Did Descartes think machines could ever think?

René Descartes, the French philosopher and mathematician, asked in 1637 a question that sounds strikingly modern: if someone built a machine that looked and moved exactly like a person, how could we tell it was not a real human being? His answer was that we could always tell, and he gave two tests to prove it.

The passage is only a page long, tucked into Part V of his Discourse on the Method. Yet philosophers and computer scientists keep returning to it, because it names the two abilities that still mark the line between clever machinery and real understanding: flexible language and general-purpose reasoning.

When and where it was published

The Discourse on the Method, whose full title is "Discourse on the Method of Rightly Conducting One's Reason and of Seeking Truth in the Sciences," was published in 1637 in Leiden, in the Netherlands. Descartes wrote it in French rather than Latin, the usual language of scholarly books. It served as an introduction to three scientific essays on optics, weather and geometry, and a Latin translation followed in 1656.

The book is best known for a line in Part IV, "I think, therefore I am." The section on machines comes in Part V, where Descartes explains his views on the body, the heart and the difference between humans and animals.

The body as a machine

Descartes starts from a bold claim. The human body, he says, can be understood as a machine made by God, far better arranged than any machine people can build. He points to the automata of his day, mechanical figures that moved by themselves, and notes how much they achieve with very few parts compared with the bones, muscles and nerves of a living body.

From there he makes a surprising concession. If someone built a machine that looked and behaved exactly like an ape or any other animal without reason, we would have no way to tell it from the real animal. For Descartes, animals are, in effect, natural machines.

Humans are different. A human-shaped machine, he argues, could always be exposed by two tests.

The two tests

Here are the tests in plain terms, with what Descartes thought a machine could and could not do:

Test	What a machine could do	What it could never do
1. Language	Utter words; even say something when touched in a certain place, or cry out that it is hurt	Arrange words in new ways to give a fitting reply to whatever is said in its presence
2. General reason	Perform some tasks as well as people, or even better	Act sensibly across all the situations of life

His reasoning for the second test is the most interesting part. Each part of a machine, he says, must be specially arranged for each particular action. Human reason, by contrast, is "an universal instrument that is alike available on every occasion." A machine would need an impossible number of special parts to handle everything life throws at it.

He adds two observations. Parrots and magpies can pronounce words, yet they do not show that they understand what they say, while deaf people invent signs of their own to share their thoughts. And a clock made only of wheels and weights keeps time more accurately than any person, which shows that being better at one task does not prove a mind.

Try it: Ask a voice assistant three questions: one simple fact, one follow-up that depends on the first answer, and one odd question it is unlikely to have been prepared for. Note where it answers fittingly and where it falls back on a stock reply. You are running Descartes' first test.

Why it matters: a test before Turing

In 1950 the mathematician Alan Turing proposed his famous "imitation game," now called the Turing test: if a machine can hold a written conversation that a judge cannot tell apart from a person's, we should take seriously the idea that it thinks. Philosophers have long pointed out that Descartes got there first, in a sense. The Stanford Encyclopedia of Philosophy notes that the Turing test is sometimes said to be prefigured in the Discourse, and a paper by the philosopher Gunderson made the connection as early as 1964.

The two men used the same yardstick, flexible conversation, but drew opposite conclusions. Descartes was sure no machine could ever pass. Turing set aside the bare question "can machines think?" in favor of a practical game, and he expected digital computers to do well at it before too long.

How do Descartes' two tests hold up against AI today?

His first test is under real pressure in 2019. Voice assistants answer spoken questions every day, and last month OpenAI described GPT-2, a language model that can write surprisingly fluent paragraphs from a short prompt. Fluency of that kind would have astonished Descartes. Still, these systems often lose track of a conversation, repeat themselves or state things that make no sense, so whether they truly give a fitting reply to whatever is said remains a fair and open question.

His second test holds up better. Today's AI is impressive in narrow areas: game programs beat world champions, image systems recognize faces, translation tools handle dozens of languages. But each system is built or trained for its own task. A program that plays Go cannot fold laundry or plan a trip. That is very close to Descartes' point that machines need a special arrangement for each action, while human reason works everywhere. Researchers call the missing piece general intelligence, and it remains one of the field's biggest unsolved goals.

For a beginner, Descartes offers two good questions to ask of any new AI system: can it reply sensibly to something truly unexpected, and can it handle a task it was never built for? Those questions are nearly four hundred years old, and they are still the right ones.