AgoraAn open square of thought

Debate

Can Machines Think or Feel?

Language models now hold conversations that people often cannot tell from a human’s. Does that mean they understand anything – and could they ever feel, or only make us believe they do?

Large language models now write essays, code and poems, and many people talk to them every day as if to a person. In a 2025 study by Cameron Jones and Benjamin Bergen, people who chatted for five minutes with a human and with GPT-4.5 at once took the machine for the human 73 per cent of the time. Turing’s test, in this form, is passed; what that shows is the open question. If such systems could feel, we could wrong them; if they only seem to, our tendency to see minds everywhere misleads us. Behind this lies the unresolved question of what consciousness is – and beside it a debate about what these systems do to judgement, work and attention.

Open in the star map

Positions

Simulation is not understanding

A computer program manipulates symbols by their form alone: it has syntax but no semantics. Searle’s Chinese Room is meant to show that following such rules can produce convincing answers in a language without understanding a word of it. Thinking, on this view, requires causal powers at least equivalent to those of brains; running the right program is never enough by itself – whatever the hardware. Emily Bender and Alexander Koller (2020) make a related point about language models: a system trained only on the form of language has no access to what its words are about.

Held by

Mind is what a system does

Functionalists hold that mental states are defined by their role – what causes them and what they cause – not by the material they are made of, an idea Putnam introduced in the 1960s and later came to doubt. If so, a machine organised in the right way could in principle think, as Turing had already suggested. Dennett adds that consciousness is not an extra inner glow but a bundle of functions that can be explained – while warning that systems built to imitate people can deceive us long before they have minds.

Held by

The hard problem remains open

Even a complete account of information processing seems to leave out what it is like to have an experience – Nagel’s point about the bat. Chalmers calls this the hard problem and, unlike Searle, does not tie consciousness to brain-like causal powers, so he treats machine consciousness as a serious open question. In 2023 he judged it unlikely for the language models of the time, but a real possibility for their successors. A report by Patrick Butlin, Robert Long and colleagues (2023) derived indicator properties from scientific theories of consciousness and found that no current system had them all, while seeing no obvious technical barrier to building one that does.

Held by

Minds that extend

Clark and Chalmers argued in 1998 that thinking does not stop at the skull: a notebook that someone consistently relies on can be part of their memory. On this view the more interesting question is not whether a machine thinks on its own, but how humans and machines form hybrid thinking systems. Clark sees generative AI as the latest such extension – promising, but demanding new skills in knowing when to trust it.

Held by

Act responsibly under uncertainty

These thinkers disagree about what the machines are, but agree that the ethical questions cannot wait for metaphysics. Metzinger argues that because we cannot rule out creating artificial beings that suffer, there should be a global moratorium until 2050 on research that aims at or risks synthetic consciousness. Floridi describes today’s systems as “agency without intelligence” – able to act successfully without understanding anything – and Vallor sees them as mirrors of past human data, warning that the greater danger is that we lose trust in our own judgement and virtues.

Held by

Read more

As of 10/2026. Summaries for orientation – the sources give the full arguments. · How Agora is made

Open in the star map