03 · What You Need to Know
Follow What Happens to the Data From Input to Deletion
Start by Identifying What the Tool Collects
Do not assume the service collects only the words you type into the prompt box.
Depending on the product, collected information may include prompts, uploaded files, generated outputs, feedback, account information, device or browser information, usage data, logs, approximate location information, cookies, integration data, or information obtained through connected services.
For researchers, the critical question is whether any of those categories can contain research information or personal data.
A transcript uploaded as a “file” may contain names, voices converted to text, health information, demographic details, institutional affiliations, or combinations of attributes that make individuals identifiable. The category label in the privacy policy may sound generic while the actual research content is anything but.
Check Why Each Type of Data Is Processed
Collection is only half the question. What does the provider do with the information?
Potential purposes can include providing the requested service, maintaining accounts, preventing abuse, monitoring security, improving products, developing or evaluating models, conducting analytics, personalizing services, complying with legal obligations, or other purposes described by the provider.
Privacy guidance from the UK Information Commissioner's Office illustrates the broader transparency principle: organisations should explain why personal data are processed, the legal basis where applicable, retention, recipients, transfers, and individual rights.
Researchers should therefore read beyond “we value your privacy” and identify the operational purposes that actually apply to prompts and uploaded content.
Find Out Whether Your Inputs May Be Used to Train or Improve AI
This is often one of the first questions researchers ask, but the terminology can vary. A provider may refer to training, model improvement, service improvement, product development, evaluation, human review, quality assurance, or similar activities.
Determine whether prompts, files, outputs, or feedback may be used for these purposes and whether the conditions differ according to product, plan, settings, API use, enterprise agreement, or institutional account.
If an opt-out or account control exists, understand what it actually changes. An option that prevents use for model training may not necessarily eliminate temporary retention, security logging, abuse monitoring, or other processing.
Watch Out
Do not reduce the privacy review to one question: “Is my data used for training?” Information can still be collected, retained, reviewed, shared, transferred, or otherwise processed even when it is excluded from model training.
Check How Long Prompts and Files Are Retained
Retention matters because deletion from your visible chat history and deletion from the provider's systems may not be the same event.
Look for stated retention periods or criteria used to determine them. The ICO identifies retention periods, or the criteria used to determine them, as core privacy information.
Researchers should ask whether retention differs for active conversations, deleted content, backups, security logs, API requests, temporary processing, or organizational accounts.
Where a research protocol promises deletion or restricted retention of participant information, these details can become methodologically and ethically consequential.
Identify Who Can Receive the Information
An AI service may rely on cloud providers, subprocessors, contractors, affiliated companies, safety reviewers, analytics providers, or other third parties.
A privacy policy should help you understand recipients or categories of recipients. ICO guidance similarly identifies recipients and categories of recipients among the information that should be disclosed for personal-data processing.
For confidential research, this matters because uploading information to one interface can result in processing by more than one organization.
Check Where the Data May Be Processed or Transferred
Cloud services can process information across jurisdictions. Privacy rules, institutional requirements, contractual commitments, or research agreements may constrain international transfers or require particular safeguards.
Look for information about countries or regions of processing, international transfers, and safeguards where applicable. The ICO's privacy-information checklist includes details of transfers to third countries or international organisations when relevant.
Do not assume the company's headquarters tells you where your research data will be processed.
Look for Human Access to Content
Researchers sometimes imagine AI processing as entirely automated. Depending on the service and circumstances, authorized personnel or contractors may have access to some content for support, safety, abuse investigation, quality review, legal compliance, or other stated purposes.
Check whether the policy or supporting documentation describes human access and under what circumstances it may occur.
This can be particularly important for material subject to confidentiality commitments. “Processed by AI” does not necessarily mean “never visible to another person.”
Understand Your Deletion and Account Controls
Look for controls governing conversation history, deletion, temporary chats or equivalent modes, model-improvement settings, account deletion, file removal, and data-access requests where available.
Then determine what those controls actually mean. Does deleting a conversation remove it immediately? Is it scheduled for deletion after a defined period? Are legal, security, or backup exceptions described?
User-interface labels are useful, but the governing documentation should explain the underlying treatment of data.
Check the Rights Available to Individuals
Where privacy law applies, individuals may have rights relating to access, correction, deletion, restriction, objection, portability, consent withdrawal, or complaints, depending on jurisdiction and legal basis.
ICO guidance, for example, lists rights such as access, rectification, erasure, restriction, objection, and data portability among information that organizations may need to communicate.
Researchers should not assume that rights available to the account holder automatically resolve obligations owed to research participants whose information the researcher uploads. Your role in collecting and disclosing participant data remains relevant.
Check Whether Different Products Have Different Privacy Conditions
A provider may offer free consumer access, paid individual subscriptions, APIs, team products, enterprise services, educational arrangements, or institutionally negotiated accounts.
Do not assume one privacy statement applies identically across all of them.
Look for product-specific terms, data controls, enterprise documentation, contractual addenda, or institutional agreements. This is one reason free and paid AI services should not be compared through price alone. A different offering may change data conditions, but you must verify that rather than infer it from payment status.
A Privacy Policy Is Not the Same as a Security Assessment
Privacy and security overlap, but they are not identical.
A privacy policy primarily explains how information is collected, used, shared, and managed. Security documentation may address encryption, access controls, certifications, incident response, authentication, infrastructure, or other technical and organizational safeguards.
If your project involves sensitive or regulated information, a public privacy policy may be only one part of the assessment required by your institution.
A Privacy Policy Is Also Not the Same as the Terms of Service
Privacy documentation explains data practices. Terms of service usually govern the contractual relationship between the user and provider and may address intellectual property, licenses, permitted use, liability, account restrictions, warranties, dispute terms, and other matters.
Researchers should therefore separately examine what the AI tool's terms of service allow and require.
Your Research Obligations Can Be Stricter Than the Provider's Policy
A provider can truthfully state that it processes information according to its privacy policy while your own research protocol still prohibits the upload.
For example, participant consent may restrict disclosure. An ethics approval may specify particular storage arrangements. A data-use agreement may prohibit transfer to third parties. A collaborator may have provided confidential information under contractual conditions. Institutional policy may permit only approved services.
UNESCO's guidance on generative AI in education and research identifies data privacy as a major concern and calls for protection of personal data and institutional validation of generative AI tools. The European Commission's updated 2026 guidelines likewise maintain accountability, transparency, responsibility, and research integrity as central principles and address risks arising when AI is involved in research information management.
| Privacy-policy question |
What to look for |
Why researchers should care |
| What is collected? |
Prompts, files, outputs, account and usage information |
Research content may contain personal or confidential information |
| Why is it processed? |
Service delivery, security, analytics, product or model improvement |
The same data may be used for purposes beyond answering your prompt |
| Is it used for model improvement? |
Default treatment, opt-outs, product-specific exceptions |
May affect whether non-public research material is appropriate to provide |
| How long is it retained? |
Retention periods, deletion timelines, exceptions |
May conflict with research data-management commitments |
| Who receives it? |
Affiliates, service providers, subprocessors, human reviewers |
Uploading may disclose information beyond the company named on the interface |
| Where is it processed? |
International transfers and applicable safeguards |
Jurisdictional or institutional restrictions may apply |
| What controls exist? |
Deletion, opt-outs, temporary modes, access and privacy settings |
May allow researchers to reduce unnecessary processing |
| Which policy applies? |
Consumer, API, team, enterprise, or institutional conditions |
Different account types may have materially different data practices |
Privacy Policies Change
AI services and their data practices can evolve. Privacy documentation may be revised as features, providers, laws, or business models change.
For an AI tool used materially in a long research project, record or otherwise document the policy and product conditions relevant when the decision was made, particularly where those conditions were important to ethics, governance, or data management.
Re-check them when the service changes materially or before introducing a new category of research data.
