Download: AI data-use contract terms checklist (PDF)
One printable page: the five contract points on this page, what to look for in each, and the provider terms and law-firm guidance behind them. No form or sign-up. Free to copy and use inside your organization. Issues to check with counsel, not legal advice.
Companion: AI vendor data-handling questionnaire (32 questions to send a vendor, Excel)
Start with the terms you are using AI under
Consumer AI accounts and business or API accounts are governed by different terms. OpenAI's help centre says that when you use its services for individuals (Free, Plus and Pro) it may use your content to train its models unless you opt out, and that by default it does not use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu or its API to improve its models. Anthropic says that by default it will not use inputs or outputs from its commercial products, such as Claude for Work and the Anthropic API, to train its models. Microsoft says that prompts, completions and training data sent to the models it sells through Azure, including Azure OpenAI, are not used to train generative AI foundation models without your permission or instruction. The first control is simply making sure research staff use company-managed accounts on those terms, and that personal accounts are not used for work.
A help-centre page is a description, not a commitment. The commitment sits in the contract your organisation accepted: the provider's services agreement or commercial terms, its data processing addendum (DPA), any service-specific terms, and any order form or amendment you negotiated. The next section lists the contract points worth checking and quotes what the major providers' published terms say about each, as checked on 30 October 2026.
Contract terms to check
These are issues to check, not legal advice, and not a statement of what your own contract says. Published terms change (OpenAI's and Anthropic's commercial terms both allow updates on notice), negotiated agreements can differ from the published versions, and some deal and research agreements add their own restrictions. Ask counsel to confirm how each point applies to your agreements before you rely on it. The five points below are the main places where the contract, rather than a person, decides what happens to proprietary research data.
Training use. Check that the agreement itself, not only a help page, says the provider will not use your content to train or improve its models, and what the exceptions are. OpenAI's Services Agreement (section 4.2) says OpenAI will only use customer content as necessary to provide the services, comply with law, enforce its policies and prevent abuse, and will not use it to develop or improve the services unless the customer explicitly agrees. Anthropic's Commercial Terms (section B) say Anthropic may not train models on customer content from the services. Google Cloud's Service Specific Terms (General Service Terms, section 18, Training Restriction) say Google will not use customer data to train or fine-tune any AI or machine-learning model without the customer's prior permission or instruction. The exceptions are the issue to check. Anthropic says that if a user explicitly reports feedback, for example with a thumbs-up or thumbs-down button, it may use that conversation to train its models and may keep feedback data for up to five years; OpenAI lets API organisation owners enable data sharing. Decide whether feedback buttons and data-sharing settings should be switched off for research work. The law firm Tucker Ellis calls the training restriction perhaps the single most important provision in an AI vendor agreement, and Morgan Lewis lists the alternatives when a vendor asks for training rights: prohibit training entirely, require opt-in consent, exclude confidential or proprietary information, require anonymisation first, and require training to stop at termination.
Retention and deletion. Training is not the only way data stays with a provider. Check how long prompts, outputs, files, logs and fine-tuning data are kept during the contract, and how and when they are deleted after it ends. During the contract, OpenAI's enterprise privacy page says it may keep API inputs and outputs for up to 30 days to provide the services and identify abuse, unless an endpoint or feature is listed otherwise; Anthropic's privacy centre says API inputs and outputs are deleted from its back end within 30 days by default, with listed exceptions, that content flagged as violating its Usage Policy may be kept for up to two years and its safety classification scores for up to seven, and that, if your contract permits it, Anthropic may anonymise your organisation's data for research or statistical purposes and keep it longer. That last clause is one to look for and, for proprietary research data, usually to exclude. After the contract ends, the published terms differ. OpenAI's Services Agreement (section 11.3) says it will delete customer content within 30 days of termination unless the law requires otherwise or the customer agreed otherwise in writing, and that OpenAI may keep 'abusive' customer content (inputs or outputs that break its use restrictions) as reasonably necessary to protect the services or others from harm. Anthropic's DPA (section H.1) says that within 30 days of termination it will return the data on request and delete it, except where the law requires storage, to resolve a dispute, or where retention is necessary to combat harmful use of the services. Microsoft's Products and Services DPA (May 2026) keeps customer data in a limited-function account for 90 days after a subscription ends and then deletes it within an additional 90 days. Google Cloud's data processing addendum (section 6.2) allows a recovery period of up to 30 days after the term ends and then commits to deletion within a maximum of 180 days, so up to about 210 days in all. Tucker Ellis advises setting retention periods for every category, including prompts, outputs, inference logs and metadata, and treating 'in due course' or 'within a reasonable period' as insufficient. Check also which features keep data longer by design, such as stored conversation state, file stores and fine-tuning files, which OpenAI says are kept until you delete them.
Derived or fine-tuned models. If you fine-tune or customise a model with research data, check who may use the resulting model and what happens to it and its training files at the end. OpenAI's enterprise privacy page says fine-tuned models are for your use alone, are never served to or shared with other customers or used to train other models, and that fine-tuning data is retained until you delete it. Microsoft says fine-tuned models are exclusively available to the customer whose data created them and can be deleted by that customer at any time. Google's Service Specific Terms (section 15) give the customer sole access to models it has tuned, including after the agreement ends, and section 20 says generated output is customer data and that Google does not assert ownership of new intellectual property in it. Owning an output is not the same as it being patentable or protected by copyright; that is a separate question. The less obvious issue is derived value. Morgan Lewis points out that even when a company keeps ownership of its raw data, broad training rights can let a vendor build models that incorporate learnings from that data indefinitely, and Loeb & Loeb reports that vendors often seek to keep rights to 'derivative data', which has become a point of negotiation. Check the definitions of usage data, aggregated data and derived data, not only 'customer content'. Check, too, the restrictions that run the other way: OpenAI's agreement (section 3.3) bars using output to develop models that compete with OpenAI except for a narrow permitted exception, and Google's terms restrict using output to create models similar to Google's. These matter if your team plans to train its own models on AI outputs.
Licence and sub-processors. Ownership clauses are usually favourable: OpenAI's agreement (section 4.1) and Anthropic's terms (section B) both say the customer keeps its inputs and owns its outputs. The issue to check is the licence the provider takes back and who else it may pass data to. OpenAI's agreement (section 9.2) grants OpenAI only the limited rights reasonably necessary to deliver the services and says that permission extends to subcontractors and sub-processors. Each provider keeps a sub-processor list and a notice rule, but check what data the rule covers. OpenAI's DPA (sections 2.9 and 2.10) and Anthropic's DPA (sections A.9 and C.3, notice and 15 days to object) apply to personal data, so they may say nothing about a sub-processor that handles only non-personal research data; for that, look to the main agreement's confidentiality and sub-contractor terms. Microsoft's DPA covers customer data generally and gives at least six months' notice of new sub-processors, or 30 days for sub-processors that support artificial-intelligence functionality, with the ability to disable use of that sub-processor until at least six months after the notice. Google's DPA (section 11.4) gives at least 30 days' notice before a new sub-processor starts processing, and the way to object is to terminate the agreement within 90 days. Tucker Ellis advises that any licence the vendor keeps be narrowly scoped, time-limited and explicitly stated, and that equivalent data-protection obligations flow down to each sub-processor. Bristows notes that some vendors ask for a perpetual licence to use all customer inputs to improve the tool, and that granting a licence to confidential data could be treated as disclosing it.
Flow-down to collaborators and CROs. Your own contract with an AI provider does not cover what a partner does with your data. Contract research organisations, academic collaborators, sites and other vendors may run your unpublished results through their own AI tools, under their own terms. Freshfields' 2024 programme materials on clinical trial agreements noted that sites and vendors use AI to negotiate agreements, recruit patients and enter study data, and suggested considering whether to ask the CRO to disclose its AI tools and to audit them in advance. There is a regulatory reason as well as a confidentiality one: ICH E6(R3) says that where trial activities are transferred to service providers, responsibility for the trial, including the quality and integrity of trial data, stays with the sponsor, and that the sponsor should oversee those activities, including ones the provider subcontracts further. In the UK, the Health Research Authority's commercial model agreements (including the model clinical trial agreement and its CRO version), updated in September 2026, now let participating NHS organisations use sponsor and CRO confidential information in their own secure AI tools under conditions stated in the contract without seeking consent; the HRA gives Microsoft 365 and Google Workspace as examples of such tools, and use outside those conditions still needs the owner's consent. Because these templates are meant to be used unaltered with NHS sites, a sponsor's task there is to understand the conditions and plan around them, not to negotiate them. Trethowans suggests that confidentiality agreements consider whether to restrict recipients from feeding confidential information into AI models, and whether the definition of confidential information should cover AI-generated outputs made from it. Where agreements are negotiable, check four things in each outbound agreement: whether the partner may put your confidential information into AI tools at all, and on what terms (no training, retention limits, enterprise accounts only); whether outputs derived from your information are your confidential information; whether the restriction also binds the partner's representatives and the service providers it subcontracts to, which Trethowans and ICH E6(R3) both point towards; and whether you may ask which AI tools are used. Check inbound agreements too. Trethowans suggests deciding your own policy for processing confidential information you receive using AI, and the HRA templates show the default position: AI use outside the agreed conditions needs the owner's consent. Material you receive under a confidentiality, material-transfer or collaboration agreement may limit how it can be processed, including by a cloud AI service, whatever that service's own terms say.
Two further points sit next to these five. First, change of terms: OpenAI's agreement gives at least 30 days' notice of updates it judges material, and Anthropic's updated terms take effect 30 days after posting or notice; both make an exception for changes required by law, which can take effect sooner. Keep a dated copy of the version you rely on and re-check when notified. Second, intellectual property. United States trade-secret law protects information only if its owner has taken reasonable measures to keep it secret, so the contracts and controls on this page can be evidence of those measures; whether they were enough is decided case by case. Whether disclosure to a provider affects patent rights in a particular invention is a question for patent counsel, not for a checklist; expect counsel to ask whether the provider was bound by confidentiality and no-training terms at the time.
| Contract point | What to look for | Example of a provider's published term (checked 30 Oct 2026) |
|---|---|---|
| Training use | A clause in the agreement itself; exceptions for feedback, opt-in data sharing or 'service improvement' | OpenAI Services Agreement s4.2; Anthropic Commercial Terms s B; Google Service Specific Terms s18 |
| Retention and deletion | Periods per data category (prompts, outputs, logs, files, fine-tuning data); deletion deadline after termination; carve-outs | OpenAI Services Agreement s11.3 (30 days); Anthropic DPA s H.1 (30 days); Microsoft DPA (90 + 90 days); Google DPA s6.2 (30-day recovery + up to 180 days) |
| Derived or fine-tuned models | Who may use the tuned model; deletion of tuning files; definitions of usage, aggregated and derivative data; limits on training your own models on outputs | OpenAI enterprise privacy page; Microsoft Foundry data-privacy page; Google Service Specific Terms s15 and s20; OpenAI Services Agreement s3.3 |
| Licence and sub-processors | Scope and duration of the licence back; sub-processor list, notice period and right to object | OpenAI Services Agreement s9.2; OpenAI DPA s2.9 and Anthropic DPA s C.3 (personal data only); Microsoft DPA (6 months, or 30 days for AI sub-processors); Google DPA s11.4 (30 days; object by terminating) |
| Flow-down to collaborators and CROs | Whether partners may use AI tools on your information, on what terms; outputs treated as confidential; flow-down to subcontractors; disclosure of tools | HRA commercial model agreements (September 2026); ICH E6(R3); Freshfields and Trethowans guidance |
Put a gateway between your data and the model
Terms decide what a provider may do with data it receives; they do not decide what your people and agents send. A gateway policy sits in front of every external model call and decides, per request, whether context may leave. Each request gets one of five outcomes: allow, when the request contains nothing sensitive; allow transformed, when identifiers or sensitive values are removed or replaced before the request leaves (the transformation must not change values the answer depends on, such as doses, concentrations or process parameters, and removing names does not make a sequence or chemical structure safe to send, because the structure is the proprietary information); local only, when the request must be served by a model inside your boundary; approval required, when a person decides before anything is sent; and deny, when the content may not go to an external model at all. Every decision is recorded so it can be re-derived later, which also gives you the evidence that the contract terms above were respected in practice. These controls address confidentiality only. If an AI output will form part of a GxP record or decision, that use also needs computerised-system validation, supplier assessment, audit-trail and record controls under the regulations that apply to you, and a fixed, recorded model version.
See: Qualifying an AI vendor as a GxP supplier · Does 21 CFR Part 11 apply to AI tools? · Change control for AI models and prompts in GxP
Know what is proprietary
A gateway is only as good as its definition of what to protect. A proprietary data registry records which assets the organisation considers proprietary, who owns each one, what is already public, which agreements govern it, and what AI use each permits. Recording the governing agreement matters because partner material may carry restrictions that your own policy would not otherwise impose. Persistent sensitivity labels, such as Microsoft Purview sensitivity labels, let that classification travel with the file so AI tools and data-loss prevention can enforce it.
Local, open-weight and federated options
Some material may need to stay inside your boundary whatever the provider's terms, depending on your classification and the agreements that govern it: unpublished sequences, partner-confidential data, or assay results that are central to a programme. For that material the choice is where the model runs, not which contract to sign. Two further checks apply to it. Some sequence, pathogen and toxin technology may be export-controlled (the Export Administration Regulations define what counts as an export, including release to foreign nationals, and the Commerce Control List says what is controlled), so whether sending it to a provider that processes it in other regions or with staff of other nationalities needs a licence is a question for export-control counsel. In the United States, a Department of Justice rule (28 CFR Part 202) restricts some transactions that would give countries of concern access to bulk sensitive personal data, including human 'omic data above set thresholds (for example, genomic data on more than 100 US persons); it also has exemptions, including for de-identified data needed to obtain or keep regulatory approval of a drug, biologic or device, and for data ordinarily part of FDA-regulated clinical investigations, so check with counsel whether it applies to your data and to your providers' and CROs' arrangements. Atlas Bio's answer on running AI models in your own cloud tenant or calling a vendor API compares the options provider by provider; the short version follows.
A model in your own cloud tenant, such as Microsoft Foundry, Amazon Bedrock or Google's Vertex AI, keeps the model inside your existing cloud agreement, region choices, identity and network controls. It does not run on your own servers, and the cloud provider's retention, abuse-monitoring and residency settings still apply, so the contract points above still need checking.
An open-weight model on infrastructure you control is the only option in which no model provider receives your prompts. OpenAI's gpt-oss models, for example, are published under the Apache 2.0 licence, and the model card says the larger one fits on a single 80 GB GPU. The cost is that you now run the controls a provider would run for you: access control, patching, logging and evaluation of whether the model is good enough for the task. If the servers or GPUs are rented, the hosting provider's terms apply instead. A split, with a local model for the sensitive step and a hosted model for the rest, only protects the material if the hosted step receives none of it, including summaries, extracted values or embeddings derived from it.
Federated learning is for a different problem: training a shared model across organisations without pooling the data. Each participant trains on its own data and shares model updates rather than records. The MELLODDY project trained models across the assay data of ten pharmaceutical companies this way, and its paper in the Journal of Chemical Information and Modeling (published online in 2023) reports that partners' own models improved on aggregate without the partners sharing their confidential datasets. Federated learning is not a guarantee of confidentiality, though. NIST notes that published privacy attacks can extract a surprising amount of information about training data from the shared model updates or from the finished model, so a federated project still needs its own agreement on who may see what, and protections chosen for the data involved.
See: Own cloud tenant or vendor API? AI data-control trade-offs for biotech
Keeping research data under your control starts with knowing where it goes today and which terms govern it. Atlas Bio's Secure AI for Life Sciences starts with an AI risk assessment of current AI use and data exposure. Status, October 2026. In use in Atlas Bio's own agent operations today: a separate verification step that checks agent work, and a hold that stops a task until a person decides. Built and tested on synthetic data, not yet used with a client: the AI risk assessment tooling (AI discovery, data-flow and vendor-terms mapping, and a NIST AI RMF gap report) and a record of what data an AI used, what it produced and who reviewed it. Built and tested only against a mock, never deployed: a template for a customer-controlled AI environment in your own Azure tenant. Planned, not yet built: an AI gateway and agent permissions. Each engagement states which controls are built.
How do you confirm an AI tool keeps your data out of model training before you adopt it?
Check three things before adoption. First, the contract and DPA: the no-training commitment should sit in the agreement itself, with its exceptions, such as feedback and opt-in data sharing, spelled out. Second, the vendor's own statement on training use for business or API accounts, read as a description rather than a commitment. Third, account and admin settings: company-managed accounts only, and a decision on whether feedback and data-sharing settings should be switched off for research work.
Is a business AI plan enough on its own?
No. Business terms generally bar the provider from training on your data, with exceptions such as feedback you choose to send, but they do not stop staff or agents from sending material that should never leave your boundary, and they leave retention, sub-processors and partner use to be checked separately. You also need a policy on what may be sent, and a way to enforce it.
Which contract terms matter most for proprietary research data?
Five: training use and its exceptions, retention and deletion, rights in fine-tuned and derived models, the provider's licence and sub-processors, and what collaborators and CROs may do with your information in their own AI tools. These are issues to check with counsel, not legal advice.
Does the provider delete our data when the contract ends?
The published terms say so, on different timetables and with carve-outs. OpenAI's agreement says within 30 days, Anthropic's DPA within 30 days, Microsoft's DPA after a 90-day extraction window plus up to 90 more days, and Google's DPA within 180 days after a recovery period of up to 30 days, all with exceptions such as legal requirements.
Should research data ever go to an external model?
Some can, under the right terms and after transformation; some should stay local. The decision should follow the data's classification and the agreements that govern it, made in advance, rather than each person's judgement in the moment.
Does federated learning keep our data private?
It keeps raw data where it is, which is a real gain, but NIST notes that attacks on shared model updates and on the trained model can reveal information about the training data. Treat it as one protection among several, with its own agreement between the participants.
What about patient data?
Health and genetic information needs its own explicit rule and an agreement with the provider that covers it, under the law that applies to you. The UK regulator, the ICO, says pseudonymised data is still personal data in the hands of anyone who holds the additional information needed to re-identify it, and may be anonymous for an organisation that receives it without that information; so coded trial data is personal data for the site that holds the key, and whether it is also personal data for a sponsor or an AI provider without the key is a question for counsel. Participant consent and study agreements can also limit secondary use. Treat it as local-only or approval-required unless the provider agreement, the consent and study agreements, and de-identification that meets the applicable legal standard are all in place.
- OpenAI Help Center: How your data is used to improve model performance
- OpenAI: Services Agreement (effective 1 January 2026)
- OpenAI: Data Processing Addendum (effective 1 January 2026)
- OpenAI: Enterprise privacy
- OpenAI API docs: Data controls in the OpenAI platform
- Anthropic: Commercial Terms of Service
- Anthropic: Data Processing Addendum
- Anthropic Privacy Center: Is my data used for model training?
- Anthropic Privacy Center: How long do you store my organization's data?
- Microsoft Learn: Data, privacy, and security for Foundry Models sold by Azure
- Microsoft: Products and Services Data Protection Addendum (DPA)
- Google Cloud: Service Specific Terms
- Google Cloud: Cloud Data Processing Addendum
- Tucker Ellis LLP: The AI Vendor Contract Is Becoming More Important Than the Privacy Policy (August 2026)
- Morgan Lewis: Key Considerations When Allowing a Vendor to Train Its AI Models on Customer Data (June 2026)
- Loeb & Loeb: Navigating the Challenges of AI Contract Negotiations (May 2026)
- Bristows: Integrating generative AI into the life science ecosystem - five tips for success
- Freshfields (ACC programme materials, 20 March 2024): Hot Topics in Global Clinical Trial Agreements
- Trethowans: Non-disclosure agreements (NDAs) in the new world of AI
- ICH E6(R3) Guideline for Good Clinical Practice (final, January 2025)
- Health Research Authority bulletin: updated commercial model agreements and AI use of confidential information (7 September 2026)
- 18 U.S.C. 1839: definitions, including 'trade secret' (Cornell LII)
- eCFR: 28 CFR Part 202, access to US sensitive personal data and government-related data by countries of concern
- eCFR: 15 CFR Part 734, scope of the Export Administration Regulations
- eCFR: 15 CFR Part 774, the Commerce Control List
- ICO (UK): Pseudonymisation guidance
- Microsoft Learn: Learn about sensitivity labels
- Microsoft Learn: Microsoft Purview data security and compliance protections for generative AI apps
- Hugging Face: openai/gpt-oss-120b model card
- Heyndrickx et al., MELLODDY: Cross-pharma Federated Learning at Unprecedented Scale (J Chem Inf Model 2024, 64(7); online 2023; PMC)
- NIST: Privacy Attacks in Federated Learning (January 2024)
- NIST AI 600-1: Generative Artificial Intelligence Profile
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →