Safety and acceptable use
The rules your app and its users follow, when a conversation ends for a policy violation, and what happens after repeated violations.
Every conversation your API keys start follows the Terms of Service and the Acceptable Use Policy. These rules apply to you and to everyone who talks to the AI persona through your app.
What you are responsible for
- Your prompt. You are responsible for the
personality.system_promptyou send. Do not write one that asks the persona to break the Acceptable Use Policy, for example to give medical, legal or financial advice, or to speak as a real person without their consent. We review each new prompt, and your tools, the first time you mint a session token with them. That first mint takes a few seconds longer; later ones with the same prompt don't wait.- A prompt that breaks the policy is refused with HTTP 400 and a reason. Examples are posing as a company or person it is not, asking for codes, card details or passwords, pushing for gift cards or wire transfers, inventing threats, or claiming to be human.
- A prompt that asks the persona for something it doesn't do is refused with HTTP 400 and a reason that says what to change. The persona speaks English only, so a prompt or tools with non-English characters, or that ask it to speak another language, are refused. So is a prompt that asks it to name the AI model or company it runs on, collect sensitive personal data (codes, passwords, ID numbers, card or bank details, income, health records), give medical or mental health guidance or legal, tax or financial advice, take a political side or persuade people how to vote, promote gambling or crypto, or give religious or spiritual guidance. These refusals never count toward a pause.
- Past a daily limit on new prompts, a new prompt's token is refused with HTTP 400 saying it is waiting for our review, until our review reaches it, within a day. Then it works, or is refused.
- A prompt our review can't finish in time at mint works until the review is done, usually within 10 minutes. When the review then refuses it, we email the person who made your organization (at most once a day), and your organization's overview on the platform lists every prompt that is refused now, with the reason.
- Your users. You choose who can start a conversation in your app. Your server mints each session token, so only mint one for a person you allow to use your app.
- AI labels. Do not hide, crop or remove anything that shows the persona is an AI.
Not a crisis service
The persona is not a medical, mental health or crisis service. Do not offer your app to people in an emergency. Point them to emergency services (911 in the United States), the 988 Suicide and Crisis Lifeline (call or text 988), or the crisis service where they live. Show a local crisis line in your app itself: a tool result that gives one can end the conversation.
When a conversation ends for a policy violation
A conversation may end because of a policy violation. You see it as:
| Where | What you get |
|---|---|
| Events route | ended with reason: "policy_ |
nuance_ topic | session. with reason: "content" and a message |
The message is the same for every policy violation, whichever check found it: "This conversation ended because something in it broke our Acceptable Use Policy." Tell the person the conversation ended, and do not start a new one for them right away. Show them the message, or these words when the event has none.
When a conversation ends because of an age check
The persona's camera may show only people who look under 18, with no adult beside them. The conversation may then end. A person under 18 with an adult in view can keep talking. The age check is a backstop, not an age gate: keep your own age requirement. You see the end as:
| Where | What you get |
|---|---|
| Events route | ended with reason: "age_ |
nuance_ topic | session. with reason: "age_ and a message |
The message is: "This conversation ended because it looks like someone under 18 is using it without an adult." An age check is not a policy violation: it does not count toward the pauses below. Tell the person the conversation ended, and do not start a new one for them right away. The check itself keeps none of the images it reads.
After repeated violations
When several conversations end for policy_violation within a few hours, new conversations are refused for a while. retry_after_seconds says how long. What the count covers depends on the session token:
| Session token | What the count covers |
|---|---|
With end_ | That person's conversations in your organization. |
Without end_ | Every conversation your organization started without an end_, through any of its API keys. Conversations that name their end user keep working. |
| Where | What you get |
|---|---|
| HTTP | 403 with code: "policy_, retry_, and a Retry-After header |
The HTTP detail opens with words to show the person, which say how long to wait, then says which conversations the pause covers.
When we pause your organization
We pause new conversations for your whole organization, and look at it ourselves, when either happens:
- Prompts that break the policy. Several of your prompts, or sets of tools, are refused by our review in a short time. A refusal or two leaves room to find and fix what tripped it, and prompts refused only for something the persona doesn't do don't count.
- Many of your users break the rules. A large share of the people who used your app recently are paused for repeated violations at once.
We email the person who made the organization when it is paused. The platform shows the pause on your organization's pages. While it is paused, your API keys keep working, but minting a session token and starting a conversation answer 403 with code: "organization_paused". Write to support@nuancelabs.ai if you think it is a mistake. After we look, we lift the pause or suspend the organization.
Conversations that are already running are not affected.
Send end_user_id with every session token you mint: your own id for the person, such as a hash of your user id (letters, digits and _ . : -, up to 128 characters; never an email address or a name). Without it, one person who keeps breaking the rules can pause new conversations for all your users, and the token response's warnings says so. Either way, stop starting conversations for a person whose conversations end for policy_violation.
Reporting
To report content or conduct that you believe breaks the Acceptable Use Policy, contact us as the Terms of Service describe.