In Dan Brown’s Origin, Edmond Kirsch imagines humanity and technology merging. I have a smaller question: when that happens, will we notice?

Five numbers bring the question closer.

3 in 10TURING · 1950

Alan Turing replaces “can machines think?” with a more concrete game. A judge questions two participants in writing and tries to tell them apart. Then he imagines putting a machine in one of their places.

THE IMITATION GAME

JUDGE

PERSON

MACHINE

Written questions. Which answer comes from a machine?

Turing’s prediction was precise: after five minutes, the judge would identify the human correctly no more than 70% of the time. The machine might fool the judge about three times in ten. A prediction, not proof of consciousness.

73%GPT-4.5 · WITH A PERSONA

In a study published in PNAS in 2026, a person and an AI system conversed with a judge. When GPT-4.5 was prompted to play a convincing person, judges chose it as human in 73% of cases. Without that prompt: 36%.

That is the result of this experiment, under those conditions. It does not show that the machine thinks; it shows how an imitated personality can sway human judgment.

WHEN THE JUDGE PICKS THE MACHINE

Turing · prediction 30%

GPT-4.5 · with a persona 73%

Two different experiments, 76 years apart: an illustrative comparison, not identical conditions.

1966ELIZA

Joseph Weizenbaum created ELIZA, a program that simulated a psychotherapist. His secretary had watched him build it; after talking with it, she asked him to leave the room. In the 2026 PNAS test, ELIZA was chosen as human in 23% of cases.

99.8%CAPTCHAS

The “T” in CAPTCHA stands for Turing: here, the machine tries to distinguish us from bots. A 2023 study of distorted-text challenges reports human accuracy of 50–84%. The 99.8% bot figure comes from earlier research, not a new bot test by those authors.

A TEXT CAPTCHA, IN PIXELS

Sometimes it is the robot that has to recognise the human.

53%WEB TRAFFIC · 2025

Thales estimates that in 2025 bots generated 53% of observed web traffic: 40% classified as bad bots and 13% as benign automation. This reflects data observed by the company, not every website. “Bot” does not always mean “attack.”

The Turing test does not tell us whether a machine thinks. It points to a more everyday limit: how much can we trust our ability to recognise who we are talking to?

Perhaps the confusion will not begin when machines become just like us. It will begin when telling the difference no longer seems important.

One practical suggestion: choose a family password. If someone calls using the voice of a loved one and asks for something unusual, ask for it.

Have a good Sunday from A margine: a small curiosity to carry with you.

Giovanni

NOTES FOR THE VOYAGE

The Imitation Game

Ti è piaciuto?