ChatGPT can’t think

There has been shock around the world at the rapid rate of progress with ChatGPT and other artificial intelligence (AI) created with large language models (LLMs).

These systems can produce text that seems to display thought, understanding and even creativity.

But can these systems really think and understand? This is not a question that can be answered through technological advancement, but careful philosophical analysis and argument tells us the answer is no.

And without working through these philosophical issues, we will never fully comprehend the dangers and benefits of the AI revolution.

In 1950, the father of modern computing, Alan Turing, published a paper which laid out a way of determining whether a computer thinks.

This is now called ‘the Turing test’.

Turing imagined a human being engaged in conversation with two interlocutors hidden from view: one another human being, the other a computer.

The game is to work out which is which.

If a computer can fool 70% of judges in a five-minute conversation into thinking it’s a person, the computer passes the test.

Would passing the Turing test show that an AI has achieved thought and understanding?

CHESS CHALLENGE

Turing dismissed this question as hopelessly vague, and replaced it with a pragmatic definition of thought, whereby ‘to think’ just means passing the test.

Turing was wrong, however, when he said the only clear notion of understanding is the purely behavioural one of passing his test.

Although this way of thinking now dominates cognitive science, there is also a clear, everyday notion of understanding that’s tied to consciousness.

To understand in this sense is to consciously grasp some truth about reality.

In 1997, the ‘Deep Blue’ AI beat chess grandmaster Garry Kasparov. On a purely behavioural conception of understanding, Deep Blue had knowledge of chess strategy that surpasses any human being.

But it was not conscious: It didn’t have any feelings or experiences.

Humans consciously understand the rules of chess and the rationale of a strategy.

Deep Blue, in contrast, was an unfeeling mechanism that had been trained to perform well at the game.

Likewise, ChatGPT is an unfeeling mechanism that has been trained on huge amounts of human-made data to generate content that seems to be written by a person.

It doesn’t consciously understand the meaning of the words it’s spitting out.

If ‘thought’ means the act of conscious reflection, then ChatGPT has no thoughts about anything.

TIME TO PAY UP

In the 1990s, neuroscientist Christof Koch bet philosopher David Chalmers a case of fine wine that scientists would have entirely pinned down the ‘neural correlates of consciousness’ in 25 years.

By this, he meant they would have identified the forms of brain activity necessary and sufficient for conscious experience.

It’s about time Koch paid up, as there is zero consensus that this has happened.

This is because consciousness can’t be observed by looking inside your head.

But there are multiple ways of interpreting the data.

Some scientists believe there is a close connection between consciousness and reflective cognition – the brain’s ability to access and use information to make decisions.

This leads them to think that the brain’s prefrontal cortex is essentially involved in all conscious experience.

Others deny this.

Scientists have good understanding of the brain’s basic chemistry.

We have also made progress in understanding the high-level functions of various bits of the brain.

But we are almost clueless about the bit in-between – how the high-level functioning of the brain is realised at cellular level.

PAUSE IN DEVELOPMENT

As I argue in my forthcoming book ‘Why? The Purpose of the Universe’, consciousness must have evolved because it made a behavioural difference.

Systems with consciousness survive better than systems without consciousness.

If all behaviour was determined by underlying chemistry and physics, natural selection would have no motivation for making organisms conscious – we would have evolved as unfeeling survival mechanisms.

My bet, then, is that as we learn more about the brain’s detailed workings, we will precisely identify which areas of the brain embody consciousness.

This is because those regions will exhibit behaviour that can’t be explained by currently known chemistry and physics.

While the processing of LLMs is too complex for us to fully understand, we know it could in principle be predicted from known physics. On this basis, we can confidently assert that ChatGPT is not conscious.

There are many dangers posed by AI, and I fully support the recent call by tens of thousands of people, including tech leaders Steve Wozniak and Elon Musk, to pause development to address safety concerns.

The potential for fraud, for example, is immense.

LLMs aren’t intelligent.

They are systems trained to give the outward appearance of human intelligence.

Scary, but not that scary.

• Philip Goff is an associate professor of philosophy at Durham University.

– This article first appeared on The Conversation.