Despite the common comparison between artificial neural networks and biological brains, today’s approaches to artificial intelligence involve energy-hungry models that consume enormous resources, in stark contrast to the efficiency of human learning. Both systems—human brains with their billions of neurons and large language models with simulated ones—remain enigmatic, particularly in their natural facility for producing language. Scientists see promise in pushing artificial intelligence to not just mimic but unravel the mysteries of human cognition.
Recent studies published in Nature highlight efforts to bridge psychology and artificial intelligence. At the forefront is Centaur, a large language model fine-tuned on data from 160 psychology experiments. Unlike traditional models relying on basic equations, Centaur achieved significant improvements in predicting human behavior across tasks like memory recall and decision-making. Beyond prediction, researchers hope that unpacking how Centaur arrives at its choices will reveal mechanisms of thought, potentially guiding the development of new psychological theories. However, some cognitive scientists urge caution, noting that good behavioral mimicry doesn’t guarantee cognitive similarity; complex outputs don’t always reflect similar underlying processes—much like comparing a calculator to a human solving math problems.
To address the challenge of interpretability, other researchers are experimenting with much smaller neural networks. These micro-models, sometimes containing only a single artificial neuron, show remarkable predictive accuracy in behavioral experiments with animals and humans. Their small size allows for transparency: scientists can track each neuron’s activity and draw clearer connections between model structure and output. Still, each of these networks is limited to narrowly defined tasks, unlike the broader versatility of Centaur. The field faces a fundamental trade-off: larger models excel at prediction but elude understanding, while smaller models offer insight yet handle only simple problems. As artificial intelligence advances, the pace of predictive breakthroughs continues to outstrip our ability to interpret them, highlighting the enduring gulf between knowing what will happen and understanding why.
