Before adding an AI bot, decide what text it may receive, who can use it, and what happens when its answer is wrong. “AI” is not a permission model or a privacy promise.
Complete this privacy worksheet first
Keep the answers with a link to the developer’s current policy or documentation and the date you read it. Mark missing answers unknown; do not assume that “private channel” means “no external processing.”
- Job and input: what task needs AI, and what exact text, attachments, or conversation history does the feature receive?
- Processing: which developer or model provider receives that input? Does the feature send data outside Discord?
- Retention and training: what does the policy say about storing prompts and outputs, using them for model improvement, and deleting them?
- Access: which channels can the app access, which members can trigger it, and where will replies appear?
- Exit and human help: who can stop the workflow, request data deletion, and answer a member without using AI?
Discord’s Data Access guidance distinguishes baseline server data from app-specific access and direct interactions. Check the bot’s profile on desktop or browser, then read its developer’s policy. For the difference between channel permissions and Message Content Intent, use our message-access explainer. Neither an invite prompt nor the intent setting explains a provider’s retention policy.
Test one explicit command with synthetic text
Start with a low-risk draft or question in a restricted test channel—not passive reading across the server. Grant only the access the feature needs. Discord command permissions let you review access for roles, members, channels, and individual commands on desktop.
- Useful output: ask a question whose answer you already know. Check the answer rather than accepting confident wording.
- Wrong or missing information: use a made-up example with a deliberate ambiguity. Check that a moderator can reject or correct the response.
- Untrusted instructions: include harmless quoted text asking the bot to ignore its task. Check whether it stays within the intended job. One successful test is not a security guarantee.
- Audience: check which ordinary members can run the command and who sees the result. Owner access is not a substitute for this check.
- Fallback: confirm that a member can reach a human if the bot fails or gives an unsuitable answer.
Keep AI output advisory. Have a human approve public announcements, sensitive summaries, and moderation decisions. Never test with passwords, recovery codes, private customer data, or confidential moderation evidence.
Write a member notice that matches verified facts
Do not publish a promise that inputs are deleted, never used for training, or never leave Discord unless the developer’s current terms support it. If a data practice is unknown, resolve that uncertainty before using the app with member content.
Record results without copying private prompts
Record the test date, app version if available, settings, synthetic test result, and reviewing moderator. Do not turn a staff log into a second collection of member messages or full sensitive prompts.
Stop the workflow if output appears in the wrong place, access expands unexpectedly, or staff cannot correct harmful answers. Removing an app from Discord does not by itself establish that the developer deleted previously received data; use the policy’s deletion route.
Official Discord documentation reviewed October 19, 2026. The worksheet and tests are suggested checks, not results of an independent AI-product test.