In 2026, the question every organisation faces is not whether to adopt AI, but how to do so in a way that strengthens rather than hollows out its people. That question ran through the recent INSEAD AI Forum Americas at our San Francisco Hub for Business Innovation, where over 200 business leaders, scientists, founders and investors gathered to explore this challenge.  

Over two days, speeches from Marc Tessier-Lavigne of Stanford University and Xaira Therapeutics, Joshua Cohen of Apple University and Yossi Matias of Google Research shared a common thread: AI's promise is inseparable from the discipline of implementation. But it was self-proclaimed “mad scientist” Vivienne Ming, Chief Scientist at Possibility Sciences, who brought the event to a rousing climax by pushing the focus squarely back to humans.

The trap of the well-posed problem

I first encountered Ming’s work by chance in a Wall Street Journal article, which featured her experiments with University of California, Berkeley students on why some people collaborate so much better with AI than others. I wrote to her cold, not expecting a reply. She answered almost immediately, and that exchange has since turned into a research collaboration. With my INSEAD colleagues Ville Satopää and Alexandra Roulet, we are now working with Ming to replicate and extend her study. So, I listened to her keynote as both host and research partner. Here is what I believe leaders should take from it.

Ming, whose latest book is titled Robot-Proof: When Machines Have All the Answers, Build Better People, began with how we measure AI. Almost every benchmark tests models with what she calls well-posed problems: hard questions with known answers, such as the GMAT (Graduate Management Admission Test), Bar exams and Mathematical Olympiad questions.  

But the problems that matter most to organisations and societies are ill-posed. Nobody knows the answer in advance, and people often don’t even agree on the question. How do you measure whether AI helps people with those? And, more important, what kind of human intelligence factors determine the quality of human-AI collaboration? 

To find out, Ming used prediction markets. She took questions from Polymarket with enough liquidity to provide a credible baseline. One was the future price of oil, another the likelihood of new conflicts. Then she asked University of California Berkeley students to forecast them. 

On their own, the students did only slightly better than chance. AI models did far better: even a small, eight-billion-parameter open-source model, on its worst question, “was vastly better than the best human”. The interesting part is what happened when humans and machines worked together.

Automators, validators and cyborgs

Three patterns were observed. The largest group, roughly 60% to 70% of participants, were automators. They asked the model for an answer and pasted it in. Their results were slightly worse than the AI's alone. More troubling, EEG (electroencephalography) recordings of some participants, which measure electrical activity within the brain, showed a sharp drop in gamma-band activity, a marker of cognitive effort. As Ming noted from her Alzheimer's research, how much we exercise our minds over a career predicts our later risk of cognitive decline. 

The second group were validators, the “human in the loop” that’s often held up as the responsible model by company leaders and policymakers. These participants stayed engaged, but they asked the AI to confirm what they already believed, and the models were happy to oblige. The result? Validators did worse than humans working alone. The familiar promise that AI handles the boring work so people can do the interesting part is, said Ming, really a case of “the AI does the hard stuff so that the human doesn't understand the problem”. 

The smallest group, around 5%, were cyborgs. They explored, challenged and asked the model to tell them why they thought as they did – and sometimes told the model it was wrong. Reading their transcripts, you couldn’t tell who had ultimately made the decision. This hybrid intelligence outperformed the best humans and the best models, and came close to the collective judgement of Polymarket traders – the wisdom of crowds. 

It’s the human, not the model

Here’s the finding that should reshape how we think about AI investment. The models’ benchmark scores barely predicted cyborg performance; small open-source models worked nearly as well as frontier ones. What did predict it were human intelligence factors: fluid intelligence, curiosity, intellectual humility and perspective-taking. The capacity to ask “why”, to be open-minded, to understand other minds, and to create a path to a solution when even the question is initially ill-defined.

Ming drew on years of mentoring graduate students to explain why. Today’s models, she said, are like those students on steroids: “They know everything but understand nothing.” Supplying the understanding is the human’s job, and that takes intellectual humility which, she stressed, doesn’t mean deferring to the machine – it means being willing to expose your ideas, find where you’re wrong and learn from it.

When an AI simply hands us the right answer, our brain registers “I knew that”, and little learning occurs. Being told we are wrong feels like going backwards which, paradoxically, moves us forwards.

Designing for productive friction

True to form, Ming took the weakest model in the study and trained it never to give answers. She called it Socrates. Students “hated it”, she said, yet it produced the highest average hybrid intelligence of any condition, because it quadrupled the share of students working in cyborg mode, from about 5% to 20%. 

Ming closed with a challenge I will extend to every leader reading this: “Challenging people makes them better. Invest in better people by investing in technology that challenges us to be better.”

Join us for the upcoming INSEAD AI Forum Asia in Singapore from 30 to 31 October, and the INSEAD AI Forum Middle East in Abu Dhabi in January 2027.