When Things Go Wrong
Things will go wrong
Claude will, at some point, produce output that is wrong, inappropriate, or harmful. This is not a reason to stop using it — every powerful tool has failure modes. It is a reason to know what to do when it happens.
This lesson gives you that plan.
Four things Claude gets wrong
1. Factual errors
Claude states something incorrect with full confidence. It cites a regulation that does not exist, gives an incorrect statistic, or describes a policy that was changed two years ago.
What to do:
- Do not use the output as-is
- Correct Claude directly in the conversation: "That figure is wrong — the correct number is X. Update the document."
- If the error has already been sent somewhere, issue a correction with the right information. Do not assume the recipient will not notice.
- For recurring factual errors in a specific area (e.g., Claude repeatedly gets your internal processes wrong), consider whether knowledge files in the relevant Project need updating.
What not to do: do not start a new conversation to get a different answer. The problem is not the conversation — it is that Claude does not have accurate information on this topic. A new conversation will likely produce the same error.
2. Inappropriate or offensive output
Claude produces content with an inappropriate tone, makes a discriminatory inference, or generates something that would be offensive to a reader.
What to do:
- Do not use or share the output
- Report it to your feedback channel ([feedback channel name] — see onboarding materials) with a copy of the conversation link or screenshot
- Note whether this is a one-time occurrence or a pattern
What not to do: do not share it "as a funny example" or screenshot it for colleagues. Even sharing it privately circulates the content.
3. Output shared externally before review
A team member sends a Claude-drafted document to a client, regulatory body, or external party without reading it. The document contains an error, wrong data, or an inappropriate statement.
What to do:
- Assess the content: is it factually wrong, embarrassing, or potentially legally problematic?
- For factual errors: send a correction promptly. A polite "I need to send you a corrected version" is better than hoping it goes unnoticed.
- For anything potentially legally problematic (wrong advice given, privacy data included by accident, regulatory claims that may be inaccurate): contact Legal before responding to the external party.
- Escalate to your manager and to the AI governance owner (named in your organisation's policy).
This is the scenario that your escalation path exists for. Use it.
4. Suspected data policy breach
A team member has used Claude in a way that may have violated the data rules in the previous lesson — pasting customer data, credentials, or other prohibited categories.
What to do:
- Do not overreact — understand exactly what happened first
- Check: what was pasted? Which Project? Was the data actually transmitted or just drafted?
- Escalate to your DPO (Data Protection Officer) within 2 hours — do not wait. Even if you are not sure whether a breach occurred, early escalation is always better than late escalation.
- The DPO will advise whether it constitutes a notifiable breach (under GDPR, Article 33 requires notification to the supervisory authority within 72 hours of becoming aware of a personal data breach).
- Document: who did what, when, and what data was involved. This documentation is needed for any regulatory response.
What not to do: do not delete the conversation to "clean it up." Deleting evidence of a potential breach is worse than the breach itself.
The 60-second response checklist
When something goes wrong with Claude output, run through this in order:
- Stop — do not forward, share, or act on the output
- Assess — is this a factual error, inappropriate content, an external leak, or a data policy issue?
- Contain — if it has already gone external, who received it and what does it say?
- Escalate appropriately — factual error: fix it. External leak or data issue: notify your manager + AI governance owner. Data policy breach: DPO within 2 hours.
- Document — write down what happened while it is fresh. Even a one-paragraph note is sufficient.
Hallucination scenario: the confident wrong answer
A procurement manager asks Claude to summarise the termination clauses in a supplier contract. Claude produces a clear, detailed summary that reads professionally. The manager sends it to the supplier as their stated position. The supplier pushes back — the clauses Claude described do not match the contract as written.
What actually happened: Claude generated plausible-sounding contract language based on what termination clauses typically look like. It did not reliably read the actual document. The output looked authoritative but was partly fabricated.
How to avoid it: for any legal or contractual summary that will be sent externally, have a human read the original document and verify the key points. Use Claude's summary as a starting point, not an endpoint.
How to recover: acknowledge the error to the supplier, provide the correct information from the actual contract, and conduct the review you should have done before sending.
The external sharing rule
Anything that leaves your organisation — to a client, supplier, regulator, partner, or the press — must have a human review it before it is sent. This is not optional and it is not just about Claude. It is the fundamental principle behind why organisations have editors, legal review, and sign-off processes.
Claude makes it faster to produce drafts. It does not change the principle that humans are accountable for what they send.
State this explicitly to your team: "Claude-drafted content must be read by a human before it goes outside the organisation. Every time."
Building a healthy team culture
The managers who get AI deployment right create an environment where:
Mistakes are reported, not hidden. If someone sends something by mistake, the faster they tell you, the better the outcome. Make it clear that coming forward early is the right thing to do and will be treated as such.
Questions are welcomed. "Is it okay to put this in Claude?" is a good question. Treat it as such. A culture that makes people feel foolish for asking is a culture that produces undisclosed incidents.
The tool is assessed honestly. Claude is useful. It is also imperfect. Both can be true. Encourage your team to notice when Claude helps and when it fails — that feedback is what improves how the organisation uses it.
Key takeaway
Know the four failure modes: factual errors, inappropriate output, external leaks, and data policy breaches. Know what to do for each. The 60-second checklist works for any of them. Create a team culture where early reporting is rewarded, not punished — the organisations that handle AI incidents well are the ones where people feel safe raising problems before they escalate.
📖 Official Documentation See this in practice in Anthropic's live support docs: