opt out, a thumbs up or thumbs down on a reply means OpenAI may use that whole conversation for training. 12
Can that be switched off?
Yes. Turn off "Improve the model for everyone" under Settings > Data controls, or choose "Do not train on my content" in OpenAI's Privacy Portal; either one covers your new chats. 13
How long is it kept?
Chats stay until you delete them, then leave OpenAI's systems within 30 days unless kept for security or legal reasons, and temporary chats are kept up to 30 days for safety. Deleting a chat doesn't delete a saved memory made from it, and OpenAI still holds some April to September 2025 user data for the New York Times lawsuit. 4567
Where is it stored?
OpenAI stores content on its own and its providers' systems in the US and around the world. Personal plans can't choose Australian storage, which OpenAI offers only to new Enterprise and Edu customers and to eligible API customers. 89
What can an admin control?
No admin: a personal account. A parent or guardian linked to a teen's account can manage the training setting through Parental controls. 1
Can people at the maker read it?
Authorised OpenAI staff and trusted service providers may access chats to investigate abuse or security incidents, give support, handle legal matters, or improve models unless you opted out. Trained staff review flagged chats, and OpenAI notifies law enforcement when one shows an imminent and credible risk of harm to others. 810
BusinessWork accounts for small and growing teams
Does it train on what you type?
No, unless your organisation signs a separate agreement with OpenAI. Thumbs-up or thumbs-down feedback doesn't change that. 111
Can that be switched off?
Nothing to switch off, because training is off by default. Admins can still turn off the feedback buttons at Workspace settings > Allow response feedback. 1
How long is it kept?
Chats stay until the member deletes them, and OpenAI's pricing page lists custom retention periods only under Enterprise. Deleted and unsaved chats leave OpenAI's systems within 30 days, unless the law requires longer or OpenAI needs them to protect its services or others from harm. 121311
Where is it stored?
Business can't choose where its data is stored, as OpenAI marks data residency "No" for this plan. OpenAI's sub-processor list shows Business data can be processed by cloud providers in many countries, Australia and the US among them. 1314
What can an admin control?
Owners and admins manage members, apps and the feedback buttons; owners also run single sign-on and can switch Memory off for everyone, which deletes saved memories. Admins don't automatically see members' chats and Business has no Compliance API, but owners can export a member's data. 15161131718
Can people at the maker read it?
Authorised OpenAI staff may access stored chats for engineering support, abuse investigations and legal compliance, and contractors bound by confidentiality may review them for abuse and misuse. OpenAI's sub-processor list names TaskUs, in the Philippines, for moderating Business content. 1114
Enterprise and EduLarge organisations (Enterprise) and universities (Edu)
Does it train on what you type?
No, unless your organisation signs a separate agreement with OpenAI. Thumbs-up or thumbs-down feedback doesn't change that. 111
Can that be switched off?
Nothing to switch off, because training is off by default. Admins can turn off the feedback buttons at Workspace settings > Allow response feedback. 1
How long is it kept?
The workspace can set its own retention period, with a 90-day minimum according to OpenAI Academy; otherwise chats stay until deleted. Deleted chats leave OpenAI's systems within 30 days unless the law requires longer, and on Enterprise, temporary chats stay available to the Compliance API for 30 days. 11125
Where is it stored?
New Enterprise and Edu customers can have chats, files and memory stored in Australia at no extra cost. The AI processing can still happen overseas, because OpenAI keeps it in-region only for Europe, the US and the UAE. 9
What can an admin control?
Admins control retention, apps, memory, roles, single sign-on and SCIM, and can pull chats and audit logs through the Compliance API. Only a workspace owner can grant access to conversation messages through it. 111661319
Can people at the maker read it?
OpenAI says authorised staff only access these conversations to resolve incidents, to recover a user's chats with your explicit permission, or where the law requires. It may also run business data through automated safety classifiers. 11
API PlatformBusinesses that build their own tools on OpenAI's models
Does it train on what you type?
No. Since 1 March 2023, data sent to the API isn't used to train OpenAI's models unless the organisation opts in to share it. 2021
Can that be switched off?
Nothing to switch off. Data sharing with OpenAI stays off for every organisation until an owner turns it on in the organisation settings. 21
How long is it kept?
API inputs and outputs sit in abuse-monitoring logs for up to 30 days, while stored responses are kept at least 30 days and files until you delete them. Approved customers can get Zero Data Retention, but images flagged as possible child sexual abuse material are still kept for review. 20
Where is it stored?
Eligible customers can store API data at rest in Australia, which needs OpenAI's approval for abuse-monitoring controls and adds 10% to prices on newer models. OpenAI doesn't offer in-region processing for Australia, so it may process and briefly store that content elsewhere. 209
What can an admin control?
Organisation owners manage members, projects, API keys and data sharing, and can switch on audit logging under Settings > Data controls > Data retention. Approved organisations choose Zero Data Retention or Modified Abuse Monitoring for the whole organisation or per project. 222120
Can people at the maker read it?
Authorised OpenAI staff may access stored API data for engineering support, abuse investigations and legal compliance, and contractors may review it for abuse. OpenAI's sub-processor list names TaskUs, in the Philippines, for moderating API content. 1114
Claudeby Anthropic
Free, Pro and MaxPersonal accounts
Does it train on what you type?
Yes, unless you switch off "Help Improve our AI models". Chats flagged for safety review and chats you rate with a thumbs up or down can still be used after you opt out, but incognito chats aren't used. 23242526
Can that be switched off?
Yes. Go to Settings > Privacy and turn off "Help Improve our AI models"; chats already used stay in training runs that have started and in models already trained. 24
How long is it kept?
Deleted chats leave your history at once and Anthropic's systems within 30 days, but with training on, de-identified chats can be kept up to 5 years. Incognito chats are kept 30 days, flagged chats up to 2 years and thumbs feedback 5 years, and deleting a chat doesn't delete memories drawn from it. 272829
Where is it stored?
Anthropic's privacy policy says your data goes to its servers in the US or to other countries. Its consumer pages don't mention storing data in Australia. 23
What can an admin control?
No admin: a personal account. You control your own settings and can delete your chats at any time. 3024
Can people at the maker read it?
By default Anthropic staff can't access your chats, unless you send feedback or its Trust & Safety team must review a chat to enforce the Usage Policy. With training on, a small number of staff may review chats after they're de-linked from your account. 3132
Team and EnterpriseWork accounts (Anthropic calls these Claude for Work)
Does it train on what you type?
No, by default. Anthropic may train on a chat only if someone sends it as thumbs feedback or your organisation chooses to allow it. 33
Can that be switched off?
Yes. Training is already off, and a Primary Owner or Owner can also block thumbs feedback with "Rate chats" under Organization settings > Data and Privacy. 33
How long is it kept?
Chats stay until deleted, then leave Anthropic's systems within 30 days, and Enterprise owners can set a retention period of 30 days or more. Incognito chats are kept 30 days for safety, flagged chats up to 2 years and feedback 5 years. 343528
Where is it stored?
Anthropic says business data is stored in the US, and by default requests may be routed to the US, Europe, Asia and Australia. Claude Enterprise has no multi-region data residency, though usage-based Enterprise plans can switch on "US-only inference". 363738
What can an admin control?
Owners control connectors, memory, feedback and single sign-on, and Primary Owners can export the organisation's conversations. Enterprise adds custom retention, audit logs, SCIM and the Compliance API, which pulls chat data and files. 3929334041424344
Can people at the maker read it?
Anthropic says it may process business data for safety review, product support or incident response. Chats its automated systems flag can be kept up to 2 years, and safety analysis linked to accounts is limited to a small number of authorised staff. 363445
EducationUniversity-wide accounts (Claude for Education)
Does it train on what you type?
No. Education runs under Anthropic's Commercial Terms, where chats aren't used for training by default, and thumbs feedback starts switched off. 263346
Can that be switched off?
Nothing to switch off. Thumbs feedback is off by default, and only Anthropic support can turn it on at the Primary Owner's request. 46
How long is it kept?
Anthropic's retention page for commercial products says deleted chats leave its systems within 30 days and flagged chats can be kept up to 2 years. Its Education admin guide points owners to the custom retention controls, which take a period of 30 days or more. 34464735
Where is it stored?
Anthropic's commercial pages say business data is stored in the US and requests may be routed to the US, Europe, Asia and Australia. They don't mention Australian storage for Education. 3646
What can an admin control?
The university's Primary Owner manages accounts and data and decides which features members can use. Data exports, audit logs and thumbs feedback start switched off, and the Primary Owner must ask Anthropic support to enable them. 46
Can people at the maker read it?
Anthropic's commercial pages say it may process data for safety review, product support or incident response, and flagged chats can be kept up to 2 years. Its Education page says nothing more about staff access. 363446
API (Claude Platform)Businesses that build their own tools on Claude
Does it train on what you type?
No. Anthropic's Commercial Terms say it may not train models on customer content, and its privacy pages allow training only on feedback you send or data you choose to share. 483349
Can that be switched off?
Nothing to switch off for API traffic. In the Console, admins can turn off thumbs feedback under Settings > Privacy controls. 50
How long is it kept?
Anthropic deletes API inputs and outputs within 30 days, unless you use a feature that keeps data longer, like the Files API, or a request is flagged, when it can keep it up to 2 years. Approved customers can get zero data retention, except on "covered models" such as Claude Fable 5, which need 30 days. 345152
Where is it stored?
Anthropic's own API stores data in the US and runs requests either globally or US-only, which costs 1.1 times as much on newer models. Its regional compliance page says residency in Australia is available through cloud platforms, listing Asia-Pacific for AWS Bedrock and Google Cloud Vertex. 5354
What can an admin control?
Console admins manage members, roles and API keys, and can limit workspaces to US-only processing. Settings > Privacy controls holds the feedback switch and shows whether zero data retention applies. 55535052
Can people at the maker read it?
Anthropic may process API data for safety review, product support or incident response, and flagged content can be kept up to 2 years even under zero data retention. For covered models, staff can't read retained data by default, and only a small set of approved reviewers can, with every access logged. 365156
Microsoft Copilotby Microsoft
Updated Copilot app (from 18 August 2026)Personal Microsoft accounts, with or without a Microsoft 365 subscription
Does it train on what you type?
Microsoft says prompts, responses and file contents in the updated app aren't used to train foundation models. Its Privacy Statement adds that Copilot uses prompts to improve its services, including advertising. 5758
Can that be switched off?
The updated app's privacy controls list no training switch. Under Settings > Personalization you can turn off 'Saved memories', 'One shared experience' (which lets your chats personalise Bing, Edge and MSN) and 'Allow ads personalization'. 59
How long is it kept?
Microsoft's pages for the updated app give no storage period, and its Privacy Statement says it keeps data as long as necessary. Delete a chat with '...More' > 'Delete', or all history from the Microsoft privacy dashboard. Turning off 'Saved memories' doesn't delete saved ones. We found no temporary chat. 575958
Where is it stored?
Microsoft's Privacy Statement says personal data may be stored and processed in your region, the US or wherever Microsoft or its providers operate. Australia is one of its major data centre locations, but its pages make no promise to keep personal Copilot chats here. 58
What can an admin control?
No admin: a personal account. An employer can stop staff signing in to the Copilot app with personal accounts by using Microsoft's Tenant Restrictions v2. 5960
Can people at the maker read it?
Yes. Microsoft says people may review your Copilot conversations for safety, legal, product-improvement or troubleshooting reasons. We found no way to opt out. 61
Older Copilot app (not yet updated)Personal Microsoft accounts on a Copilot app that hasn't been updated since 18 August 2026
Does it train on what you type?
Yes, unless you opt out. Microsoft's older-app FAQ says it uses Copilot conversations, including voice and uploaded files, for AI training. It excludes under-18s, people who opt out and users in six named countries. Australia isn't one of them. 62
Can that be switched off?
Yes. Turn off 'Training on conversation activity' and 'Training on voice conversations'. On copilot.com: profile icon, your name, then Privacy. Windows or Mac app: profile icon > Settings > Privacy. Mobile: menu, profile icon, Account > Privacy. Microsoft can still use chats for safety, advertising and general product improvement. 63
How long is it kept?
Microsoft keeps conversation history for 18 months by default. You can delete single conversations or your whole history at any time. Uploaded files are kept no longer than 18 months. Images shared with Copilot Vision aren't stored after the session, though text transcripts may be. 6263
Where is it stored?
Microsoft's Privacy Statement says personal data may be stored and processed in your region, the US or wherever Microsoft or its providers operate. Australia is one of its major data centre locations, but its pages make no promise to keep personal Copilot chats here. 58
What can an admin control?
No admin: a personal account. An employer can stop staff signing in to the Copilot app with personal accounts by using Microsoft's Tenant Restrictions v2. 6360
Can people at the maker read it?
Yes. Microsoft says some conversations get automated and human review for product improvement and safety, with no opt-out. Outside research organisations may also review conversation logs. 62
Copilot Chat and the paid Microsoft Copilot licenceWork and school accounts, including students 13 and over (formerly called Microsoft 365 Copilot Chat and Microsoft 365 Copilot)
Does it train on what you type?
No. Microsoft says prompts, responses and the work data Copilot reads through Microsoft Graph aren't used to train foundation models, including those Copilot uses. Feedback isn't used to train them either. 6465
Can that be switched off?
Not needed: there's no training to switch off. Admins can still turn feedback off, or stop it collecting content samples such as prompts and files. 6566
How long is it kept?
Chats are logged in each person's mailbox and kept until deleted; uploaded files are saved to their OneDrive. People can delete their history in My Account, though Microsoft says full deletion can take time. Admins' Microsoft Purview retention policies can keep deleted and temporary chats searchable until they expire. Saved memories stay until the person deletes them. 646567686970
Where is it stored?
If your Microsoft 365 was set up in Australia, Microsoft promises to store Copilot chats here; schools need the Advanced Data Residency for Education add-on for that. Processing may happen in the US, the EU or elsewhere; the EU Data Boundary covers EU users only. Microsoft expects Australian processing by the end of 2026, excluding Anthropic and OpenAI-run models. 657172737475
What can an admin control?
Admins can block Copilot Chat, turn off web search ('Allow web search in Copilot') and memory ('Enhanced personalization'), choose who gets Anthropic and OpenAI-run models (on by default for most organisations), allow or block agents, and set retention, audit and eDiscovery in Microsoft Purview. Copilot only shows data a person can already open. 6065707475
Can people at the maker read it?
Microsoft says Copilot has opted out of the abuse monitoring, with human review, that its Azure OpenAI service offers. If someone sends thumbs-up or thumbs-down feedback with content samples, Microsoft may see it, and the organisation's IT admin can view it too. 6566
Google Geminiby Google
Free and paid personal plansPersonal Google accounts
Does it train on what you type?
Yes, by default for adults. With 'Keep Activity' on (the default from age 18), Google uses your chats and uploads to improve its services, including training generative AI models. With it off, future chats aren't used for training unless you send feedback. Temporary chats aren't used for training. 7677
Can that be switched off?
Yes. At gemini.google.com go to Settings & help > Activity, select 'On' near the top, then 'Turn off' (or 'Turn off and delete activity'). For a one-off chat, select 'Temporary chat' next to 'New chat'. Some connected apps, including Google Workspace, stop working with Keep Activity off. 777876
How long is it kept?
With Keep Activity on, chats are deleted after 18 months unless you choose 3 or 36 months or no auto-delete. With it off, or in a temporary chat, Google keeps chats up to 72 hours. Chats picked for human review are kept up to three years, even if you delete them. 7677
Where is it stored?
Google's Privacy Policy says your information may be processed on servers outside the country where you live. The Gemini Apps Privacy Hub doesn't say where chats are stored, and we found no Australian storage option for personal accounts. 7976
Yes. Google says trained reviewers, including outside service providers, read a subset of chats, disconnected from your account. Turning Keep Activity off, or using a temporary chat, stops review for improving Google's AI. Google says chats can still be reviewed to protect it, its users and the public. 76
Workspace Business, Enterprise and Education editionsWork and school Google accounts using the Gemini app or Gemini in Gmail, Docs and other Workspace apps
Does it train on what you type?
No, unless your organisation gives permission. Google says chats, uploaded files, prompts and Workspace content aren't used to train generative AI models outside your organisation without permission. Feedback isn't used to train them either. 808182
Can that be switched off?
Not needed: Google doesn't train on this content without your organisation's permission, so there's no switch to find. Admins can still turn the Gemini app and Gemini in Workspace apps off. 8081
How long is it kept?
Gemini app chats are kept 18 months by default; admins can pick 3 or 36 months, or turn 'Gemini conversation history' off, leaving chats held up to 72 hours. Side-panel chats in Gmail and Docs are kept for a period admins set, from 90 days to indefinitely. People can delete chats where admins allow. Temporary chat isn't offered. 818078
Where is it stored?
Google's data regions offer only the United States or Europe, so there's no Australian option. Business Standard and Plus can keep Gemini prompts and responses stored in one of those; Enterprise Plus can also keep processing there; Education Standard and Plus get storage only. Otherwise Google's pages don't say where Gemini processes your data. 838480
What can an admin control?
Admins can switch the Gemini app, and Gemini in Gmail, Docs, Drive, Meet and Chat, on or off for chosen groups, set conversation history and retention, control whether Gemini can reach Workspace apps, and see Gemini activity in audit logs. Some Gemini app features are restricted by default for students marked under 18. 8081
Can people at the maker read it?
No, on core-service editions: Google says chats and uploaded files won't be read by human reviewers without permission. On editions where the Gemini app is only an 'additional service', such as Essentials Starter and Workspace Individual, Google says chats may be reviewed by humans and used to improve its products. 8081
Gemini Notebook (formerly NotebookLM)by Google
Free and paid personal plansPersonal Google accounts
Does it train on what you type?
No, unless you send feedback. Google says notebook files, outputs and chats aren't used to directly train its foundational AI models unless you give thumbs up or down feedback. Chats with a notebook inside the Gemini app follow your Gemini 'Keep Activity' setting instead. 8586
Can that be switched off?
Not needed: there's no training switch. Google names thumbs up or down feedback as the exception, and feedback sends your prompts, sources, uploads and outputs, including Gemini chats in the notebook, even with Keep Activity off. 85
How long is it kept?
Notebooks, sources and chat history stay until you delete them; we found no set time limit. Clear chat with 'Delete Chat History' in the chat panel's three-dots menu. Deleting a notebook removes its sources. Turning off Gemini's Keep Activity doesn't delete notebook data. Reviewed feedback is kept up to three years. 878685
Where is it stored?
Google's Privacy Policy says your information may be processed on servers outside the country where you live. Gemini Notebook's privacy page doesn't say where data is stored, and we found no Australian option. 7985
Yes, if you send feedback: Google says trained teams review thumbs up or down feedback and the content sent with it, disconnected from your account. Chats with a notebook inside the Gemini app can be human-reviewed under Gemini's rules. Google also says it may process data to prevent fraud and abuse. 85867688
Workspace Business, Enterprise and Education editionsWork and school Google accounts where Gemini Notebook is a core service
Does it train on what you type?
No. Google says uploads, questions and the model's responses in Gemini Notebook won't be used to train AI models where it's a core Workspace service. 808589
Can that be switched off?
Not needed: Google says this content isn't used for training, and thumbs up or down feedback doesn't change that. 85
How long is it kept?
Google's Workspace Privacy Hub says notebook questions and answers aren't kept after the session ends. Uploaded files and notebooks follow Google's Workspace data-deletion terms, and people can delete or export them. Files added from Drive are copied and stored with the notebook, not in Drive. 80
Where is it stored?
Google says your organisation's data region settings don't apply to Gemini Notebook, so even a US or Europe choice doesn't cover it. Its pages don't say where notebook data is processed or stored, and we found no Australian option. 8083
What can an admin control?
Admins can turn Gemini Notebook on or off for everyone or chosen groups, apply Context-Aware Access, and export data with the Data Export tool. It isn't connected to Workspace data loss prevention (DLP), but Information Rights Management can stop Drive files being added. 80
Can people at the maker read it?
No. Google says uploads, questions and responses won't be read by human reviewers, even when you send thumbs up or down feedback. On editions where it's only an 'additional service', such as Essentials Starter and Workspace Individual, Google may review feedback in full context. 8589
Perplexityby Perplexity AI
Free, Pro, Education Pro and MaxPersonal accounts
Does it train on what you type?
Yes. On Free, Pro and Max, the 'AI data retention' setting starts switched on, and while it's on Perplexity uses your data to train its AI models. Content saved in Memory can be used too. Perplexity says outside model makers such as OpenAI and Anthropic can't train on Perplexity data. 909192
Can that be switched off?
Yes. Turn off 'AI data retention' under Account settings > Preferences. It only covers data collected after you switch it off; anything already collected for training can't be removed. Each person sets it for their own account. 909391
How long is it kept?
Perplexity keeps your threads while your account is active, until you delete them; deleting your account removes personal data within 30 days. Uploaded files are kept 30 days. Incognito threads expire within 24 hours and logged-out threads after 14 days. Cleared memories may be logged for up to 30 days. 9194959697
Where is it stored?
Perplexity says your data may be stored and processed outside your home country, 'including, but not limited to, the United States'. Its privacy notice doesn't mention Australia, and we found no Australian storage option. 98
What can an admin control?
No admin: a personal account. Perplexity's consumer terms licence these plans for 'personal, non-commercial use only'. 99
Can people at the maker read it?
Perplexity's privacy notice doesn't say whether staff read conversations. Its terms let it 'remove, screen, edit, or delete' your content at any time, and its security page says engineers get temporary access to sensitive resources only when needed, such as for debugging. 9810099
Enterprise Pro and Enterprise MaxWork accounts
Does it train on what you type?
No. Perplexity's enterprise terms say it won't use customer content, or let anyone else use it, to train or improve any generative AI model. Its help centre says Enterprise data, including Memory, is never used for training. 10190102
Can that be switched off?
Not needed: there's nothing to switch off. If your organisation drops to a Free or other non-Enterprise plan, the consumer policy applies and 'AI data retention' is on by default until each person turns it off. 90
How long is it kept?
Files attached to sessions are deleted after 7 days; sessions stay until deleted. Organisations with 50+ seats or one Enterprise Max seat can set a session retention window, and admins can enforce Incognito Mode (sessions kept 24 hours). After the contract ends, personal data is deleted within 30 days. 103104105106107
Where is it stored?
Perplexity's Trust Center lists 23 subprocessors, including AWS, Google Cloud and Microsoft Azure, all located in the US. We found no data residency option and nothing about Australia. 108
What can an admin control?
Admins manage members and roles, enforce SSO, set file, sharing and connector permissions, switch off Memory and enforce Incognito Mode. With 50+ seats, or any Enterprise Max seat, they also get session retention controls, audit logs sent to a webhook, SCIM and a usage dashboard. 102107109110
Can people at the maker read it?
Perplexity says access to production data needs authentication and access control, and engineers get temporary access only when needed, such as for debugging. Its enterprise terms also require your organisation to tell staff their inputs and outputs may be reviewed by the organisation. 100101111
Meta AIby Meta
FreePersonal accounts in the Meta AI app, WhatsApp, Instagram, Facebook and Messenger
Does it train on what you type?
Yes. Meta says your chats with its AI, including questions and images you ask for, can be used to train AI models. It also uses them to personalise your experience and ads. Private messages with friends aren't used unless someone shares them with Meta AI. 112113114115
Can that be switched off?
There's no setting. Meta's Privacy Centre links 'objection request' forms (one for your AI chats and public posts, one for WhatsApp AI messages), but Meta says objections are possible 'in certain countries' without naming them. We couldn't confirm they cover Australia. 113116117
How long is it kept?
Meta decides retention case by case and gives no set period for AI chats. To delete Meta's copy, type /reset-ai in a Messenger, Instagram or WhatsApp chat with an AI (/reset-all-ais for all); deleting a WhatsApp chat alone doesn't. The Meta AI app has 'Delete all chats and media'. Incognito chats aren't saved. 112118119120121122
Where is it stored?
Meta's privacy policy says your information is stored and processed where it has data centres, 'including the United States, Ireland, Denmark and Sweden', and in other countries. It doesn't mention Australia or offer local storage. 118
What can an admin control?
No admin: a personal account. Meta AI comes with personal Meta, WhatsApp, Instagram and Facebook accounts under terms between you and Meta; we found no plan that lets an employer manage staff use. 112
Can people at the maker read it?
Yes. Meta's AI terms say it sometimes reviews your conversations with its AIs, automatically or by people, sometimes through outside vendors. Reviewers work at Meta or for trusted service providers. Incognito chats and other Private Processing features can't be read by Meta. 112118120123
Yes. DeepSeek's privacy policy says it uses your inputs to train and improve its models. Its terms say it may use a 'minimal' amount of de-identified inputs and outputs unless you turn off 'Improve the model for everyone'. 124125126
Can that be switched off?
Yes. Turn off 'Improve the model for everyone'. The terms name the switch but not where it sits in settings. The privacy policy also lists a right to opt out of training, depending on where you live, by emailing privacy@deepseek.com. 124125
How long is it kept?
DeepSeek keeps your inputs for as long as you have an account, and longer where it has a legal or business reason. You can delete chat history in settings. After you delete your account it can still keep data the law requires. It names no time periods. 124125
Where is it stored?
DeepSeek says it collects, processes and stores your personal data in the People's Republic of China. We found no option to keep it in Australia or anywhere else. 124
What can an admin control?
No admin: a personal account. DeepSeek's paid service is its developer API platform; we found no business plan for the chat app. 124
Can people at the maker read it?
DeepSeek doesn't say plainly whether staff read chats. Its terms let it use 'technical means to review the behavior of users', and its training notes say its research team builds some training examples from de-identified user input. 125126
Grokby SpaceXAI (xAI)
Free, SuperGrok and The X Subscription (Plus, Premium, Super, Ultra)Personal accounts in the Grok app and on grok.com
Does it train on what you type?
Yes. SpaceXAI says it may use your prompts and Grok's replies to train its models unless you opt out. Private Chat isn't used for training. If you use Grok without logging in, in some regions outside the EU and UK you can't opt out. 127128
Can that be switched off?
Yes. In the app: Settings > Data Controls > 'Improve the model'. On grok.com: Settings > Data > 'Improve the Model'. It applies to new conversations only, and feedback you choose to send, such as a thumbs rating, may still be used for training. 129127
How long is it kept?
Chats stay in your history until you delete them. Deleted chats, deleted accounts and Private Chat conversations are removed within 30 days, unless de-identified or kept for safety, security or legal reasons. SpaceXAI may delete accounts left inactive for more than 120 days. 129127128130
Where is it stored?
SpaceXAI is a US company and its privacy policy names no storage location. Its published subprocessor list puts its cloud providers, including AWS, Google Cloud and Oracle, in the US. We found nothing on Australian storage. 129131
What can an admin control?
No admin: a personal account. But if you sign up with a work email, your account may be linked to your employer's Grok Enterprise subscription, and its admin may access, monitor and control it. 128
Can people at the maker read it?
Yes. SpaceXAI says a limited number of authorised staff may review your Grok conversations to improve the model, investigate security incidents or misuse, and meet legal obligations. Automated classifiers also check for misuse. 127128130
Grok on XAny X account, free or paid (run under X Corp.'s terms)
Does it train on what you type?
Yes. X says it may share your Grok chats on X, and your public posts, with xAI to train and fine-tune Grok unless you opt out. Making your account private also keeps your posts out of training. 132
Can that be switched off?
Yes. In X: Settings > Privacy & Safety > Data sharing and personalization > Grok & Third-party Collaborators, then untick 'Allow your public data as well as your interactions, inputs, and results with Grok and xAI to be used for training and fine-tuning.' Feedback you send may still be used. 132
How long is it kept?
X gives no set period for Grok chats; its privacy policy keeps usage information for the life of your account. Delete all Grok history at once under Privacy & Safety > Data sharing and personalization > Grok > 'Delete Conversation History'. Deleted chats go within 30 days unless kept for security or legal reasons. 132133
Where is it stored?
SpaceXAI says Grok on X falls under X's privacy policy. X says it moves information across borders and uses data centres and providers 'in many parts of the world'. For Australians the data controller is X Corp. in the US. Nothing on Australian storage. 129133
X doesn't say whether people read your Grok chats on X. It says Grok was trained on data sets 'reviewed and curated by AI Tutors who are human reviewers', and your chats can go to xAI for training unless you opt out. 132
Grok Business (Teams) and Grok EnterpriseWork accounts
Does it train on what you type?
No by default. The enterprise terms say SpaceXAI won't use your content to train any AI model, 'subject to disclosures to Customer and Customer-controlled user settings'. SpaceXAI's guide ties these terms to the team workspace, so chats in a personal workspace may not be covered. 134135
Can that be switched off?
Not needed for the team workspace. Staff should check the workspace selector on grok.com: SpaceXAI's guide says to use the team workspace so conversations are 'protected under business plan terms'. 135
How long is it kept?
The enterprise terms delete content within 30 days after each session, unless a different period is agreed, chosen by your organisation, shown in the product, or needed for legal or safety reasons. Teams and Enterprise both list 'Custom Chat data retention'. 134136
Where is it stored?
SpaceXAI's subprocessor list puts its cloud providers in the US, with one support subsidiary in the UK, and its DPA allows processing in the US and other countries. Enterprise sales offer 'Data residency' but name no regions. Nothing on Australia. 131137136
What can an admin control?
Teams: billing, usage analytics, team-wide privacy and chat-retention settings, sharing limits, and admin approval of connectors; the pricing page also lists SSO. Enterprise adds SCIM, audit logs, spend and MCP controls, switching off personal workspaces, and Enterprise Vault with customer-managed keys. 136135138139140
Can people at the maker read it?
SpaceXAI's enterprise FAQ, written for API customers, says classifiers check for misuse and a limited number of authorised staff may review business data when legally required, such as for security incidents. Your organisation's admin may access and monitor accounts on your work domain. 128134141
What we couldn’t establish.
OpenAI's Enterprise privacy page says Business admins can control how long data is kept, but its pricing page lists custom retention only under Enterprise and its help article on removed members says Business chats are kept indefinitely; we found no Business retention setting.
OpenAI's Enterprise privacy page says Business admins can view, export and delete members' chats, while its Business help articles say admins don't automatically see private chats but owners can export a member's data; we used the help articles.
The 90-day minimum for Enterprise retention appears on OpenAI Academy; the help centre article on removed members gives 90 and 180 days only as examples.
OpenAI's training article names Free, Plus and Pro; Go appears only on OpenAI's pricing page, which shows the same opt-out.
OpenAI's New York Times page says it holds limited April to September 2025 user data but doesn't say which plans it comes from; the earlier court order covered Free, Plus, Pro, Team (now Business) and API data without Zero Data Retention.
OpenAI's Enterprise privacy page, Services Agreement and Data Processing Addendum don't say which country stores Business chats by default; its sub-processor list gives only the countries where its cloud providers process data.
Anthropic's consumer training article and its 28 Aug 2025 Consumer Terms announcement describe training as your choice, made at sign-up, without saying which way the switch starts; its privacy policy says it trains unless you opt out, and the pop-up pictured in that announcement shows the switch on, so we marked it yes-default.
Anthropic names the consumer training switch three ways: "Help Improve our AI models" (settings guide), "Model Improvement" (training article) and "You can help improve Claude" (2025 pop-up); we used the settings guide.
Anthropic's privacy centre says API inputs and outputs are deleted within 30 days, while its API and data retention docs say conversation content isn't retained by default except for covered models; we used the privacy centre figure.
Anthropic's custom retention article says the feature is for Enterprise plans, and its commercial retention page doesn't name Education; we applied both to Education because the Education admin guide links to the retention controls and Anthropic calls Education a commercial product.
Anthropic's commercial pages don't list when its staff may read Team, Enterprise or API chats beyond safety review, product support and incident response.
Anthropic's Console feedback page doesn't say whether the thumbs feedback switch starts on or off.
Anthropic offers no Australian storage in its own apps or API; its March 2026 Sydney office announcement says only that it is exploring more computing capacity in Australia, with no date.
How long the updated personal Copilot app keeps chats: Microsoft's activity history and privacy controls pages and its Privacy Statement give no period.
Whether Microsoft trains any models other than foundation models on chats in the updated personal Copilot app: its activity history page rules out foundation models only, while its Privacy Statement says Microsoft may use data to develop and train its AI models.
Whether deleted chats in either version of the personal Copilot app are held for a time before they're erased: neither the older-app Privacy FAQ nor the new activity history page says.
Whether the updated personal Copilot app has a temporary chat or any opt-out from human review: its privacy controls page lists neither, and only the older-app Privacy FAQ addresses human review, saying there is no opt-out.
When the older personal Copilot app stops working: Microsoft's pages say the updated app arrived on 18 August 2026 but give no end date for the older version.
How long Microsoft takes to erase a work account's Copilot history after the person deletes it when no retention policy applies: its deletion page says only that it might take some time.
Whether Australian schools whose Microsoft 365 is already hosted here keep Copilot data in Australia without the Advanced Data Residency for Education add-on: Microsoft's Product Terms Data Residency page lists Copilot as covered for education but also lets Microsoft store education data in the EU or North America without the add-on.
Whether Microsoft has started processing Copilot in Australia: the newest statement we found is the 3 April 2026 editor's note on its in-country processing announcement, which expects it by the end of 2026.
Whether Anthropic or OpenAI staff can see Copilot prompts when those models are used: Microsoft's two subprocessor pages cover terms and retention but don't mention human review.
Where Google stores or processes personal Gemini and Gemini Notebook data: the Gemini Apps Privacy Hub and the Gemini Notebook privacy page don't say, and Google's Privacy Policy only says it keeps servers around the world.
Where Gemini processing happens for Workspace customers without data regions: the Generative AI in Google Workspace Privacy Hub says only that it runs inside Google's own infrastructure.
Whether Gemini Notebook keeps chat history for Workspace accounts: the Workspace Privacy Hub says prompts and responses aren't kept after the session, but the 'Use chat in Gemini Notebook' help page says chat history is kept until you delete it, so we used the Privacy Hub for the Workspace row.
How long Google keeps personal Gemini Notebook notebooks, sources and chat history, and how long deletion takes: its privacy and chat help pages give no period.
What permission would let Google train on Workspace data, and how an organisation would give it: the Workspace Privacy Hub says 'without permission' and mentions new feedback and data donation flows added on 1 July 2026 without describing them.
Google words the Workspace training promise two ways: the Gemini app admin page's comparison table and the Education FAQ say chats won't be used to improve AI models, while the Privacy Hub adds 'outside your domain without permission', and we used the Privacy Hub's wording.
Whether Keep Activity is on by default for personal Gemini users under 18: Google's activity help page gives the default only for people aged 18 and over.
Perplexity's 'Data Collection at Perplexity' page names Free, Pro and Max as training by default but not Education Pro; we grouped Education Pro with them because Perplexity's 'What is Education Pro?' page says it includes everything in Pro.
Perplexity's privacy notice, terms and help centre don't say whether staff read conversations on personal plans for safety or feedback.
Perplexity's 'Account Deletion' and 'Self-Serve Data Deletion' pages don't say how long a deleted thread is held before it's purged; the only period given is 30 days for deleting a whole account.
Perplexity's help centre doesn't say whether searches made while logged out are used for training; its 'Where did my sessions go?' page only says logged-out threads expire after 14 days.
We found no data residency option for Perplexity Enterprise, and no mention of Australia, on its Trust Center, enterprise terms, data processing addendum or help centre.
Meta's Privacy Centre guide 'AI at Meta' shows 'Submit an objection request' forms, but Meta's help page 'Information for people who don't use Muse' says objections apply 'in certain countries' without listing them, and the form we opened is tagged for the EU region in its page code, so we couldn't confirm Australians can use it.
Meta's privacy policy and AI terms give no retention period for Meta AI chats or say how long Meta's copy lasts after a reset; the AI terms warn that deleting messages, threads or your account may not delete Meta's copy.
Meta's help pages say incognito chats aren't available everywhere, and WhatsApp's 'About Private Processing' page says only English, Indonesian, Portuguese and Spanish are supported; neither says whether incognito chats are offered in Australia.
WhatsApp's help page 'About Meta One AI benefits' describes Meta One Core and Premium subscriptions with higher AI image and video limits in 'limited locations'; we found no separate data terms for them and couldn't confirm they're sold in Australia.
Meta's Muse agent, announced on 8 September 2026 with its own privacy policy and paid plans, is rolling out in the US first according to Meta's newsroom, and isn't covered here.
DeepSeek's terms name the 'Improve the model for everyone' switch, but no DeepSeek page we found says where it sits in the app or website, or states plainly that it starts switched on; the wording 'opt out by turning off' implies it does.
DeepSeek's 'Model Mechanism and Training Methods' page sends readers to the privacy policy for how to opt out of training, but the policy only lists an opt-out right and an email address; the switch is named only in the terms of use.
DeepSeek's privacy policy and terms give no retention period for chats, don't say how long deleted chats are held, don't mention a temporary or incognito chat, and don't say whether people review conversations for safety.
The Australian Government's PSPF Direction 001-2025 (published 4 February 2025, updated 14 February 2025) requires Australian Government entities to prevent the use or installation of DeepSeek products, applications and web services and to remove existing instances from government systems and devices: https://www.protectivesecurity.gov.au/publications-library/direction-001-2025-deepseek-products-applications-and-web-services
xAI's site now names the company SpaceXAI LLC ('SpaceXSI') and carries a post titled 'xAI joins SpaceX' (2 February 2026), while X's help pages still say xAI; we found no formal rename notice, so we've written 'SpaceXAI (xAI)'.
SpaceXAI's pricing page calls the team plan 'Teams' while its business page and docs say 'Grok Business'; both are priced per user and we've treated them as one plan.
SpaceXAI's Grok Business user guide says custom retention policies are for the Enterprise tier, but its pricing page lists 'Custom Chat data retention' for Teams as well; we used the pricing page.
SpaceXAI's organisation docs say SSO is for Enterprise subscribers only, while its pricing page lists SAML/OIDC SSO for Teams and its security page offers SSO for 'Business Tier accounts'.
SpaceXAI's Grok Business announcement promises 'no training on it, ever', but its enterprise terms make the promise 'subject to disclosures to Customer and Customer-controlled user settings' and its enterprise FAQ mentions free credits in exchange for permission to train; we used the terms.
SpaceXAI's Grok Business guide doesn't say which terms cover chats in a personal workspace; it only says to use the team workspace so conversations are 'protected under business plan terms'.
SpaceXAI's statement on staff reviewing business data comes from its enterprise FAQ, dated February 2025 and written for API customers; we found no review statement specific to Grok Business chats.
SpaceXAI's pricing page offers Enterprise customers 'Data residency' through its sales team but names no regions, and its legal and security pages mention no Australian option.
X's 'About Grok' help page and X Privacy Policy don't say whether people review Grok conversations on X or how long Grok chats are kept if you don't delete them.
SpaceXAI's Consumer FAQs page shows the date 12 May 2025 but already uses the SpaceXSI name, so it appears to have been edited since without a new date.
Sources.
Every source was read on the date shown. “Updated” is the date the page itself gives.