Try AI where nothing is at stake

When someone tells me they don’t know where to start with AI, I ask where they’re trying to start. The answer is almost always work.

Work is the hardest place to begin. The stakes are real, other people are watching, and one bad answer can feel like a verdict on you. You might also not be sure what you’re allowed to paste in.

So here’s what I’d do instead. Start somewhere nothing is at stake.

Why practicing outside work helps

There are two reasons, and neither needs a psychology degree.

The first is confidence. Psychologist Albert Bandura spent his career studying what makes people believe they can do something. He found that the biggest builder of that belief is trying something and having it work. Watching others do it helps, and so does being cheered on, but neither beats a win of your own. Later research also notes that early failures can shake your confidence, especially when you decide the problem is you. So your first few tries matter, and a high-stakes first try is a bad way to get a win.

The second is how people react. Researchers at Duke ran four experiments with more than 4,400 people and found that employees described as using AI at work were rated as lazier and less competent than employees getting other kinds of help. People imagining themselves in that spot expected the judgment and said they’d be less likely to tell their manager or colleagues. You can read the study here, and it’s free. It looked at workplaces only, so I can’t say how friends and family would react. My guess is that a dinner plan carries far less risk to your reputation than a work deliverable.

A simple way to run an experiment

Keep it to one small loop over one week:

• Pick one small, real problem.
• Guess what will happen before you try.
• Try it, and count one or two things you can actually measure.
• Decide whether to keep it, change it, or drop it.

Four experiments to try

• The group chat test. Take a plan your family or friends are already making, like a trip, a reunion, or a gift. Use AI to write a quick recap after each burst of messages and to keep a running list of what’s decided and what’s still open. Count how many “wait, what did we decide?” messages show up compared to the last time you planned something.
• The week of dinners. Tell it what’s in your fridge and what your household will actually eat, and ask for a plan. Track how long it took you and how many extra grocery trips you made.
• The explain-it test. Pick something you’ve been curious about and ask for an explanation. The next day, try explaining it back without help. That tells you whether you learned it or just read it.
• The second opinion. Take a small decision, like what to buy or how to word an awkward text. Ask AI for the strongest case against your plan before you ask it to help carry the plan out. It’s a good way to see how easily the answer you want gets handed back to you.

Keep three lines of notes

What I tried. What happened. What I’d change. That’s enough, and it’s what turns playing around into learning.

If you do have to start at work

Sometimes you don’t get a choice. Your company rolled out a tool, or your manager wants results by next month. In that case, the thing that matters most is whether you have room to get it wrong.

Harvard researcher Amy Edmondson studied 51 work teams and found that teams where people felt safe taking small social risks, like asking a question or admitting a mistake, did more learning. She calls it psychological safety. In plain terms, it’s whether you can try something, mess it up, and say so without it costing you.

If you’re the one trying, start where a miss is cheap. Pick a task where AI obviously fits and a mistake doesn’t matter much, like a first draft of an internal note, a meeting recap you’ll check, or a list of ideas. The Duke study found the penalty for using AI was offset when the task was one where AI was clearly useful. Don’t do it in secret, either. Use the tools your company has approved, keep anything confidential out, and tell your manager what you’re trying, because a trial you’ve announced is much safer than one you’re hiding. Then give it a time limit. Two weeks is plenty. Share what worked and what didn’t, including the flops, since the flops are where most of the learning is.

If you manage people, you’re the one who makes the space. Edmondson’s advice includes treating the work as a learning problem, not a performance test, and admitting your own mistakes. Say out loud that it’s an experiment. Share one of your own bad prompts. When someone shows you a first attempt, ask what they learned from it before you judge how it turned out.

What you walk into work with

After a few of these, you won’t go into work with an opinion about AI. You’ll go in with evidence about what it does, what it doesn’t, and how you handle both. That confidence comes from having done it, which makes it the kind that lasts.

You learn faster when a bad answer is a data point and not a verdict.

Happy Learning,

KP