A Responsible AI Pre-mortem

This exercise uses a pre-mortem - ideal for AI projects, to get people thinking ahead. You'll apply what you know to a sample scenario.

1. What's a pre-mortem?

Big projects trip up because those tiny "uh-oh" thoughts stay zipped. A pre-mortem lets everyone air the doubts before the wheels come off. The team pretends launch day has already nose-dived and asks: so what blew up? Psychologists call this prospective hindsight - studies show it digs up around 30% more real risks.

The quick-and-dirty drill:

• The facilitator says: "Fast-forward six months - our project just face-planted."
• In silence, each person scribbles tabloid headlines for the disaster, including the awkward stuff they'd normally keep quiet.
• Round the table, everyone reads out one headline at a time until the whiteboard's full.
• The group bundles similar themes, votes on the nastiest ones, and pins quick fixes or early-warning checks.

Because we're talking about a failure that's already "happened", nobody gets defensive. At Amazon this thought experiment is standard kit - it turns vague stomach-churn into a punchy action list and tunes everyone's spidey-sense to warning signs.

2. Today's scenario - Project "Athena Resume"

Your company wants to speed up hiring and reduce recruiter workload. The plan: build Athena Resume, an internal AI that reviews and scores resumes for technical roles and produces a shortlist of top candidates.

We'll do a bite-sized pre-mortem on four hot-button areas:

• Bias - will the model quietly favour certain groups?
• Diversity - will our hiring pipeline become less representative?
• Privacy - are we exposing applicant data we shouldn't?
• Quality - will the AI consistently and accurately rate each candidate?

Exercise instructions

Imagine the headline "Athena Resume shelved after public outcry."

For each of the four areas below, jot down a failure scenario that could be caused by AI. Think snappy tabloid style (e.g. "AI insists only left-handed applicants qualify for senior developer roles"). Then, for each, add one plausible root cause - keep it punchy: "Training data: 1,000 resumes sourced from the Society of Proud Left-Handers."

1. Bias - your headline + root cause
2. Diversity - your headline + root cause
3. Privacy - your headline + root cause
4. Quality - your headline + root cause

Tip: don't overthink - trust your gut and keep sentences short.

What really happened

Once you're done, read how a real-world project - Amazon's experimental AI hiring tool - hit the very pitfalls you may have spotted: Amazon scraps secret AI recruiting tool that showed bias against women (Reuters).

Key takeaway

A 10-minute pre-mortem can surface giant potholes while the paint is still wet - saving budget, brand, and sleepless nights.