Artificial Summer School

I turned my three teenagers into an AI experiment. Here’s what the field report covers.

This summer I turned my three teenagers into a small, poorly controlled AI experiment. There was no control group. The subjects were related to the researcher. The researcher bought lunch, changed the curriculum while it was running and was occasionally the largest obstacle to progress. The Institutional Review Board consisted primarily of my wife asking what exactly I was doing to the children.

I was trying to teach them AI, but that was only half of it. I run a company that builds AI tools for credit unions, and for a couple of years we have been chasing a question I have become mildly obsessed with: how much expertise can you safely build into the environment around a person, so that someone without years of specialized training can do sophisticated work?

Three teenagers seemed like a sufficiently hostile test environment.

They built things. They broke things. We drove to an AI hackathon and built Gandalf, an autonomous learning agent. Warren later applied the approach to fantasy football, which is probably not the application the autonomous-agent research community had in mind but is exactly why teenagers make useful test subjects. Violet built a horse-valuation app. Jasper built a study platform and then got stuck for a week. I became an approval bottleneck in a program designed to increase autonomy.

This is not really a series about school

Education is where the collision is easiest to see, because schools are built around artifacts: essays, homework, tests, projects and grades. AI makes most of those artifacts radically cheaper to produce. But the same question is coming for every workplace. An employee can now draft a policy, analyze a spreadsheet, write code or research a regulation with AI doing much of the typing. The boss faces the same design problem as the teacher: how do you give people dramatically more capability without outsourcing their judgment, their accountability, or their compliance obligations along with it?

If your company’s AI policy is basically “don’t use it unless Legal says yes,” while employees quietly use it on their phones, congratulations: you have recreated my teenagers’ school AI policy. You do not have AI governance. You have an AI don’t-ask-don’t-tell program.

Over the next eleven pieces I want to work out what humans still need to know when machines can produce so much of the output; whether AI makes people lazy or ambitious; how teachers should teach and employers should train; what we should grade; what we should automate; and what has to stay human. The models will keep getting better. The more interesting question is what happens to us.

Sources and further reading

• Kirk Drake, “The AI skills gap is growing while schools wait for perfect answers,” The Next Web, 2026.
• OECD, Digital Education Outlook 2026, 2026.
• PwC, “The AI Workforce Planning Gap in Financial Services,” 2026.
• NCUA, Artificial Intelligence Resource Center for Credit Unions.