After a recent study found that human children completely crush AI tools at basic problem-solving and thinking tasks, scientists have determined that AI has a serious blind spot: innovation.
The study was conducted by researchers at the University of California, Berkeley and published in the journal Perspectives on Psychological Science, provides a fascinating glimpse of the ability – or lack thereof – of currently available large language model (LLM) programs to generate truly innovative ideas. Trained on countless gigabytes of human-generated data, AI tools are incredibly good at predictive, statistics-focused tasks. After all, AI is not humans, but its job is to imitate them, and imitation naturally contradicts truly fresh, imaginative and inventive thinking.
But while this talent for imitation and pattern-discovery makes AI programs good at some tasks, it apparently makes them really bad at other tasks that are extremely rudimentary. On the other hand, human children? Unlike the AI, their innovative problem-solving skills were spot-on – suggesting there’s little more to human understanding and reasoning than the mass consumption of internet-scraped data (at least for now. , AI boosters might argue).
The study focused heavily on the use of tools, particularly tool innovation, as a means of testing problem-solving skills in its subjects. Tool innovation, as the researchers put it, “may involve designing new tools from scratch, but it can also refer to discovering and using old tools in new ways to solve novel problems.”
In other words, the ability to innovate does not just mean inventing a new thing like a lightbulb or an airplane. It also means being able to use existing equipment in unconventional ways. And so, to test each group’s innovation abilities, the researchers presented several AI models — a list that included models like OpenAI’s GPT-4 And GPT-3.5 Turbo, Anthropic’s Cloud, and Google’s FLAN-T5, among others, as well as several children, all aged three to seven, who have “a series of problems involving finding a target in the absence of specific tools” has to be executed.”
For example, in one such test, researchers asked subjects to draw a circle – which is simple enough! But instead of being provided with more traditional circle-drawing tools like a compass or stencil, participants were given the option of using a ruler, teapot, or stove to complete the task. The study found that 85 percent of the time, children made the correct choice, selecting the round-based teapot as a temporary stencil for their work.
Meanwhile, resource-gathering AI continued to reach out to the ruler; The only program that came close to success for children was GPT-4, which had a success rate of 76 percent.
It is noteworthy that the AI continued to choose the ruler. For most humans, apparently even children, the ruler is clearly not helpful for this specific goal, given its famously straight edges. But then again, the AI is predictable, and among the given options, the ruler is the only object that is traditionally used to draw some shapes. Meanwhile, tea utensils are made for tea only. So, by AI logic, Ruler makes the most sense.
“Discovering new functions in everyday devices is not just about finding statistical nearest neighbors from literal co-occurrence patterns,” the researchers write. “Rather, it is about appreciating more abstract functional analogies and causal relationships between objects that do not necessarily belong to the same category or be connected in the text.”
AIs also failed tests of the ability to infer novel causal structures, meaning the ability to detect new cause-and-effect means of achieving certain goals.
According to the study, researchers introduced each group to a virtual “blicket detector,” a silly object that lights up and plays music when placed over certain objects, but not others. It doesn’t matter why something is or isn’t a brick. UnderstandingBut that’s kind of the point – the goal is to test whether an AI or child can make studied observations about a phenomenon, and infer cause and effect from their experiments and observations.
Once again, during these tests, human children performed excellently, with the researchers writing that “even 4-year-old children automatically worked on the systems and discovered their structure – they discovered that “What were the bricks and what were they used to run the machine.” In contrast, according to the study, the AI ”struggled to produce relevant causal structures, even after vast amounts of training, compared to children.”
The study has some caveats, the most notable being the fact that it is difficult to measure and compare human cognition versus AI when there are no widely agreed definitions of flexible concepts like “intelligence.”
But the research is fascinating, and may well reinforce the notion that although AI models are largely designed in the image of the human brain – not to mention trained on countless human outputs – their reasoning processes naturally occur. Are different from humans. Similar in some ways, sure, but not definitively Same, And as we move into this emerging AI era, understanding those differences may prove essential to knowing where and how to use AI – and where it may ultimately not be so useful.
“Although we do not know the details of children’s learning algorithms or data,” the researchers wrote, “we do know that, contrary to older language and language-and-vision models, children are curious, active, self-observed, and internally Are inspired by.”
More on AI: OpenAI says GPT-5 is officially coming