When an AI result is not right
What to do when an AI result comes back wrong, stuck, or not worth the credits: fix the source and the prompt, check files and permissions, then regenerate once.
Review AI output before using it
If an AI result is wrong, vague, incomplete, off-tone, based on the wrong source, or stuck in progress, treat it as a recoverable draft problem. Start with four checks: give better context, ask a more specific question, choose the right output type, or stop regenerating and edit manually.
AI can be fluent and still wrong. Review generated output before sharing it with teammates, clients, students, or customers. For preview guidance and webpage-style sharing, read Review AI output and share web pages.
Before you start
| Item | Details |
|---|---|
| Applies to | AI Studio, Object Chat, selected item AI actions, voice input, and generated canvas output. |
| Available for | Web app and desktop app. AI actions appear only in canvas, object, and workspace states that expose the relevant entrypoint. |
| Who can recover | People who can open the relevant canvas or object can review existing output. New AI actions require send or edit permission on that canvas or object. AI credit purchase and Auto-reload recovery require a workspace admin. |
| What can block recovery | Read-only canvas access, unsupported object type, incomplete or failed attachments, insufficient credits, a deleted processing placeholder, or workspace deactivation. |
Identify what changed from your intent
| Problem | Likely cause | Best recovery |
|---|---|---|
| Result is too generic | Prompt did not include audience, source, format, or constraints. | Rewrite the prompt with specific context. |
| Result uses wrong facts | Wrong source selected, missing file, failed upload, or AI inferred unsupported details. | Attach or select the correct source and ask it to use only that source. |
| Result is the wrong format | Output type or prompt did not specify structure. | Ask for a document, table, checklist, image, audio, webpage, or exact sections. |
| Result is too long or too short | Prompt did not set length. | Ask for a word count, number of bullets, or one-page format. |
| Tone is wrong | Prompt did not name audience or tone. | Specify client-facing, internal, executive, classroom, technical, or casual tone. |
| Item answer references wrong object | Wrong object selected or object context unavailable. | Select the correct item and use Object Chat again. |
| File context missing | File still uploading, failed, unsupported, too large, or not included. | Wait, remove failed upload, convert file, or reattach. |
| Placeholder stuck | Generation is still running, the server task failed, the placeholder was deleted, or the connection is stale. | Refresh, check the placeholder, and avoid duplicate submissions. |
| Credit block appears | The workspace did not have enough credits, so the action was blocked before it started. | Admins can open Billing to buy credits or manage Auto-reload. Non-admin members, guests, and external collaborators need a workspace admin. |
Rewrite the prompt
A strong prompt includes goal, audience, source, format, constraints, and what not to change.
Weak prompt: Summarize this.
Better prompt: Summarize the selected customer interview notes for a product manager. Use five bullets: pain points, requested features, objections, exact quotes, and recommended follow-up. Do not add facts that are not in the notes.
Weak prompt: Make a table.
Better prompt: Create a comparison table from the attached files. Columns: feature, customer impact, engineering effort, risk, owner, and next step. If a value is not in the source, write "Not stated".
Weak prompt: Rewrite this.
Better prompt: Rewrite the selected note for a client-facing update. Keep the date and numbers unchanged. Remove internal team names. Use a calm professional tone.
Clean up the source context
AI Studio and Object Chat work best when the source is explicit. Select the objects you want AI to use. Attach the right files. Remove irrelevant files. Paste only safe, relevant URLs. Wait for uploads to finish.
If the result ignored a file, check whether the file actually uploaded. A file that is still processing or failed is not included in the AI request.
If the source material is huge, ask a narrower question. For example, instead of asking for a full strategy from five documents, ask for the risks in one document or a comparison across two selected sections.
Choose the right AI feature
| Need | Best tool |
|---|---|
| Generate a new canvas object | AI Studio. |
| Ask about one selected item | Object Chat or item-level AI action. |
| Ask about broader canvas context | Select the relevant canvas objects and use AI Studio, or use Object Chat when the question is about one selected item. |
| Review or revise generated output | Object Chat, AI Studio follow-up, or manual editing. |
| Get a teammate decision | Comments, Team chat, or an audio/video call. |
| Keep feedback attached to one object | Comments, not AI chat. |
Using the wrong tool can make the result feel wrong even when AI works correctly.
Regenerate or edit by hand
Regenerate when the model misunderstood the task, used the wrong source, chose the wrong format, or produced an unusable draft.
Edit manually when the result is mostly correct and only needs judgment, tone, wording, or small factual corrections. Manual editing is often faster and avoids spending more credits on repeated attempts.
Ask Object Chat for a targeted critique when you are unsure what to fix: Review this generated draft and list the top five issues before I share it.
When generation stalls or fails
If a placeholder is pending, wait first. File-heavy, image, audio, and webpage work takes longer than plain text.
If the placeholder stays stuck, refresh the canvas. If it still shows as pending, check whether credits ran out, the attached file failed, or the network dropped. Avoid submitting the same request repeatedly until you know whether the first one completed.
Deleting a processing placeholder cancels the server task from the canvas. Submit again only after deciding that another credit spend is acceptable.
If the output failed, simplify the request. Remove questionable attachments, use a smaller source file, choose a simpler output type, or ask for text first before asking for a complex visual or webpage-style result.
Check attached files
If a file fails, try these steps:
| Problem | Recovery |
|---|---|
| Upload never finishes | Check network, refresh, and upload again. |
| File is unsupported | Convert to a supported format such as PDF, image, text, Markdown, CSV, JSON, SRT, DOCX, or Excel. |
| File is too large | Split it, compress it, or ask about a smaller excerpt. |
| File is password-protected | Remove protection before uploading if policy allows. |
| Spreadsheet result is wrong | Specify sheet, columns, units, and desired table output. |
| URL fails | Check whether the URL is public, safe, accessible, and not too large for processing. |
Do not assume AI used a source file just because you intended to attach it. Confirm upload completion.
Check credit issues
If credits are insufficient, AI actions are blocked before they start. Workspace admins can buy credits or manage Auto-reload in Billing. Non-admin members, guests, and external collaborators need a workspace admin to add credits.
If you spent credits on an unsatisfactory result, review the source and prompt before trying again. Deleting a generated object does not return credits because the AI work already ran.
For details, read Understand AI credits.
Review before sharing
Before sharing AI output externally, verify:
| Check | Why |
|---|---|
| Facts | AI can invent or misread details. |
| Dates and numbers | These are easy to get subtly wrong. |
| Names | People, companies, files, and projects can be confused. |
| Permissions | Shared canvases show generated output to everyone with access. |
| Source completeness | Failed or missing attachments can lead to incomplete answers. |
| Tone | Client-facing, classroom, executive, and internal tones differ. |
| Sensitive content | Remove private or unnecessary data before sharing. |
What people often miss
Improve the prompt instead of regenerating the same vague request several times.
AI can only verify facts it can access, so attach or select the source of truth.
Do not share a fluent answer without checking the source.
Approval belongs to a person, not to AI chat. Mention the teammate or use Team chat.
Do not ignore credit warnings. Repeated attempts can spend more workspace credits.
What to include when contacting support
Contact support when AI Studio is available but consistently fails, a completed purchase does not update credits, a generation appears to spend credits without creating any visible placeholder or result, or a pending result cannot be recovered after refresh.
Include the workspace name, canvas link, output type, approximate time, prompt summary, attachment types, whether the placeholder appeared, whether credits were blocked or spent, and screenshots of the error. Do not send passwords, verification codes, private credentials, or unnecessary sensitive source material.
Related guides
- Use AI Studio in a canvas
- Create content with AI Studio
- Ask about a canvas item
- Review AI output and share web pages
- Understand AI credits
- Contact support with the right details