Is AI taking entry-level jobs, and should a student start building?

Two sources, both read on 6 November 2026.

The Stanford Digital Economy Lab paper is called Canaries in the Coal Mine. It uses payroll records from ADP, a large United States payroll company, to follow employment by age and job. Its August 2026 revision covers data through June 2026.

An AI-exposed job is one where much of the daily work is the kind current AI tools can do, such as writing code or answering customer questions. In those jobs, employment of workers aged 22 to 25 stood about 19 percent below where it would have been had it kept pace with less exposed peers. Experienced workers in the same jobs show no comparable gap. The authors first documented the gap in August 2025, and it has widened since.

The Federal Reserve Bank of New York tracks the job market for recent college graduates, aged 22 to 27. For the third quarter of 2026 it reports 5.2 percent unemployment for recent graduates. It also reports 42.5 percent underemployment, meaning more than four in ten worked in jobs that do not usually need a degree.

They do not say that AI caused the drop. The Stanford authors call their results early, descriptive indicators. The paper tests other explanations, such as interest-rate rises and remote work, and the gap holds. It shrinks when the authors account for education. Some of it began before generative AI spread. It is also larger in this one payroll company's data than in national surveys.

They do not say that degrees are worthless or that graduate jobs are gone. At 5.2 percent unemployment, more than nine in ten recent graduates in the labor force had a job. The Stanford drop sits in a set of jobs and one age band. Treat these figures as a direction, not a forecast for you.

They also do not say that starting a company is the alternative. Many first builds find no users. That is normal and cheap within a term. It is expensive after a loan.

Entry-level jobs used to teach a set of skills on the job. A small real build now teaches the same ones. You find a problem someone has, talk to that person, scope the smallest version, ship it and read what happened. None of those is a coding skill. The tools now help with the code. They cannot pick the problem or sit across from the person who has it.

A side project built for yourself teaches the tools. A small thing built for other people, who can refuse it, teaches the rest.

Example, not a real student: a second-year nursing student sees every placement start with the same scramble over shift swaps in a group chat. She can see the problem and reach the people who have it. That shape, close and visible, is the one to look for.

1. Pick a problem you can watch. Your course, your campus, your part-time job or a family business. You need to see the problem happen and reach the people it happens to, without a marketing budget. 2. Write the riskiest assumption. One sentence that starts with "I believe". For the nursing example: "I believe students will post a swap in a tool instead of the group chat." If that is wrong, nothing else matters. 3. Test it before you build. A form, a spreadsheet, or a message asking ten people to try it by hand. The test fits in seven days and costs only your time. 4. Build the minimum only after people act. An MVP (minimum viable product) exists to learn, not to impress. If nobody posted a swap by hand, a nicer app will not change that. 5. Set a stopping rule on day one. Write down what result, by what date, means you stop. A term is long enough to learn and short enough to walk away.

If you want help with step one, nommi's Idea Finder starts from your background. You can add your hours and budget if you like. It suggests where to look.

The job figures come from the United States, from one payroll company and one survey. Other countries and fields may look different. The advice assumes you can spare a few hours a week, which not every student can. A term of building shows you can find and test a problem. It does not show you should start a company. Many people who do this well will still take a job, and the habit goes with them.

• Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, revised 12 August 2026, payroll data to June 2026. Read 6 November 2026. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
• Federal Reserve Bank of New York, The Labor Market for Recent College Graduates, third quarter of 2026. Read 6 November 2026. https://www.newyorkfed.org/research/college-labor-market