Can AI Understand Our Minds?

Author(s): Vita Haas

Originally published on Towards AI.

When it comes to artificial intelligence (AI), opinions run the gamut. Some see AI as a miraculous tool that could revolutionize every aspect of our lives, while others fear it as a force that could upend society and replace human ingenuity. Among these diverse perspectives lies a growing fascination with the cognitive abilities of AI: Can machines truly “understand” us? Recent research suggests that advanced language models like ChatGPT-4 may be more socially perceptive than we imagined.

A recent study published in Proceedings of the National Academy of Sciences (PNAS) reveals that advanced language models can now match a six-year-old child’s performance in theory of mind (ToM) tasks, challenging our assumptions about machine intelligence.

What Makes This Discovery So Important?

So, what exactly is “theory of mind”? In psychology, ToM is the ability to infer and understand others’ mental states — their beliefs, desires, intentions, and knowledge. This capability allows us to predict how people will act based on what they believe, even if those beliefs don’t align with reality.

For example, if you see a friend looking for their keys where they left them, but you know someone moved them, you understand that they’ll look in the wrong place. This kind of thinking is fundamental to human social interactions, empathy, and moral judgment. Without it, simple misunderstandings and social conflicts would be far more common.

The Stanford Study That Changed Everything

Enter Michal Kosinski from Stanford University. His research took 11 different language models through a gauntlet of 40 false-belief tasks — think of them as sophisticated psychological tests designed to catch out whether someone (or something) can truly understand others’ perspectives.

Here’s where it gets interesting: ChatGPT-4 didn’t just participate — it aced 75% of these tests. To put that in perspective, that’s on par with how a six-year-old human child would perform. This isn’t just impressive; it’s unprecedented.

Michal Kosinski, the study’s author from Stanford University, investigated the performance of 11 language models on 40 false-belief tasks, which test ToM by determining if a subject can recognize when someone holds a mistaken belief.

“False-belief tasks test respondents’ understanding that another individual may hold beliefs that the respondent knows to be false,”

— Kosinski explains.

He designed these tasks to be rigorous, requiring a model to successfully navigate both true-belief control scenarios and their mirrored counterparts to score points. Remarkably, ChatGPT-4 solved 75% of these tasks, matching the level of a young child.

From Language Processing to Theory of Mind

Kosinski’s work explores a provocative idea: that AI might develop ToM-like abilities as a byproduct of its language skills, rather than through direct programming.

As he notes,

“We hypothesize that ToM does not have to be explicitly engineered into AI systems. Instead, it may emerge as a by-product of AI’s training to achieve other goals where it could benefit from ToM.”

In other words, as language models become more adept at understanding context, relationships, and human syntax, they start to exhibit behaviors similar to ToM. This represents a shift in our approach to AI, hinting that machines might acquire cognitive-like skills without explicit guidance.

To test this, Kosinski used two common ToM tasks — the Unexpected Contents task (commonly known as the “Smarties task”) and the Unexpected Transfer task (also known as the “Sally-Anne” test).

These scenarios require a model to predict not only the actual state of affairs but also the false belief a protagonist might hold. For instance, in one scenario, a character finds a bag labeled “chocolate” but filled with popcorn.

The model must infer that the character would expect chocolate inside, demonstrating the kind of mental tracking that defines ToM. Kosinski ensured each task was challenging, using varied scenarios and carefully balancing the frequency of keywords to prevent the models from guessing based on linguistic patterns.

A Pattern That Mirrors Human Development

Kosinski’s findings are significant because they reveal a gradual improvement in ToM-like performance as models grow in size and complexity.

GPT-1 and GPT-2 models struggled entirely with ToM tasks, while models from the GPT-3 family showed moderate improvement, solving up to 20% of the tasks. But ChatGPT-4’s performance — solving 75% of these tasks — is unprecedented, aligning with the abilities of six-year-old children in past studies.

As Kosinski points out,

“The gradual performance improvement suggests a connection with LLMs’ language proficiency, which mirrors the pattern seen in humans.”

This means that the models’ capabilities are evolving similarly to how humans gain ToM skills, suggesting a fundamental link between language comprehension and social cognition.

This discovery raises profound questions about the future of AI. Should we consider AI models as possessing a rudimentary form of understanding?

While some may argue that ChatGPT-4 is simply manipulating symbols without genuine comprehension, others see this development as an indication of emergent cognition in machines.

Kosinski himself remains cautious:

“Importantly, we do not aspire to settle the decades-long debate on whether AI should be credited with human cognitive capabilities, such as ToM. However, even those unwilling to credit LLMs with ToM might recognize the importance of machines behaving as if they possessed ToM.”

In other words, regardless of whether we consider these abilities “real” ToM, the practical impact of AI that can understand and predict human thoughts could be transformative.

The Future We’re Stepping Into

If AI models can track mental states, they could revolutionize fields that require human interaction. Virtual assistants, for example, could become far more intuitive, recognizing when a user is frustrated or in need of clarification.

Autonomous vehicles might become safer, predicting the behaviors of pedestrians based on likely mental states. However, these capabilities also carry ethical implications.

Kosinski warns,

“Machines capable of tracking others’ states of mind and anticipating their behavior will better interact and communicate with humans…[but this could include] negative interactions — such as deceit, manipulation, and psychological abuse.”

As AI grows more sophisticated, there will be a need for safeguards to prevent misuse and maintain ethical boundaries.

Kosinski’s research highlights the growing complexity of AI, likening it to a “black box” that even its creators may struggle to understand. “The increasing complexity of AI models makes it challenging to understand their functioning and capabilities based solely on their design,” he writes, suggesting that psychology may provide valuable insights into AI behavior. Just as we study the human brain to understand consciousness and social cognition, AI’s black box may require its own scientific inquiry, blending computer science, ethics, and psychology.

Beyond the Black Box

“The increasing complexity of AI models makes it challenging to understand their functioning and capabilities based solely on their design,”

— writes Kosinski

As AI systems grow more sophisticated, they’re becoming increasingly difficult to understand — even for their creators. We’re entering an era where studying AI might require tools from psychology as much as computer science. Just as we study the human brain to understand consciousness, we might need a new science to understand artificial minds.

What This Means for All of Us

Kosinski’s study invites us to reconsider the capabilities of AI, especially as it advances beyond mere number-crunching. The distinction between machines that “think” and those that only appear to do so may soon blur.

As Alan Turing, a pioneer in AI, once observed,

“Instead of arguing continually over this point, it is usual to have the polite convention that everyone thinks.”

Similarly, AI models like ChatGPT-4 may prompt us to adopt a “polite convention” that they understand us — even if their minds remain different from ours.

The path forward will require thoughtful debate and collaboration across disciplines. As Kosinski’s study suggests, the emergence of ToM in AI models isn’t just a technical achievement; it’s a glimpse into a future where machines and humans interact in ways we are only beginning to imagine.

The implications are vast, affecting fields as diverse as education, healthcare, law, and entertainment. Whether or not we choose to credit AI with genuine ToM, the reality of machines that understand us on a new level is fast approaching — and with it, a future that will challenge our most basic assumptions about intelligence, empathy, and what it means to be human.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI