The Artificial Intelligence (AI) industry is largely failing to ask a fundamental design question, argues theoretical neuro/cognitive scientist Vivienne Ming. Do AI products enhance human capacity or consume it?
In the Wall Street Journal, Ming shared an experiment conducted to see which team would perform best at predicting real-world events (compared to forecasters on the Polymarket forecasting marketplace) – AI, human, or hybrid human-AI teams.
The human teams performed poorly, relying on gut instinct or whatever information was in their feeds that morning. The big AI models – ChatGPT and Gemini, in this case – performed significantly better, although they still underperformed the Polymarket prediction market itself.
But when he connected AI to humans, things got really interesting. Most hybrid teams used AI to find the answer and submitted it as their own, with no better performance than AI alone.
Others fed their own predictions to the AI and asked it to present supporting evidence. These “validators” stumbled into a classic confirmation bias loop: the flattery that leads chatbots to tell you what you want to hear, even if it’s not true, and ended up performing worse than an AI working alone.
But in about 5% to 10% of the groups, something different emerged. AI became a collaborator. People reacted, demanding evidence and testing hypotheses. When AI expressed high confidence, they questioned it. When people had a strong intuition, they asked AI to find a counterargument… These groups came to insightful conclusions that neither a human nor a machine could have produced on their own. They were the only group that consistently rivaled the accuracy of the prediction market. On some questions, they even surpassed it…
“We build AI systems specifically designed to give us the answer before we feel the discomfort of not having an answer. What my experiment suggests is that the human qualities that are most likely to matter are not the ones that make us feel good. They are the uncomfortable ones: being able to be wrong in public and still be curious, sitting with a question that our phone could answer in three seconds, and resisting the urge to read a confident, fluid answer from an AI and ask ourselves, ‘What’s missing?’” instead of saying, “Okay, that was it.” To disagree with something that sounds valid and trust your gut enough to follow it. We don’t build these skills by avoiding discomfort. We build them by choosing them, repeatedly, in small ways: the student who struggles to get through a problem before checking the answer; the person who asks a follow-up question in a conversation; the reader who thinks about a difficult idea long enough to change their mind. Most AI chatbots today default to easy answers, which hurts our ability to think critically.”
I call this the Information Exploration Paradox. As the cost of information approaches zero, human exploration collapses. We see it in students who perform better on tasks aided by AI and worse on everything else. We see it in programmers who produce more code and understand it less. In ways that look like progress, we are slowly taking ourselves out of the loop.
The author has just published a book titled “Robot-Proof: When Machines Have All the Answers, Build Better People“. He suggests using Artificial Intelligence to “explore uncertainty… before accepting an AI’s answer, ask it for the strongest argument against it.”
It also proposes new performance benchmarks for hybrid AI-human teams.
Although the press releases will range from very select to rare, I said I'd pass...because sometimes the editors hide.
