Troubleshooting
Almost every frustrating moment with an AI assistant traces back to a handful of causes — and each has a quick fix. When something feels off, find the symptom below, check the likely cause, and apply the fix. None of this needs technical skill, just a small change to how you ask.
Quick Diagnosis
| What you are seeing | Most likely cause | Go to |
|---|---|---|
| Confident answer that is simply wrong | Missing facts, or the AI guessed | Wrong or Made-Up Answers |
| Output reads like a generic article | No context or specifics | Output Too Generic |
| Tool refuses your document | File type, size, or privacy block | The Tool Will Not Process Your File |
| A wall of text, or two thin lines | No length or format guidance | Responses Too Long or Too Short |
| Replies have drifted off-track | Conversation overloaded or muddled | Recovering a Derailed Conversation |
Wrong or Made-Up Answers
A "hallucination" is when AI states something false as fact — an invented statistic, a fake source, a wrong date. It sounds confident, so it is easy to miss.
| Problem | Likely cause | Fix |
|---|---|---|
| Invented figures or quotes | It filled a gap rather than admit a blank | Paste the source numbers; say "use only the figures I provided" |
| Fake citations or links | It predicted what a source might say | Ask it to "list only sources you are certain exist" |
| Out-of-date facts | Knowledge has a cutoff date | Provide current data, or use a tool with web search |
| Subtle errors in a variance analysis | It estimated, not calculated | Ask it to "show the working so I can check it" |
Verification prompt — paste the draft back and ask:
List every factual claim, figure, and source above, and mark
each one certain, unsure, or guessing. Flag what I should
verify before I send this on.
Golden rule: treat AI output like a draft from a bright but new colleague. Verify any number, name, or claim before it leaves your desk. Governance has a pre-send checklist.
Output Too Generic
If the result could have been written for anyone at any company, the AI lacked enough about your case.
| Problem | Likely cause | Fix |
|---|---|---|
| Status report reads like a template | No project specifics | Paste the actual milestones, owners, and dates |
| Advice states the obvious | No goal or constraints | Add what you want to achieve and what you cannot do |
| Email sounds robotic | No audience or tone defined | Name the recipient and the relationship |
| Ignores your industry | No domain context | Say "I work in [field]; assume that knowledge" |
Add these four ingredients and the blandness goes away:
Role: Act as my finance business partner.
Context: The Q2 budget review for the sales team.
Specifics: Revenue 8% under plan; headcount frozen.
Output: A 4-bullet summary for our director.
The more real detail you paste, the more tailored the result. Pasting last month's report and saying "match this style" beats describing it.
The Tool Will Not Process Your File
When an upload is rejected or misread, it is almost always a format, size, or content issue, not your fault.
| Problem | Likely cause | Fix |
|---|---|---|
| "Unsupported file type" | The tool reads some formats only | Save as PDF, plain text, or DOCX and re-upload |
| Scanned PDF comes back blank | It is an image, not real text | Use a searchable PDF, or paste the text directly |
| Spreadsheet read incorrectly | Merged cells or complex layout confused it | Export the tab as a CSV, or paste rows as a simple table |
| File too large to upload | Exceeds the size limit | Split it, or paste only the relevant section |
| No upload button at all | The tool does not support files | Copy and paste the content into the message |
| "I cannot access that link" | It cannot open shared drives | Download the file and upload it, or paste it in |
When in doubt, copy and paste beats uploading. And before you upload, confirm it is allowed — client lists, financials, or personal data may be off-limits under your company's rules. Check Governance first.
Responses Too Long or Too Short
The AI cannot guess how much you want. Without a length and shape, you get its default — a wordy essay when you wanted three bullets.
| Problem | Likely cause | Fix |
|---|---|---|
| Endless preamble | No format requested | Add "3 bullets, no introduction" |
| Reply too short to use | Request was thin | Ask it to "expand each point with an example" |
| Wrong level of detail | Audience not set | Specify "for a board" or "for a new hire" |
| Drops important parts | Squeezed into a length limit | Ask for it "in two parts" or raise the word count |
Controls you can drop into any request:
- Keep it under 150 words.
- Use a short table, not paragraphs.
- One line per meeting action, owner in bold.
Pin down length, structure, and audience and you will rarely fight this again. See the Prompting Rules and the Cheat Sheet for templates.
Recovering a Derailed Conversation
A long chat can drift: the AI contradicts itself, forgets your instructions, or fixates on a wrong assumption. This is normal — it has too much in it.
| Problem | Likely cause | Fix |
|---|---|---|
| Repeats a mistake you corrected | The wrong version is still in the thread | Start a fresh chat with only the good material |
| Forgot your original instructions | Too much has scrolled past | Restate the key rules in one clear message |
| Lost the thread of a long task | The conversation got overloaded | Ask for a summary, then continue from it |
| Fixated on a wrong assumption | It locked onto something early | Say "stop — reset on this point" and correct it |
| Tone or quality slowly drifted | Accumulated clutter | Begin again with a clean, structured prompt |
The reset-and-summarise move — ask for this, then paste the summary into a new chat:
Summarise everything we have agreed so far: the goal, the
decisions, and the current draft. Make it complete and
self-contained so I can continue in a fresh chat.
For a smaller wobble, redirect in place: "Stop. Ignore the last few messages. The correct brief is [restate it]. Work only from this."
Starting over is a skill, not a failure. Most problems are solved not by a better tool but by a clearer request — change one thing, try again.