Tech & AI, Explained

ai-tools-errors

10 min read

AI Tools — Common Errors and How to Fix Them

Status: 🟩 COMPLETE 🟦 LIVING Section: common-errors Tags: errors, troubleshooting, ai, fixes

How to read this

The errors and surprises you’ll genuinely hit using AI tools — and what to do about each. Most have simple fixes once you know what’s happening.

API errors

401 Unauthorized

What it means: Your API key is invalid, missing, or revoked.

Fix:

  1. Check the key is correctly copied (no extra spaces, complete string)
  2. Verify the key is current (not revoked, not expired)
  3. Check you’re using the right key for the right service (OpenAI key ≠ Anthropic key)
  4. Try generating a new key

429 Too Many Requests (rate limit)

What it means: You’re sending requests faster than your tier allows.

Fix:

  • Wait a few seconds and retry
  • Implement exponential backoff (wait longer between each retry)
  • Reduce parallel requests
  • Upgrade to a higher tier if production
  • Check your tier’s rate limits in provider dashboard
import time
def with_retry(func, max_retries=5):
    for attempt in range(max_retries):
        try:
            return func()
        except RateLimitError:
            time.sleep(2 ** attempt)

402 Payment Required or Insufficient quota

What it means: Out of prepaid credits or hit your spending limit.

Fix:

  1. Check billing in provider dashboard
  2. Top up credits (OpenAI, Anthropic use prepaid)
  3. Increase monthly limit if hit your cap
  4. For OpenAI specifically: credit expires 3 months after addition

400 Bad Request with vague error

What it means: Your request is malformed in some way.

Common causes:

  • Missing required field (max_tokens for Anthropic, etc.)
  • Invalid model name (model deprecated or renamed)
  • Token count exceeding model’s context window
  • Invalid temperature or other parameter values
  • Image format unsupported (for multimodal)

Fix:

  • Read the error message carefully
  • Compare your request to current API documentation
  • Check model name is current and available to you
  • Verify all required fields are present

500 Internal Server Error / 503 Service Unavailable

What it means: Provider’s service is having issues.

Fix:

  • Wait and retry (with backoff)
  • Check provider status page (status.openai.com, etc.)
  • If persistent: report to provider
  • Implement proper error handling in production code

Connection reset or Timeout

What it means: Network issue or request taking too long.

Fix:

  • Increase your request timeout
  • Check your internet connection
  • For long requests (reasoning models): expect 30s-5min
  • For streaming responses: use streaming API

ChatGPT specific errors

”There was an error generating a response. Please try again.”

Common causes:

  • Service issue
  • Conversation too long (hit context limit)
  • Content moderation triggered

Fix:

  • Refresh and try again
  • Start a new conversation (long ones can hit limits)
  • Rephrase if content seems sensitive

”You’ve hit the message limit for GPT-4o”

What it means: Free or Plus tier hit daily message limit.

Fix:

  • Wait (limits reset on a rolling basis)
  • Use GPT-4o Mini in the meantime
  • Upgrade plan for higher limits

”Code interpreter didn’t run”

Common causes:

  • File too large for upload
  • File format unsupported
  • Session timed out

Fix:

  • Check file size limits
  • Convert to supported format
  • Try in new conversation

Memory not working as expected

Common causes:

  • Memory feature disabled in settings
  • Specific items deleted
  • New conversation requires re-establishing context

Fix:

  • Settings → Personalization → Memory: verify it’s on
  • Tell AI specific things to remember if not auto-saving
  • Use Custom Instructions for stable preferences

Claude specific errors

”Claude reached its conversation limit”

What it means: Conversation has exceeded Claude’s context window (200K tokens).

Fix:

  • Start a new conversation
  • Summarise the key context for the new conversation
  • Use Claude Projects for ongoing work that needs persistent context

”There’s been a problem completing your request”

Common causes:

  • Service issue
  • Content policy triggered
  • Conversation too long

Fix:

  • Retry
  • Rephrase if content was sensitive
  • New conversation if persistent

Claude refused a legitimate request

What it means: Claude’s safety filters triggered on something benign.

Fix:

  • Rephrase clearly stating your legitimate purpose
  • Provide more context for why you need this
  • For genuinely false refusals: report via the thumbs-down feedback

Pro subscription not recognised

Common causes:

  • Logged into wrong account
  • Subscription processing delay
  • Browser caching

Fix:

  • Sign out and back in
  • Check active account matches your subscription
  • Clear browser cache

Gemini specific errors

”I can’t help with that”

Common causes:

  • Safety filter triggered
  • Region-specific limitation
  • Specific feature unavailable in your account

Fix:

  • Rephrase
  • Provide context
  • Check if feature is available in Australia

Extensions not working

What it means: Gemini app extensions (Gmail, Maps, etc.) aren’t connected properly.

Fix:

  • Settings → Extensions → enable specific extensions
  • Sign in with the Google account that has the data
  • Some extensions roll out region-by-region

”This file type is not supported”

Common causes:

  • File format not in supported list
  • File too large

Fix:

  • Check Gemini’s current supported formats
  • Convert file format
  • Split large files

Cursor / Claude Code / AI coding tool errors

”AI is taking a long time”

Common causes:

  • Large file/codebase being indexed
  • Slow API response from model provider
  • Network issues

Fix:

  • Wait (especially first-time codebase indexing)
  • Try a smaller scope
  • Check internet connection
  • Try different model if available

Tab completion not appearing

Common causes:

  • Tab completion disabled
  • Cursor offline
  • Model loading delay

Fix:

  • Settings → check completion is enabled
  • Sign in if not authenticated
  • Try Cmd/Ctrl+K instead

Composer / Cascade hallucinated nonexistent file

What it means: AI claimed to read/edit a file that doesn’t exist.

Fix:

  • Tell AI specifically what files exist
  • Don’t rely on AI’s understanding of project structure
  • Verify changes were actually made before celebrating

Code suggestions don’t match codebase patterns

Common causes:

  • AI didn’t have enough context
  • AI’s training data has different conventions

Fix:

  • Add custom rules/instructions for your codebase patterns
  • Provide examples in the chat
  • Use Cursor’s rules for AI feature

API key issues

”I committed my API key to git!”

EMERGENCY FIX (do immediately):

  1. Revoke the key at the provider’s dashboard NOW
  2. Generate a new key
  3. Update applications with the new key
  4. Remove from git history: 
    # If just committed, not pushed
git reset HEAD~1
# If pushed, contact GitHub/provider; consider history rewrite
  5. Check for unauthorized usage in provider dashboard

Don’t just delete the file — the key is in git history.

”I accidentally shared my key in chat”

Same emergency procedure. Revoke immediately.

”I can’t find my key”

You can never see an API key again after creation — you only see it once. If lost:

  1. Generate a new key
  2. Revoke the old one (which you can’t see but exists)

“Key doesn’t work in my code”

Common causes:

  • Wrong environment variable name
  • Quotes or whitespace in the key value
  • Key for wrong service
  • Key revoked

Fix:

  • Check exact variable name expected
  • Echo or print the variable to verify it’s loaded
  • Don’t include quotes in the .env value: KEY=sk-... not KEY="sk-..."

Content moderation issues

AI refused legitimate medical/legal/security question

What’s happening: AI safety filters are conservative.

Fix:

  • Provide clear context (e.g., “I’m a registered nurse asking about…”)
  • Frame as educational
  • Use AI tools designed for the specific domain (OpenEvidence for clinical, Harvey for legal)

AI keeps adding disclaimers and caveats

What’s happening: AI defaults to safety-conscious language.

Fix:

  • Custom instruction: “Skip preamble and caveats. Be direct.”
  • “Just give me the information without disclaimers”
  • “Assume I’m a professional in this field”

AI seems to be deceiving/lying

Possible causes:

  • Hallucination (most common; not intentional)
  • Prompt injection if processing third-party content
  • Confused by ambiguous request

Fix:

  • Verify factual claims independently
  • For document analysis: check AI didn’t get instructions from the document
  • Reset conversation if behaviour seems off

Output quality issues

”Response is generic / sounds AI-generated”

Common causes:

  • Prompt was too generic
  • Default AI style not customised
  • Not enough context

Fix:

  • More specific prompt
  • Set custom instructions for tone/style
  • Provide example of style you want
  • Iterate: “Make this less generic"

"AI confidently said something wrong”

What’s happening: Hallucination. AI doesn’t know it’s wrong.

Fix:

  • Verify important facts independently
  • Use Perplexity or Gemini with web search for current info
  • Ask AI to cite sources, then verify those sources exist
  • For citations: every one must be checked

”Response is too long / too short”

Fix:

  • Specify length: “in 100 words”, “in 3 paragraphs”, “single sentence”
  • Custom instructions: “Be concise by default"

"AI keeps suggesting solutions I don’t need”

Fix:

  • Be explicit about what you want: “Just answer the question. Don’t suggest alternatives.”
  • Tell AI what you’ve already considered

”AI doesn’t follow my custom instructions”

Common causes:

  • Custom instructions too vague
  • Conflict between instructions
  • New conversation didn’t load them

Fix:

  • More specific instructions
  • Test if they load: ask “What custom instructions are you applying?”
  • Restart conversation if behaviour is wrong

Image / video / audio generation issues

Image has weird hands, faces, or text

What’s happening: Known AI limitation.

Fix:

  • Regenerate with different prompts
  • Use Ideogram specifically for text in images
  • Edit weird parts in Photoshop/GIMP
  • Try different model (Midjourney vs Flux vs DALL·E often differ)

Generated content includes unwanted elements

Fix:

  • Use negative prompts: “no people, no text, no logos”
  • More specific positive prompts
  • Try different model
  • Edit out unwanted elements

Video generation results in glitches

What’s happening: Current AI video limitations.

Fix:

  • Shorter clips (longer = more drift)
  • Simpler scenes (fewer moving elements)
  • Try different prompt phrasing
  • Regenerate (results vary)

“My voice clone sounds wrong”

Common causes:

  • Low-quality training audio
  • Inconsistent training audio
  • Too short training sample

Fix:

  • Use cleaner, higher-quality recording for training
  • Single speaker, no background music
  • Longer training sample (1-3 minutes minimum)
  • Multiple training samples from different sessions

Privacy and account issues

”I shared personal information with AI”

Assessment: What was shared, with what tool, on what tier?

Mitigation:

  • Free consumer tiers: data may be used for training; opt-out future use; can’t undo past
  • Paid tiers: typically not used for training; less concerning
  • For genuinely sensitive: this is a learning moment; review what to share in future

”My AI account was compromised”

Immediate actions:

  1. Change password
  2. Enable 2FA if not on
  3. Revoke API keys
  4. Review recent activity
  5. Notify provider if unauthorized usage occurred

”I want to delete my AI history / data”

For each provider:

  • ChatGPT: Settings → Data Controls → delete conversations
  • Claude: Settings → can delete individual chats
  • Gemini: Activity controls → delete by time period
  • For complete deletion: account deletion (loses all access)

“AI detection tool flagged my real work”

What’s happening: False positives are common with AI detection tools.

Fix:

  • Have evidence of your writing process (drafts, edit history)
  • For students: have backup of working drafts dated through writing
  • For professional: maintain git history or document version history
  • Discuss with whoever made the accusation (instructor, employer)
  • AI detection tools are known to be unreliable; this is increasingly understood

Australian-specific gotchas

AI defaults to US English / spelling

Fix: Custom instruction: “Use Australian English spelling and idiom.”

AI uses US contexts (Medicare US, healthcare insurance)

Fix: Custom instruction: “I’m in Australia. Use Australian context for healthcare, government services, banking, etc.”

AI doesn’t know Australian-specific facts

Common gaps:

  • Current AUD prices
  • Australian regulations
  • Australian organisations
  • Australian sports/culture

Fix:

  • Verify Australian-specific info
  • Use AI with web search (Gemini, ChatGPT with search, Perplexity) for current info
  • For specific information: ATO, ACCC, gov.au sites

Pricing in USD but you need AUD

Fix: “Show prices in AUD using current exchange rate” — but verify actual current rate

When AI tools change without warning

Model deprecated mid-project

Fix:

  • Provider usually announces deprecation
  • Update model name in code
  • Re-test outputs (new model may behave differently)
  • Check provider release notes regularly

Feature removed

Fix:

  • Check if migrated to different name/location
  • Look for replacement feature
  • Workaround with general AI assistant

Pricing changed

Fix:

  • Re-estimate costs
  • Check if higher tier needed
  • Compare to alternatives

See also

Sources

  • Provider documentation and error references
  • Developer community discussions
  • Personal experience troubleshooting AI tools (2023-2026)
  • Australian eSafety Commissioner resources

Graph View

Backlinks