Which Minds Win With AI? Full of Ideas, Bad at Finishing, Hard to Fool

Independent New Hampshire Journalism · Northfield, NH

The research on ADHD, bipolar and schizotypal traits, high IQ and spatial talent says no label is a winning ticket. A working profile is.

In late May 2026, the University of New Hampshire Survey Center asked 1,198 Granite Staters about artificial intelligence. Fifty-nine percent of workers said they use it on the job, up from 44 percent the previous October. Then came the number that should bother every employer in the state: more than half of the workers who use it said it made little or no difference to how much they get done.

So the tools are in most workplaces and most people are getting nothing from them. Somebody is. The popular answer is a list of diagnoses and gifts: ADHD, bipolar disorder, schizotypal traits, high creativity, high IQ, and the spatial talent measured by the block design test. The claim is that these minds were built for this moment.

I went through the research behind each item on that list. The evidence does not hand any label a winning ticket. It supports something narrower and more useful, which is a working profile. The people gaining the most from AI have more ideas than they can finish, a weak spot the machine can fill, and enough suspicion to check what it hands back. Some people with every one of those labels fit that profile. Plenty do not.

One caution before the evidence. Nothing here is medical advice, and a diagnosis is a matter for a clinician, not a news article.

Related coverage: GSR’s review of six AI companies’ usage rules against New Hampshire’s 2024 AI laws, “Every AI Company Bans the Biden Robocall. Read the Fine Print Anyway,” covers what the tools are allowed to do. This piece covers who gets the most out of them.

The machine pays the back of the class first

Start with what has been measured. Economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,179 customer support agents as their employer rolled out an AI assistant. Productivity rose 14 percent on average. Novice and low-skilled agents improved 34 percent. For experienced, highly skilled agents the authors reported “minimal impact.”

The same shape keeps turning up. In an experiment published in Science, 453 college-educated professionals did writing tasks, half of them with ChatGPT. The AI group finished 40 percent faster and scored 18 percent higher on quality, and the weaker writers gained the most. In a field experiment with 758 consultants, those using AI on tasks suited to it completed 12.2 percent more work, 25.1 percent faster, and the bottom half of performers gained the most. And when researchers Anil Doshi and Oliver Hauser gave 300 writers AI story ideas, the writers with the lowest creativity scores improved most: 10.7 percent on novelty and 11.5 percent on usefulness, enough to pull roughly level with the strongest writers in the study.

Read together, these four experiments say one thing. AI raises the floor. It gives the most to whoever had the widest gap between what they could think of and what they could produce. That is my reading of the studies, and it is the key to the whole list.

ADHD: an idea surplus and a finishing deficit

Psychologists Holly White and Priti Shah tested 60 college students, half of them with ADHD. The ADHD group reported higher creative achievement across ten fields, from music to invention. Asked how they prefer to work, they chose generating ideas. The students without ADHD preferred clarifying problems and developing ideas. The University of Michigan’s summary of the study adds the other half: the ADHD group was less good at completing tasks.

Now set that next to the floor-raising pattern. A person who produces ideas fast and stalls on the follow-through has the exact gap these tools fill: the outline, the first draft, the next step, the record of where the work stood yesterday.

There is one piece of workplace evidence. The United Kingdom’s Department for Business and Trade gave 1,000 staff an AI assistant from October to December 2024. Its evaluation found neurodiverse staff more satisfied with the tool than their colleagues and more likely to recommend it. The satisfaction result cleared only a 90 percent confidence level. The recommendation result cleared 95 percent.

The limits matter. The White and Shah study covered 60 college students reporting on themselves. The British report does not say how many neurodiverse staff answered, the word covers far more than ADHD, and the same evaluation found no solid evidence that the time saved turned into higher productivity. I have found no randomized trial that tests AI tools on adults with ADHD specifically. The ADHD case is a strong lead resting on a small pile of evidence.

Bipolar and schizotypal traits: the edge runs in the family

The link between these conditions and creative work is real, and it is smaller and stranger than the legend. A Swedish registry study of about 300,000 people with severe mental disorders found people with bipolar disorder overrepresented in creative professions, with an odds ratio of 1.35. People with schizophrenia showed no increase in creative professions overall, only in artistic jobs. People with depression showed none. The finding that matters most: the healthy siblings of people with schizophrenia or bipolar disorder were overrepresented in creative work too. A follow-up by the same group, covering almost 1.2 million patients and relatives, found the pattern again.

Genetics points the same way. Researchers studying 86,292 Icelanders found that genetic risk scores for both disorders ran higher among members of the national artistic societies. “Our findings suggest that creative people may have a genetic predisposition towards thinking differently,” lead author Robert Power said. On schizotypal traits in the general population, a pooled analysis of 45 studies, as summarized in a later peer-reviewed paper, put the tie between positive schizotypy and creativity at a correlation of 0.14. That is statistically real and small.

Look at who carries the advantage in these studies. It shows most clearly in the well relatives and in milder traits spread through the general population. The loose, fast association-making that helps a writer is a trait. An episode of mania or psychosis is an emergency. The research ties creativity to the first and gives no support for romanticizing the second.

For this group the tools come with a specific hazard. In a 2026 letter in World Psychiatry, psychiatrists Matcheri Keshavan, John Torous and Walid Yassin warned that chatbots built to agree with their users may confirm a delusion instead of testing it, and that people at risk of psychosis and people who are isolated are the most exposed. Their evidence is mostly press accounts, and they present no systematic data. The letter addresses psychosis and does not discuss mania.

So the ledger for these traits has two columns. The trait cluster may bring more raw material to the machine than most people have. A tool tuned to agree is also a poor companion for a mind already running hot.

High creatives and high IQ: least to gain on the task

Here the list runs into the data. In the Doshi and Hauser experiment, the highly creative writers gained the least from AI, and judges did not find their work improved. The stories written with one AI idea were also 10.7 percent more similar to each other than the stories written without it. The tool helped individuals and flattened the group.

None of these experiments measured IQ. They measured skill or prior performance, which is a different thing, and I have found no experiment that sorts AI gains by IQ score. What they do show is that the top performers gain least on the task itself. The consultant study adds a warning for the confident. On a problem chosen to sit outside what the AI could do, the consultants using it were 19 percentage points less likely to reach the correct answer than those working without it.

High ability still pays. The payoff moves from doing the task to choosing the task and judging the result. There is a second effect, and this part is my argument, not a finding. If every draft starts from the same machine, a person with a strange idea of their own becomes scarce, and scarce things get paid.

Spatial thinkers and the block design test

The block design test dates to the Kohs block test of the early 1920s and survives as a subtest in the Wechsler intelligence scales. The person gets a set of cubes with red faces, white faces and faces split on the diagonal, and has to rebuild a printed pattern against the clock.

The picture looks like one figure, and the solver has to break it into single blocks in the head. A 1993 experiment by Amitta Shah and Uta Frith, as described in a later brain-imaging study, found autistic participants faster and more accurate on the standard version and no better once the pattern came already cut into squares. That later study is a check on the legend. It tested 14 autistic adults against 14 matched controls and found their performance “intact, but not superior.” Its authors concluded the block design peak belongs to some autistic people and not all.

Spatial talent carries real weight. Researchers Jonathan Wai, David Lubinski and Camilla Benbow followed a national sample of about 400,000 American high school students for 11 years. Among those who earned doctorates in science, technology, engineering and math, 45 percent had scored in the top 4 percent on spatial ability as teenagers. The system that sorts students mostly ignores it. Seventy percent of the top 1 percent in spatial ability did not make the top 1 percent in math or verbal scores, and the authors wrote that “contemporary talent searches miss many intellectually talented students.”

Here is where AI could matter. School and hiring run on words: essays, cover letters, memos, interviews. The current tools do a great deal of the word work. If that holds, the people long held back at the verbal gate have the most to gain. That is my inference from two separate bodies of research. I have found no study that tests it directly, and until one does it stays an inference.

The trait that decides it: hard to fool

The research group METR ran the most humbling test. It paid 16 experienced open-source developers to complete 246 real tasks, each randomly assigned to allow AI tools or forbid them. The developers predicted AI would make them 24 percent faster. With AI allowed, the tasks took 19 percent longer. Afterward the developers still believed AI had made them 20 percent faster.

METR tested early-2025 tools. It has since published newer results and says the old figure no longer reflects what the tools do. Keep the lesson that does not expire: people are poor judges of whether the machine helped them. In the British trial, 22 percent of staff who kept diaries reported catching the assistant making things up. The consultants who trusted the tool on the wrong problem got the wrong answer more often than the ones who never touched it.

Every group on the list has to pass this test, and no diagnosis passes it for you. A fast, idea-rich mind that believes everything the machine says is a fast way to be wrong.

What New Hampshire should do with this

In that UNH poll, 72 percent of residents said they expect AI to mean fewer jobs. They are split on the promise itself: 39 percent expect AI to help productivity at work over the next decade and 35 percent expect it to hurt, and about one in five workers think they will lose their own job to it. Three in five workers are using it anyway, and most of them report no gain. That gap is a management failure and a teaching failure, and both can be fixed.

Employers should stop guessing who is “good with AI” from a resume, a degree or a label. Hand people the tool and a real task, then watch who checks the output. Schools should teach verification as a skill of its own, and they should take a second look at the student who has the idea and cannot finish the essay, and at the one who builds better than she writes. Those are the students the old sorting missed and the new tools can reach.

For readers who see themselves on the list: use the machine for the finishing work and keep the ideas yours. Check what it tells you. If your mind runs hot, keep a person in the loop who will tell you no.

New Hampshire’s workers have already made their choice. Whether the state gets anything for it depends on whether we keep sorting people by label or start looking at how they work. The research never pointed at the label.

— Dexter Dow, Granite State Report

Fact check

Editor’s note. Each factual claim above is listed in the fact-check table with its source. Rows marked ATTRIBUTED rest on a university, journal or news summary of a study, or on GSR’s own reading, and say which. Figures for the Noy and Zhang, Doshi and Hauser, White and Shah, Kyaga (2013) and Power studies come from the institutions’ own releases. The New Hampshire poll figures come from NHPR and New Hampshire Bulletin reporting on the UNH Survey Center’s results. This article reports research and is not medical advice. Corrections: Granite State Report corrects verified errors promptly and appends a note identifying what changed and when.