- Subject
- Nutrition
- Filed
- Source
- Sara Talpos
- Loans
- 333
What Scientists Mean When They Ask Whether AI Is Intelligent
Five experts say AI can learn and perform impressive tasks, but intelligence has many forms—and fluent answers don’t reveal how a system reached them.

AI systems can learn from data, adapt to feedback and produce answers that look like reasoning. Whether those abilities amount to intelligence depends on what the word means—and on how much we can infer from a system’s behavior. Five researchers in linguistics, computer science, cognitive science and psychology describe different ways to draw that line. Their shared caution is that intelligence isn’t a single, settled measure.
Language can sound meaningful without showing understanding
Emily M. Bender, a linguistics professor at the University of Washington, argues that language models work with the form of language without grounding words in meaning as people do. A model processes patterns in text; the meaning a reader finds in its output may reflect the reader’s own knowledge and interpretation.
Understanding language involves more than matching words to definitions. People consider what a speaker knows, what they intend to communicate and what they believe their audience will understand. Readers make these judgments almost automatically, which can make machine-generated text seem purposeful even when its production does not involve an author’s intentions.
Bender points to this tendency as a reason to be careful when judging what language models can do. In a 2021 paper, “On the Dangers of Stochastic Parrots,” she and her coauthors wrote that people interpret text through their own language skills and tendency to see coherent meaning and intent in communication, regardless of how the text was produced. That response can make convincing writing look like evidence of a mind behind it.
Her view is skeptical of claims that large language models think in a human-like way. The text they generate can prompt people to infer understanding, but that impression alone doesn’t establish that the model shares human cognitive experiences or knows what its words refer to.
Everyday tasks draw on more than one kind of ability
Melanie Mitchell, a computer scientist and professor at the Santa Fe Institute, questions the common division between cognitive intelligence and physical intelligence. Some definitions of artificial general intelligence focus on cognitive tasks, such as the kinds of work a chatbot can do, while leaving out physical activity. But human thinking is closely tied to bodies and social life, she says.
Even a seemingly straightforward repair shows why the boundary can blur. A chatbot may describe how to fix a dishwasher, but the person carrying out those instructions has to identify the right button or part in a real machine. The words don’t identify it for them in every situation; interpreting and applying them calls on the person’s knowledge of the physical setting.
Mitchell also points to social intelligence: coordinating with people to get things done. Someone who can’t fix a car might know whom to ask for help and how to describe the problem. She says it isn’t clear why AI systems often struggle with this kind of coordination. They may lack suitable training, or the experience of taking part in a social world.
That distinction matters when assessing claims about what AI can do. A system may produce useful information about a task without being able to carry it out, recognize what a particular situation calls for, or recruit the right help.
Learning and adaptation can count as intelligence
Anima Anandkumar, a professor of computing and mathematical sciences at Caltech, offers a broader definition: intelligence is the ability to learn from data and adapt to an environment. Under that definition, some AI systems already display intelligence. Chatbots can respond to people and be adjusted through feedback, including reinforcement learning, in which a system learns to pursue specified goals.
Human language is only one area where AI may help make sense of complex patterns, Anandkumar says. Models can also work with information from natural systems, from interactions between atoms and chemical reactions to weather at planetary scales. That work could support research into materials, medicines and forecasts of extreme weather.
She says her team created an AI-based weather model about five years ago that ran tens of thousands of times faster than traditional forecasting approaches and later exceeded their accuracy. Faster forecasts could help researchers assess risks and study extreme events. She also describes the prospect of predicting longer-term climate while accounting for physical laws, including in conditions the system hasn’t encountered before.
For Anandkumar, this work widens the conversation beyond whether machines resemble human thinkers. Intelligence can also describe the patterns and processes present throughout nature, even when people can’t readily see them.
There may be many intelligences, not one
Alison Gopnik, a psychology professor at UC Berkeley who studies child development, says the idea of artificial general intelligence assumes that one broad capacity called intelligence can be measured across people, animals and machines. Cognitive science, she argues, points instead to many abilities with different, sometimes conflicting, purposes.
Today’s large models succeed in part by detecting patterns in the enormous amount of text and imagery people have created, then generating new material from those patterns. Gopnik describes this as cultural transmission: a way for users to draw on information produced by earlier generations. She sees it as a real capability, but distinct from the abilities that let people produce that knowledge in the first place.
She also distinguishes between pursuing a goal and exploring an unfamiliar, changing world. Reinforcement learning can help AI systems master tasks such as games or coding, and it is used in training large models after their initial development. But young children learn by actively exploring their surroundings and building new ways to understand them. That exploratory capacity can pull in a different direction from goal-driven performance.
Care is another part of the picture. Children can explore because other people take on the demanding work of caring for them. Gopnik says this dependence is part of the conditions that support human learning, and current large models don’t have the same capacities. Other forms of artificial intelligence might approach some of them in the future, she allows, but that remains an open possibility.
Behavior is evidence, but it doesn’t settle the question
Benjamin K. Bergen, a cognitive science professor at UC San Diego, says current AI systems meet some common definitions of intelligence: they respond to their surroundings, learn from data and generalize beyond what they have seen. By some measures, they can also appear to reason logically and respond to people’s context or feelings.
But observed behavior doesn’t always reveal the process behind it. Bergen uses a horse that appears to answer arithmetic questions by stomping the correct number of times. At first glance, the horse might seem to understand the question and calculate an answer. If the horse is actually responding to tiny cues from its trainer, the performance points to a different ability: learning from experience, attending to social signals and stopping at the right time.
The example doesn’t make the behavior meaningless. It shows that similar-looking actions can arise through different mechanisms, and that those mechanisms may involve different kinds or degrees of intelligence. AI systems can likewise answer difficult questions in ways that seem to require deep reasoning, even when the answer can be traced to patterns in their training material.
There’s no single test in these accounts that resolves whether AI is “truly” intelligent. The researchers instead ask what ability a system displays, how it learned that ability, and what else the task requires. For readers weighing claims about AI, the practical takeaway is to look beyond fluent answers: performance can be impressive, but it doesn’t by itself show human-like understanding, social judgment or exploration.