When AI Goes Wrong

It will happen. Despite good policy, good training, and good intentions, an AI-related error will make it past your review process and into something external. The question isn’t if — it’s when, and how you respond.

Common Incidents

A fabricated statistic in a submitted proposal. AI generated a plausible-sounding data point. The reviewer didn’t catch it. The funder did.

PII in an AI tool. A staff member pasted a case study with real names into an AI tool before anonymizing.

An AI-generated email that doesn’t represent the organization. Someone used AI to draft a funder thank-you note. It sounded generic and corporate. The funder noticed the tone shift.

A compliance claim that’s incorrect. AI stated the organization meets a specific standard when it doesn’t.

The Response Framework

1. Acknowledge immediately. When you discover an error, don’t wait. If a funder has received content with a fabricated statistic, contact them: “We identified an error in our submitted proposal. The statistic on page X was incorrect. The accurate figure is Y.”

Honesty is faster, cleaner, and more relationship-preserving than hoping no one notices.

2. Assess the scope. How serious is this? A formatting error in an LOI is different from a fabricated statistic in a federal application. Match your response to the severity:

Low severity (formatting, minor wording issues): Correct internally, note the cause, move on.
Medium severity (inaccurate data in a submitted proposal): Contact the funder, correct the record, review how it slipped through.
High severity (PII exposure, compliance violations): Escalate to leadership, consult legal if needed, document the incident formally.

3. Trace the cause. How did this happen? Common root causes:

Review process failure (someone skimmed instead of spot-checking)
Workflow gap (AI output went directly to submission without the required review step)
Training gap (the person didn’t know to verify this type of claim)
Tool issue (the AI tool’s behavior changed or a team member used an unvetted tool)

Understanding the cause determines the fix.

4. Fix the process, not just the incident. The incident is a symptom. The process gap is the disease.

The incident is a symptom. The process gap is the disease. Fix the process, not just the incident.

If review failed: strengthen the review checklist, add specific verification steps for the type of error that occurred
If the workflow had a gap: add a mandatory checkpoint
If training was missing: add the scenario to your team’s shared learnings
If a tool was the issue: flag it to your evaluator and consider discontinuing use

5. Share the learning. After the incident is resolved, share what happened (anonymized if needed) with the team. “Here’s what went wrong, why, and what we changed so it doesn’t happen again.” This is Component 4 of the MVP Policy at work — shared learning from real experience.

What Not to Do

Watch out

Don’t blame the individual. If your review process allowed an error through, the process is responsible, not the person who used AI. Blaming individuals drives AI use underground.

Don’t ban AI reactively. One incident doesn’t mean AI is broken. It means your process needs improvement. Banning AI after an incident is like banning email after someone sends a message to the wrong address.

Don’t hide it. Internally or externally. Transparency builds trust. Cover-ups destroy it.

Preparing in Advance

Don’t wait for an incident to think about response. Prepare now:

Designate who handles AI incidents. Usually the same person who manages other operational issues.
Document your response protocol. A simple one-page checklist: acknowledge, assess, trace, fix, share.
Do a tabletop exercise. “What would we do if AI generated a fabricated statistic in our federal proposal?” Walk through the response as a team. It takes 30 minutes and is worth every second.

Philip’s Take: The organizations that handle AI incidents well are the ones that expected them. Not because they were pessimistic — because they were realistic. AI is powerful and imperfect. Plan for both.

Key Takeaways

AI incidents will happen -- preparation beats reaction
Response framework: acknowledge, assess scope, trace cause, fix process, share learning
Don't blame individuals or ban AI reactively -- fix the process
Do a tabletop exercise before you need the real thing

### Track Complete

You’ve completed all five modules of the AI Risk, Policy & Leadership track. You now have:

A thoughtful case for AI engagement in grant-funded work
A clear understanding of every major risk — and how to think about each one proportionally
A four-component framework as a living practice, not just a document
A rollout model (Science Fair) and a risk spectrum (four dimensions) for evaluating AI use
Leadership strategies for culture, skeptics, measurement, and incident response

When you’re ready, take the certification quiz to earn your “AI Risk & Policy for Grant Organizations” certificate.