What Harvard’s AI Experiment in Northampton County Just Found

One of America's most interesting experiments with artificial intelligence is happening in the Lehigh Valley's Northampton County as data center politics roil the midterm elections.

And, almost nobody knows about it.

The experiment is not about self-driving cars, humanoid robots, or replacing thousands of workers with machines.

Those decisions are difficult not because government lacks information. In fact, the opposite is often true.

Caseworkers have extensive histories available to them, including prior referrals, investigations, services, placements, and family relationships. The problem is deciding which cases deserve the most attention.

Northampton County tested whether artificial intelligence could help.

Researchers from Duke and Harvard studied 3,661 child referrals between December 2023 and February 2025. The county's system assigned cases to one of five risk categories, from Very Low to Very High, based on existing administrative records. Some supervisors could see the score. Others could not. Everything else remained essentially the same.

Most important, AI did not make the decision.

The algorithm simply took information already sitting in the county's files and turned it into a clearer warning signal.

The results were striking.

When supervisors could see the AI-determined risk score, foster-care placements increased overall. But they did not increase indiscriminately. The effect was concentrated among children that AI classified as the highest risk.

For those Very High risk cases, access to the score increased foster-care placement by 5.2 percentage points, ongoing services by 10.4 points, and substantiation of abuse or neglect by 14.8 points. For most lower-risk children, the changes were generally smaller and statistically indistinguishable from zero.

In other words, AI helped government workers prioritize and focus.

Then came the most important result.

Among children whose supervisors could not see the score, roughly 25.8% were referred back into the child protection system within three months. Giving supervisors access to AI reduced those subsequent referrals by 3.8 percentage points, a 15% decline.

The study produced another result worth considering.

One of the most serious criticisms of AI is that it may reproduce racial disparities buried inside historical data. Yet the Northampton County researchers found no systematic evidence that making the risk score visible widened racial or ethnic disparities.

In fact, the estimated reduction in subsequent referrals was largest among Black children.

Perhaps the most surprising finding, however, involved the workers themselves.

The benefits of AI appeared larger among social workers who had performed better before the experiment. While not conclusive, it raises an important possibility.

Artificial intelligence may make expertise more valuable.

That is almost the reverse of how the AI debate is normally framed. We spend enormous amounts of time asking which workers computers will replace.

The Northampton County study suggests a different question.

What happens when we give good social workers better tools?

Government agencies everywhere face the same basic problem: too many cases, too much information, and too little time. Police investigators, health inspectors, fraud investigators, regulators, auditors, and social workers all must decide where to focus limited attention.

That brings us back to the midterm elections.

Across Pennsylvania, voters are being asked to think about AI largely through the physical infrastructure required to power it: data centers and enormous new demands on the grid.

Those are legitimate questions, and candidates should have to answer them.

Northampton County provides an important reminder of the other side of that ledger. AI has initially proven capable of helping county workers identify the highest-risk cases, direct scarce resources more intelligently, and make better use of the information government already possesses.

So as AI becomes an issue in the midterm elections, perhaps the debate should be larger than whether Pennsylvania wants the infrastructure that powers artificial intelligence.

We should also be asking what we want artificial intelligence to do for Pennsylvania.

Athan Koutsiouroumbas is a managing director at Long Nyquist and Associates and a former congressional chief of staff. Follow him on X at @Athan_K.