Research: AI becomes indistinguishable from humans

For a long time, the Turing test remained more of a theoretical boundary between man and machine. But a new study by scientists at the University of California, San Diego shows that modern language models have already moved into the zone where people regularly make mistakes trying to understand who is in front of them – a human or an AI.

The main participant of the experiment was the GPT-4.5 model. In a series of text dialogues, she managed to convince the participants of her “humanity” in 73% of cases. In essence, this means that the chatbot was more often perceived as a real person than real interlocutors in the same conditions.

How the experiment was conducted

The researchers used a complicated version of the classic Turing test. Participants were simultaneously shown two dialogues — with a human and an AI — without revealing who is who. After a short conversation, they had to determine where the real person was.

Unlike earlier generations of chatbots, modern models do not pretend to be boilerplate responses. They adjust the language style, use Internet slang, imitate emotions, add humor, pauses and even slight uncertainty – all that people subconsciously read as “live” communication.

The best result was shown by GPT-4.5 in the persona prompt mode – when the model is given a certain character. In this case, she imitated a young, introverted user familiar with Internet culture and the modern style of correspondence. This presentation was the most convincing.

For comparison, LLaMa-3.1-405B received 56% “human” ratings, GPT-4o – 21%, and the classic ELIZA of the 1960s – 23%. Without special behavior tuning, GPT-4.5 showed a noticeably more modest result — 36%.

Why do people think AI is human

An interesting conclusion of the study is that the participants did not evaluate the intelligence of the interlocutor, but the communication style. The decision was influenced by:

• manner of formulating thoughts;
• emotionality;
• sense of humor;
• naturalness of dialogue;
• social signals and “human” errors.

These factors turned out to be more important than the content of the answers. And modern models have learned to reproduce them especially well.

It is also important that the experiment was conducted exclusively in text format – without voice, video or visual prompts. This increased the importance of speech behavior, where AI is already almost inferior to humans.

This is not intelligence, but a risk

The authors emphasize that passing the Turing test means that the model has consciousness or understanding. GPT-4.5 does not “think” like a person – it only very accurately reproduces the human style of communication.

However, the practical consequences are obvious.

If AI is able to convincingly imitate a person in a short dialogue, it can affect many areas:

• support services;
• social networks;
• dating programs;
• educational platforms;
• political communications;
• fraudulent schemes.

The problem is that people often build trust in just a few messages, and now this process can be easily simulated by an algorithm.

The Turing Test is losing its former meaning

The Turing test was originally intended as a test of a machine’s ability to imitate human thinking. Today, the situation has changed: the question is increasingly being asked differently — whether a person can distinguish a car from another person.

And if GPT-4.5 already copes with this better than humans, the line between “real” and “artificial” in digital communication is becoming less and less visible.