It’s one of the first questions AI founders ask about a beta: the testers are generating exactly the data the model needs — real prompts, real documents, real corrections. Can the company train on it?
Sometimes, yes. But only if the right to train was clearly granted before the data came in, by someone with authority to grant it, for data the law allows to be used that way.
SaaS AI Lawyer Andrew S. Bosin, a two-time SaaS founder, sees most training problems start the same way: a beta agreement that says nothing about training, or a vague “to improve our services” line that the founder reads as training rights and the customer reads as bug fixes.
Andrew’s guide can an AI beta use customer data? covers the privacy-law threshold — whose data it is, what the customer already signed, and the startup’s role under CCPA and GDPR. This article picks up from there: how to get training consent that actually holds up, and what happens when it doesn’t.
Quick Answer
An AI startup can usually train on beta customer data and inputs only if:
- The beta agreement expressly grants training rights — “improve the services” is not enough
- Consent is obtained before the data is collected, not added later through a terms update
- The person consenting has authority over the data — including any third-party personal data inside it
- Privacy and sector laws allow it — the startup’s role under CCPA or GDPR and any health, biometric, or children’s data rules
- The startup’s model provider terms allow it for the models involved
- The startup can honor what it promised about withdrawal and deletion once the data is in the model
If any of these is missing, the safer path is to train only on de-identified data, synthetic data, or data from testers who opted in separately.
What “Training” Actually Covers
The word does a lot of work in a beta agreement, and it should be defined. The uses founders want to make of beta data range widely in risk:
- Debugging and support — reviewing inputs and outputs to fix errors
- Evaluation — using beta inputs as test cases to measure model quality
- Customer-specific tuning — adjusting prompts, retrieval, or a fine-tuned model used only for that tester
- Shared-model training — using the data to train or fine-tune a model offered to all customers
- Feedback and ratings — thumbs up/down, corrections, and preference data used for reinforcement or evaluation
The first three rarely surprise a customer. Shared-model training is where disputes start, because the customer’s data ends up improving a product their competitors can buy.
A good beta agreement defines each category separately and grants — or withholds — rights for each one. Andrew’s guide to AI beta testing agreement key terms covers how the definitions of Customer Data, Inputs, Outputs, and Feedback decide these rights.
Rule 1: Get Consent Before the Data Comes In
The most important rule is timing.
In February 2024 the FTC warned that a company that quietly changes its terms of service or privacy policy to start using previously collected data for AI training may be engaging in an unfair or deceptive practice (FTC Technology Blog, AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive).
For a beta, that means:
- Decide on training rights before the first tester signs — not after the data proves valuable
- Don’t rely on a unilateral terms update to cover data already collected under narrower terms
- If the plan changes mid-beta, get fresh, affirmative consent, and apply new rights only to data collected afterward — unless each tester expressly agrees otherwise
Rule 2: Opt-In Beats Opt-Out for Shared-Model Training
There are three common consent structures:
- Opt-in — training is off unless the tester affirmatively turns it on or signs a training addendum
- Opt-out — training is on by default, and the tester can turn it off
- Mandatory — participation in the beta is conditioned on training rights
For individual users and small-business testers accepting clickwrap terms, an opt-out buried in the terms is the weakest position if anyone later argues the use wasn’t disclosed clearly.
For enterprise testers, the question is usually moot: procurement and security teams will insist on no shared-model training by default. Andrew’s guide to the no-training clause covers how those negotiations play out.
Mandatory training can work — for a free beta where training is the main value the startup receives — but the consent must be prominent, specific, and presented before any data is submitted.
Rule 3: The Tester Must Be Able to Grant the Rights
A business tester can license its own data. It often can’t license everyone else’s.
Beta inputs routinely contain third-party data — the tester’s customers, employees, patients, or end users. A tester that grants training rights over that data may be promising something it has no right to promise, and the startup inherits the problem.
The agreement should:
- Require the tester to represent that it has the rights and notices needed to grant the license it’s granting
- Exclude sensitive categories from training — health, biometric, financial account, and children’s data — unless there’s a specific addendum
- Address the startup’s own role — under CCPA, a service provider generally can’t use customer personal data to build or improve services for other customers except within narrow limits; under GDPR, using customer data to train a shared model can make the startup a controller for that processing, which requires its own lawful basis
On the GDPR point, the European Data Protection Board’s Opinion 28/2024 on AI models says legitimate interest can be a lawful basis for developing AI models, but only after a case-by-case balancing test — and that a model trained on personal data is not automatically “anonymous.” The opinion also notes that unlawful processing during development can affect whether the model can lawfully be used later.
Rule 4: Check the Upstream Model Terms
Most AI startups fine-tune or build on a foundation model from OpenAI, Anthropic, Google, or an open-weight provider. The startup’s right to train is limited by the provider’s terms as well as the customer’s consent.
Before promising anything to testers, confirm:
- Whether the provider permits fine-tuning on the model and plan in use
- The provider’s own data use — whether it retains or uses the startup’s API data, and for how long
- Any license restrictions on open-weight models, including acceptable-use policies and limits on commercial use
A beta agreement that tells testers “your data will never leave our systems” isn’t accurate if inputs are sent to a third-party API with its own retention period. Andrew’s guide to AI vendor contracts for startups covers the upstream terms in more detail.
Rule 5: Don’t Promise Deletion You Can’t Deliver
This is the most common drafting mistake in AI training consent.
Once data has been used to train model weights, removing its influence usually means retraining from a clean dataset. Most startups can delete raw data, logs, and embeddings on request. Few can surgically remove a single tester’s contribution from a trained model.
The consent clause should be honest about the difference:
- Withdrawal stops future use — data submitted after withdrawal won’t be used for training
- Raw data is deleted on the stated timeline
- Models already trained are not recalled — unless the startup has chosen to commit to retraining
Why this matters: the FTC has repeatedly required companies to delete not just improperly obtained data but the models and algorithms built from it — a remedy often called algorithmic disgorgement — in orders against Everalbum (2021), WW International’s Kurbo (2022), and Rite Aid (2023). A startup that trains on beta data without clear consent may be putting the model itself at risk, not just the dataset.
Andrew’s article on AI model ownership — weights, training data, and outputs explains why the provenance of training data matters for model ownership and investor diligence.
Safer Paths That Still Improve the Model
If full training consent isn’t available, there are still useful options:
- Train on feedback, not content — ratings and corrections often carry most of the signal, with less of the customer’s confidential material
- Aggregate and de-identify — using a de-identification standard the law actually recognizes, not just removing names
- Separate opt-in cohort — a smaller group of testers who expressly agree to shared-model training, often in exchange for pricing or credits
- Synthetic data — generated from patterns observed in beta usage, reviewed so it doesn’t reproduce real records
- Customer-specific tuning only — improving the tester’s own instance, which most enterprise customers will accept
Sample Clause: Training Consent
Training consent is one clause in a larger agreement, and it has to line up with the definitions, feedback, and deletion terms around it. For the full set of AI-specific clauses, see this AI beta testing agreement template with annotated sample clauses.
Model Training. Customer may elect, by written notice or through the Beta Service settings, to permit Company to use Training Data to train, fine-tune, and evaluate machine learning models made available to Company’s other customers (“Shared Model Training”). Absent such election, Company will not use Customer Data, Inputs, or Outputs for Shared Model Training. “Training Data” means Inputs, Outputs, and Feedback submitted after the date of Customer’s election, excluding any Sensitive Data. Customer represents that it has provided all notices and obtained all consents required to permit Shared Model Training on Training Data.
Sample Clause: Withdrawal and Deletion
Customer may withdraw its election at any time on thirty (30) days’ written notice. Upon withdrawal, Company will cease using newly submitted Customer data for Shared Model Training and will delete Training Data held in raw form, including associated logs and embeddings, within sixty (60) days. Customer acknowledges that models trained before the effective date of withdrawal may continue to be used and will not be retrained to remove Training Data.
These are starting points, not ready-to-sign terms. The right structure depends on the product’s data flows, the testers, and the model stack.
Founder Checklist
Before the beta opens, confirm:
- Training is defined, and each use (debugging, evaluation, customer tuning, shared training) is addressed separately
- Consent is obtained before data is submitted, not added through a later terms update
- The consent structure — opt-in, opt-out, or mandatory — fits the tester type
- Testers represent that they can grant rights over third-party data
- Sensitive data categories are excluded or covered by an addendum
- The startup’s CCPA and GDPR role supports the planned use
- Model provider terms permit the planned fine-tuning
- Withdrawal and deletion terms match what the system can actually do
- The privacy policy and AI terms of service say the same thing as the beta agreement
- Confirm the rest of your beta agreement doesn’t contradict your training choice — the AI beta agreement template guide covers the ten gaps free templates leave.
Frequently Asked Questions
Does “to improve our services” give an AI startup the right to train its models?
Usually not safely. It’s ambiguous, and a customer can reasonably read it as covering bug fixes and support, not training a model sold to other customers. Training rights should be stated expressly.
Can we add training rights after the beta has started?
Only with fresh, affirmative consent, and ideally only for data collected after that consent. The FTC has warned that quietly changing terms to use previously collected data for AI training may be unfair or deceptive.
Is it enough to de-identify beta data before training?
It can be, if the de-identification meets the legal standard that applies — CCPA, HIPAA, or GDPR each set their own — and the data is actually protected against re-identification. Removing names alone usually isn’t enough.
What happens if a tester withdraws consent after we’ve trained on their data?
That depends on what the agreement promised. A well-drafted clause stops future use and deletes raw data but says plainly that existing models won’t be retrained. Without that language, a tester may argue the model itself must change.
Do enterprise beta testers ever agree to shared-model training?
Occasionally, usually in exchange for meaningful pricing concessions and with strict limits on data categories. Most enterprise testers accept only customer-specific tuning.
Can our model provider’s terms limit what we can do with beta data?
Yes. The provider’s terms govern whether fine-tuning is allowed and how the provider handles API data. The startup can’t grant testers assurances that its upstream terms don’t support.
Get Training Rights Right Before the Beta Starts
Andrew S. Bosin is a New Jersey technology attorney and two-time SaaS founder who represents AI and SaaS startups nationwide on a flat-fee basis. He drafts beta, design partner, and pilot agreements with training rights built around the product’s actual data flows and model stack — so the startup secures the data rights it needs, and the promises it makes to testers are ones it can keep.
He personally drafts every agreement, gives every client his personal cell phone number, and responds quickly.
Andrew S. Bosin LLC, 36 Highland Road, Glen Rock, New Jersey 07452 📞 201-446-9643 📧 andrewbosin@gmail.com 🌐 www.njbusiness-attorney.com
#AISaaSLawyer #AIStartupLawyer #BestAISaaSLawyer #AITraining #BetaAgreement #DataRights #AICompliance #SaaSContracts