exactly as provided. Then 6-8
somewhere.
We need to include What this is NOT
honesty section (no guarantee / 100% / never miss). So that section should say it's not a guarantee, not 100%, not never miss etc.
We need References at end: References
Understanding your AI feature scope
- Explain how to define feature boundaries, input types, output types, user expectations.
- Include list.
2. Drafting prohibited uses specific to your product
- Explain how to list prohibited uses, examples, referencing OWASP LLM01 prompt injection, EU digital strategy.
- Include list.
3. Mapping enforcement actions to violations
- Provide table: Violation type -> Action (throttle, suspend, terminate) -> Who decides.
- This is the ONE table.
4. Input and output responsibilities
- Explain user rights, checking accuracy, etc.
5. Transparency and disclosure requirements
- Explain when to disclose AI use, referencing NIST AI RMF.
6. Testing your acceptable-use section with real users
- Explain how to get feedback, iterate.
7. What this is NOT
- Explain it's not a guarantee, not legal advice, not 100% coverage.
8. Understanding your AI feature scope
I start by writing a short description of what the feature does, who the typical user is, and what data flows in and out. This description becomes the boundary for the acceptable‑use rules.
- List the exact inputs the feature accepts – text prompts, images, audio files, or URLs.
- Note the format limits – maximum length, file size, or number of items per request.
- Describe the expected outputs – generated text, edited image, code snippet, or summary.
- Identify any downstream effects – does the output get stored, shared, or used to train another model?
- State the user’s goal – are they seeking creative help, factual answers, or automation of a repetitive task?
Drafting prohibited uses specific to your product
I look at the feature map from the previous step and ask what a motivated user could try to do that would break the law, violate someone else’s rights, or undermine the safety controls we built.
- Generation of illegal content – such as hate symbols, extremist propaganda, or instructions for making harmful devices.
- Impersonation – creating text, audio, or video that pretends to be a real person without their consent.
- Unauthorised surveillance – using the feature to analyse private communications or to build a profile of someone who has not agreed.
- Spam or flood – sending massive amounts of generated messages to disrupt a service or to harass.
- Circumventing safety controls – trying to bypass filters with misspelled words, code words, or the technique described in the OWASP LLM01 prompt injection guide: https://genai.owasp.org/llmrisk/llm01-prompt-injection
For a chat‑based assistant I also add a rule against providing medical, legal, or financial advice unless a qualified professional is named in the output. For an image generator I block the creation of deep‑fakes that could be used for non‑consensual pornography.
The EU’s digital‑age strategy reminds me that regulators expect clear limits on AI that can affect public discourse: https://commission.europa.eu/strategy-and-policy/priorities-2019-2024/europe-fit-digital-age_en
Mapping enforcement actions to violations
Once the prohibited uses are written I decide what the system will do when a violation is detected. I match the severity of the act to a response that is technically feasible and easy to explain.
| Violation type |
Enforcement action |
Who decides |
| Generation of illegal content |
Immediate suspend of the account |
Automated system with human review |
| Impersonation without consent |
Throttle the user’s API calls for 24 hours |
Automated system |
| Unauthorised surveillance of private data |
Terminate the account and preserve logs |
Legal‑compliance officer |
| Spam or flood of generated messages |
Throttle to one request per minute |
Automated system |
| Attempt to bypass safety filters |
Require a captcha and log the attempt |
Automated system |
Input and output responsibilities
I split the duties into two clear groups: what the user must guarantee before sending data, and what they must do after receiving a result.
- Input rights – the user confirms they own the uploaded file or have a licence that allows the AI to process it. For text prompts they warrant that the content does not contain private data about another person.
- Input legality – the user promises the upload does not violate copyright, trademark, or privacy law.
- Output check – before relying on the generated text, image, or code the user reviews it for mistakes, bias, or unintended disclosure.
- Output use – the user agrees not to present the output as their own original work unless they have added substantial transformation.
- Output sharing – if the user shares the result with a third party they must pass along the same disclaimer about AI assistance.
I write these points in the terms using short sentences and bold the key verbs (must, warrants, agrees) so they stand out when someone reads the clause after a dispute.
Transparency and disclosure requirements
I decide where the user would not expect AI to be involved and place a clear notice there. The notice stays visible until the user dismisses it, and it reappears if the feature is used again.
- Chat interface – a small badge next to the input box that reads “Powered by AI”.
- Image editor – a tooltip that appears when the user clicks the “Generate” button, stating “Result created with generative AI”.
- API endpoint – the response header includes a field X‑AI‑Generated: true.
- Email newsletter – a footer line that says “This message contains AI‑generated suggestions”.
I also add a short explanation of what the AI does and does not do, so the user knows the limits. For example, “The model suggests edits but does not guarantee factual accuracy.”
The NIST AI risk‑management framework suggests documenting these disclosures in a way that auditors can review: https://www.nist.gov/itl/ai-risk-management-framework
Testing your acceptable-use section with real users
I treat the clause like any other user‑facing text and run a short study to see if people can find it and understand the consequences.
- Recruit five to seven participants who match the target audience.
- Give them a scenario where they try to use the feature for a prohibited purpose – for example, ask them to request a fake ID image.
- Observe where they look for the rules – do they scroll to the bottom of the terms, search the help centre, or expect a pop‑up?
- Ask them to paraphrase what would happen if they broke the rule.
- Note any confusion about the words “throttle”, “suspend”, or “terminate”.
After the session I revise the wording: replace jargon with plain verbs, add a short example, and move the enforcement table closer to the top if users missed it.
I repeat the test after each major change until at least four out of five participants can state the correct action without looking at the document.
Frequently asked questions
Is the generated policy legally binding?
It is a draft document you publish. Whether it is adequate, and how it is enforced, depends on your jurisdiction and on review by a qualified professional.
Can we just copy a competitor's acceptable-use section?
You can end up with clauses that describe a product you do not have. Draft against your own features.
Do we need to say the feature uses AI?
Where users would not reasonably expect it, transparency is expected by regulators and by app store review guidance.
What enforcement should we promise?
Only what you will actually do. Promising termination you never carry out is worse than promising a throttle.
Does this cover app store requirements?
PolicyForge drafts against common app store disclosure themes. Store review can still require changes.
How often should the terms be updated?
Whenever the feature changes materially, and whenever your enforcement practice changes.
What is policy-as-code?
Compliance rules expressed as code, reviewed and tested like software.
Which frameworks?
GDPR, EU AI Act, and custom internal controls out of the box.
How are policies tested?
Each rule ships with test cases; CI runs them on change.
Is it a DLP?
No - it governs policy logic, not data loss prevention.
Can non-engineers edit?
Pro includes a template editor; changes still go through review.
How is versioning handled?
Git-backed; each ruleset has a version for audit.
References
Try the product: PolicyForge
2026-09-20 · primary sources only · no fabricated traffic metrics · decision-support only