Will Relying on AI Stop Human Progress? A New Study Tests an Evolutionary Paradox
Overusing artificial intelligence for answers could eventually drag society’s collective knowledge down to a state worse than if the technology never existed at all. While LLMs offer cheap and instant access to information, they still depend on new human discoveries to remain useful in a constantly changing world. To understand why, a team of researchers tested what happens when humans stop learning for themselves and rely entirely on AI assistance, applying an evolutionary theory known as Rogers’ paradox to a theoretical model of human-AI interactions.
In the 1980s, anthropologist Alan Rogers proposed a thought experiment about evolution. He theorized that while learning from others — social learning — is cheap and easy, it relies on someone else putting in the hard, costly work of individual exploration. If every organism in a population chooses the easy route of copying others, a free-rider problem emerges where individuals act purely out of self-interest. No new knowledge is generated, and the population’s overall fitness plateaus at the exact same level it would have reached if social learning never existed.
In a recent paper published in Philosophical Transactions of the Royal Society A, “Revisiting Rogers’ Paradox in the Context of Human-AI Interaction,” MIT researcher Katherine M. Collins, former CDS Faculty Fellow Umang Bhatt, and CDS Faculty Fellow Ilia Sucholutsky updated this evolutionary model for the data science era. To map the paradox onto modern technology, the researchers simulated a slowly changing, stochastic environment. They then introduced an artificial intelligence agent into the simulation, treating the system as the ultimate social learner that aggregates knowledge from the entire population at once.
The team found that the paradox persists even with advanced technology. Because the system acts as an aggregator that relies entirely on existing human data, its marginal value drops as more people use it and fewer people conduct individual research. The equilibrium simply resets to a state where people are once again forced to learn individually. “The AI system sucks up all of the human knowledge we have, and doesn’t really create any new individual knowledge, at least so far,” Sucholutsky said.
The simulations revealed an even steeper drop in collective intelligence when the researchers accounted for cognitive decline. Recent studies suggest that overusing artificial intelligence can lead to de-skilling and weaken a person’s critical thinking abilities, making it harder for learners to recover their prior capabilities when the technology is taken away. When the researchers modeled this negative feedback loop, the population’s knowledge equilibrium dropped completely below the original baseline. “If it is true that AI causes cognitive decline in people who use it or overuse it, not only will that not lead to an improvement for societal discoveries and fitness, but perhaps it could even lead to this negative feedback loop that actually worsens our condition as a society,” Sucholutsky said.
The team originally set out to resolve the decades-old paradox, testing different orchestration strategies to see if they could break the equilibrium. They tried altering the models so that access would be restricted if the number of individual learners fell too low, but the simulations proved stubbornly resistant to optimization. “Pretty much every single thing we tried that we thought would be a promising way to resolve Rogers’ paradox ended up falling flat,” Sucholutsky said.
The findings highlight the pressing need to design artificial intelligence systems that act as collaborative thought partners rather than as outright replacements that stifle human discovery.
By Stephen Thomas
