Problems & fixes · AI assistants
What happens to the data you type into an AI assistant
Last updated 25 October 2026 · 2 sources listed · Reading time about 3 minutes
Short answer
Content you enter may be stored, reviewed by staff, used to improve models or surfaced through sharing features, depending on the product and plan. The risk is rarely a dramatic breach; it is ordinary retention and unintended disclosure. Learn the retention, training and sharing rules for your exact plan, and keep regulated, confidential and credential data out unless the agreement permits it.
Who this is for: Someone who has already pasted work or personal material into a chat assistant and wants to understand the exposure, or a manager worried about what staff are entering.
What to evaluate
Retention and deletion
Deleting a chat from the interface does not always mean immediate deletion from backups or logs.
How to check: Look for stated retention windows, deletion timelines and any carve-outs for abuse monitoring or legal obligations.
Training use
Whether your content can shape future models decides whether fragments could ever appear in other users' outputs.
How to check: Find the exact wording for your plan and whether training use is opt-in, opt-out or excluded by contract.
Human review
Some providers allow staff or contractors to read conversations for safety or quality purposes.
How to check: Check conditions, scope and whether review can be restricted in your agreement.
Sharing and link features
Shared-conversation links and public galleries have exposed content that users believed was private.
How to check: Test whether shared links are public, who can create them and whether admins can disable them.
Connected tools and plug-ins
Third-party extensions can receive your content under their own terms, outside the assistant provider's commitments.
How to check: List the connectors enabled, read each one's data terms and disable those you cannot justify.
Legal basis and jurisdiction
Personal data about clients, patients or staff brings data protection duties that depend on the jurisdiction.
How to check: Identify the applicable law and regulator for your location and sector, and confirm whether a data processing agreement is in place.
Comparison
| Exposure path | How it arises | Reasonable mitigation |
|---|---|---|
| Provider retention | Conversations kept for history, safety or debugging | Choose plans with short or zero retention options; delete sensitive threads |
| Model improvement | Content used to refine future models | Select plans and settings that exclude training use, and verify in writing |
| Sharing links | A user publishes a conversation to collaborate | Disable public sharing or restrict it to the organisation |
| Third-party connectors | Extensions pass content to other services | Allow only vetted connectors and review their terms |
| Local or browser exposure | Shared devices, synced browser extensions, screenshots | Use managed devices and avoid unvetted extensions that read page content |
Decision worksheet
Rate up to three options against this guide’s own criteria. It runs in your browser and saves only on this device. It is a thinking aid, not a recommendation.
1. Have you checked each criterion?
0 of 6 criteria checked.
2. Compare options (0 = fails or unknown, 3 = meets it well)
| Criterion | |||
|---|---|---|---|
| Retention and deletion | |||
| Training use | |||
| Human review | |||
| Sharing and link features | |||
| Connected tools and plug-ins | |||
| Legal basis and jurisdiction | |||
| Total (max 18) | 0 | 0 | 0 |
Questions to ask before you commit
- Is content from our plan ever used to train or evaluate models, and under what conditions?
- Who at the provider can access conversations and for which purposes?
- How long do backups and logs retain deleted content?
- Can we restrict processing and storage to specific regions?
- Will you sign a data processing agreement and list your subprocessors?
- How are security incidents notified to customers?
- Can administrators disable public sharing, memory features and third-party connectors?
- What happens to our data when we terminate the contract?
Typical trade-offs
Turning off conversation history versus keeping it
- Upside
- Less data stored and a smaller exposure window.
- Downside
- Staff lose the ability to return to earlier work, which can push them to retype material elsewhere.
Redacting inputs before use versus trusting the contract
- Upside
- Redaction protects you even if terms or settings fail.
- Downside
- It costs time, can remove context the assistant needs and is easy to skip under deadline pressure.
Central blocking of consumer tools versus education only
- Upside
- Blocking gives consistent enforcement.
- Downside
- Blocking without a usable alternative drives use onto personal devices.
Common mistakes
- Pasting credentials, keys or customer lists to debug a problem. Use placeholders and keep secrets out of any prompt.
- Assuming deletion is instant and total. Check the stated retention and request deletion through the correct channel.
- Ignoring memory or personalisation features that carry details from one chat into another. Review what is stored and clear it.
- Relying on a policy document nobody has read. Pair rules with short, concrete examples of what to keep out.
Related questions
I already pasted something sensitive. What should I do?
Identify what it was and where it went: which product, which plan, whether history and training were enabled. Delete the conversation and any shared links, then use the provider's process for requesting deletion if the content was serious. Tell whoever owns the data, because notification duties for personal data depend on your jurisdiction and sector. Do not wait to see whether anything happens. Record what occurred so your policy can be improved, and treat any credentials that were pasted as compromised and rotate them.
Does turning off training mean my data is private?
No. It addresses one use of your content. The provider may still store it for a period, process it through subprocessors, and allow limited staff access for safety or support. Other exposure paths such as sharing links, connectors and shared devices stay open. Think of the setting as one layer. For sensitive material, combine it with contractual terms, retention limits and a rule about what may be entered at all.
Do data protection laws apply to chat assistant use?
If you enter personal data about identifiable people, data protection rules such as those in the European Union and United Kingdom, and various other regimes, can apply to you as the party deciding the purpose. That generally means having a lawful basis, informing people, limiting use and securing processors under an agreement. Rules vary by country and sector, so consult your regulator's guidance or a qualified adviser rather than relying on a generic answer. Regulators in several jurisdictions publish guidance specific to generative tools.
Is the interface or the API safer for sensitive data?
They are different products with often different terms. Programmatic access is sometimes offered with stricter default retention or training exclusion, but that varies and must be verified in the agreement. Safety also depends on what you build around it: logging you add yourself can store the same sensitive content. Compare the written terms for each access route and examine your own logging, rather than assuming one route is inherently safer.
Sources and further reading
These pages were reachable when we last checked. They informed definitions and rules referred to above; they do not endorse this guide. Where rules differ by country, check the authority for your location.
- Guidance on AI and data protection — UK Information Commissioner's Office, retrieved 27 October 2026
- Data protection — European Commission, retrieved 27 October 2026
Related guides
How to choose · AI assistants
How to choose an AI assistant for a small team
Pick an assistant by the work it must do and the data it will touch, not by how impressive a demo feels. Check whether business plans keep your content out of model training, whether admins can manage seats and sign-in, which file and connector types are supported, how usage is capped, and how easily you could export your history and leave.
Comparison · AI assistants
Free consumer versus paid business AI assistant plans
The practical gap is control, not intelligence. Business tiers typically add contractual data commitments, administrative controls, audit options and support, while consumer accounts offer convenience and personal settings that you cannot enforce across staff. Choose by the sensitivity of the content, the number of users and whether you must demonstrate governance to clients or regulators.
Suitability · AI assistants
Is an AI assistant suitable for research and fact-dependent work
Assistants are strong at structuring a topic, explaining concepts, summarising supplied text and suggesting search angles. They are unreliable as a sole source of specific facts, quotations, citations and recent developments, because fluent text can be fabricated. Use them to accelerate reading and drafting, and verify every claim that carries weight against a primary source.
Requirements · AI coding tools
What an engineering policy for AI coding tools should cover
A workable policy names approved tools and plans, forbids pasting secrets or customer data, sets repository eligibility, requires normal review and testing for all code regardless of origin, defines licence and provenance handling, covers agentic tool permissions, and assigns ownership for monitoring and updates. Keep it short enough that developers actually read it.
Requirements · AI agents
Requirements before giving an AI agent access to company systems
Before granting access, you need a defined task and owner, dedicated low-privilege credentials, a data classification of everything reachable, approval rules, logging with retention, a tested stop procedure, and agreement on who answers for outcomes. Without these, the agent inherits the access of whoever set it up, and nobody can reconstruct what it did.
How to choose · AI risk and governance
How to assess an AI vendor during due diligence
Assess an AI vendor on data handling, security assurance, model and subprocessor transparency, contractual protections, testing and monitoring support, and operational resilience. Request evidence rather than assertions: independent attestations, data flow diagrams, retention terms and incident commitments. Tailor depth to the risk of the use case, and revisit when features or subprocessors change.
Requirements · B2B software
Consent and legal requirements for email marketing
Rules differ by jurisdiction. Some regimes, such as the EU and UK, generally require prior consent for marketing to individuals, while others, such as the United States federal law, allow sending if you honour opt-outs, identify yourself and include a physical address. Always keep consent records, include a working unsubscribe, and check the regulator in each place you market. This is general information, not legal advice.