Terms of Service for AI Products: The Clauses That Actually Matter

Terms of service for an AI product look like ordinary SaaS terms until you reach the hard questions. Who owns what the model produces? What happens when it is wrong? Can you learn from what customers type? What do you owe them if your model provider changes its rules overnight? Generic templates answer none of these, and enterprise buyers now ask all of them in procurement.

This guide covers the clauses that matter most for AI products and how to draft each one so it holds up.

1. Definitions that match the product

Start with precise definitions of Inputs (what users submit), Outputs (what the service returns), Customer Data, the Service and Third-Party Models. Most later clauses depend on these. A clause about “content” that does not distinguish inputs from outputs will be read against you when a dispute arrives.

2. Ownership of inputs and outputs

The market position has settled into a pattern:

• Inputs remain the customer’s. The provider receives a licence limited to operating, securing and (if agreed) improving the service.
• Outputs are assigned to or left with the customer, as between the parties, subject to compliance with the terms.
• The model, software and service remain the provider’s.

Two cautions. First, do not promise that outputs are protected by copyright. Whether copyright subsists in AI-generated material depends on national law and on human creative contribution; the U.S. Copyright Office, for example, has said that purely AI-generated material without sufficient human authorship is not protected. Second, do not promise outputs are unique. Similar prompts can produce similar results for different users, and the terms should say so.

3. Accuracy, reliance and professional advice

This is the clause most likely to be tested. It should state, clearly and prominently, that outputs are generated automatically, may be inaccurate, incomplete or out of date, must be reviewed by the user before reliance, and do not constitute legal, medical, financial or other professional advice. Where the product serves regulated sectors, require users to apply their own professional judgement.

A disclaimer is only as strong as its consistency with everything else. If your website promises “accurate legal answers every time”, a disclaimer in the terms will carry little weight with a court or regulator, and consumer protection authorities, including the US Federal Trade Commission, have taken action over unsupported AI claims. Align marketing, interface and terms.

4. Acceptable use

AI products need an acceptable use policy that goes beyond spam and malware. Address:

• illegal content and infringement of others’ rights;
• attempts to extract model weights, training data or system prompts;
• jailbreaking or circumventing safety measures;
• generating deceptive content, impersonation and undisclosed deepfakes;
• high-risk uses you do not support (for example, using the service for decisions about employment, credit or medical treatment without appropriate safeguards);
• automated access beyond rate limits;
• using outputs to build a competing model, where you choose to prohibit it.

Reserve the right to suspend accounts that breach the policy, and say how notice works.

5. Training on customer data

State the position in one unmistakable sentence. Either:

• “We do not use Customer Data, Inputs or Outputs to train or improve our models or third-party models”, or
• “We may use [de-identified] Inputs and Outputs to improve the Service, and you can opt out at [location].”

Enterprise customers often require the first. Consumer products often use the second. What fails is ambiguity, such as “we may use data to improve our services” without saying whether that means model training. The privacy policy must say the same thing, with a lawful basis under data protection law. See how to write a privacy policy for a SaaS or AI app.

6. Third-party models and the flow-down problem

If your product runs on a third-party model, your own obligations depend on that provider’s terms. Check them for usage restrictions, data retention, training, regional processing, uptime and the right to change or withdraw models. Then make sure your customer terms:

• pass through the acceptable use restrictions the provider imposes on you;
• do not promise more than the provider gives you (on availability, data location or retention);
• allow you to switch models or providers;
• name sub-processors where personal data is involved.

The free AI Vendor Terms Red-Flag Scanner highlights training, retention, confidentiality, output ownership and liability terms in a provider’s contract.

7. Confidentiality and security

B2B customers will expect inputs to be treated as confidential, a description of security measures, breach notification and, where personal data is processed, a data processing agreement. Keep the security description accurate and current. Never state certifications or standards you have not actually achieved.

8. Intellectual property indemnities

Large AI providers increasingly offer to indemnify business customers against third-party IP claims arising from outputs, subject to conditions such as using the service as intended and not disabling filters. Smaller providers may not be able to offer this. If you do, define the scope carefully and tie it to your own upstream protection.

9. Liability

Cap liability at a sensible figure, commonly fees paid over a defined period, exclude indirect and consequential loss, and carve out what the law will not let you exclude. In many jurisdictions you cannot exclude liability for death or personal injury caused by negligence or for fraud, and consumer law controls unfair terms (in the EU under the Unfair Contract Terms Directive and in the UK under Part 2 of the Consumer Rights Act 2015). A balanced cap drafted for enforceability protects you better than an extreme one a court may strike down.

10. The EU AI Act layer

If your product reaches users in the EU, the AI Act affects your terms and interface. From 2 August 2026, people must be told when they interact with an AI system unless it is obvious, and generative AI output must be marked as machine-generated. If your product could be used for a high-risk purpose, decide whether you support that use. If you do not, prohibit it in the acceptable use policy; if you do, prepare for the high-risk obligations now due from 2 December 2027. The EU AI Act timeline sets out every date.

11. Changes, suspension and termination

AI services change quickly. Reserve the right to update models and features, but commit to notice for material adverse changes. Set out data export and deletion on termination, including what happens to inputs and outputs held in logs.

12. Governing law and disputes

Choose a governing law and forum you can live with, consider arbitration for cross-border B2B customers, and remember that consumers may keep the protection of their home courts and mandatory laws regardless of your choice.

A checklist before you publish

• Inputs, outputs and customer data defined and distinguished.
• Output ownership allocated, with no promise of copyright or uniqueness.
• Accuracy disclaimer clear, prominent and consistent with marketing.
• Acceptable use covers AI-specific misuse.
• Training position stated in one sentence and matched in the privacy policy.
• Upstream model terms flowed down; switching rights reserved.
• Liability capped and carve-outs compliant with mandatory law.
• AI Act transparency reflected in terms and interface.
• DPA available for business customers.

The bottom line

Good AI terms are honest about what the product can and cannot do, precise about data and ownership, and consistent with the contracts behind them. That combination protects the business, closes enterprise deals faster and builds the kind of trust that keeps customers when competitors stumble.