Intuition in online dating no longer works

Illustrative image generated by Gemini

Scientists conducted a rigorous version of the classic Turing test and came to alarming conclusions. Modern artificial intelligence has not only learned to impersonate a human — it does so so convincingly that real people seem like bots. We tell you how the machine manages to deceive us and why.

Imagine you are chatting online with two strangers: one of them is a real person, the other is a robot. Your task is to compare the two conversations and figure out who is who. This is exactly what the classic Turing test looks like in a three-participant format, which researchers from the University of California proposed to modern language models. The study was published in the journal Psychological and Cognitive Sciences.

The result of the experiment is interesting. When artificial intelligence (GPT-4.5 model) was instructed to behave like an ordinary person, participants chose the machine in 73% of cases, while only 27% believed the real person sitting on the other side of the screen.

The main secret of artificial intelligence's success is not its phenomenal intelligence or encyclopedic knowledge, but a successful persona. When the program was given only the basic instruction "be human," it often failed. AI was recognized in 64% of cases.

But when it was given a detailed role of a young, slightly introverted person who knows internet culture well and uses slang, the results were turned upside down. It turned out that it is precisely the informal style and the impression that the interlocutor doesn't know something or evades a direct answer that makes the robot indistinguishable from us.

The most interesting thing about this study is how exactly people tried to expose the machine. Almost no one asked complex mathematical problems or logical puzzles traditionally associated with checking robots; such questions were only heard in 12% of games.

Instead, people relied on "small talk": asking about daily life, joking, and trying to feel emotional warmth. Judges made decisions based on writing style and their own intuition. One participant explained his choice in favor of the bot by saying, "it had a human vibe."

When researchers decided to complicate the task and tripled the conversation time, the results of the experiment became even more indicative. Participants were forbidden to end the game prematurely, so they were forced to communicate for a full 15 minutes, sending an average of 22 messages to each interlocutor — almost three times more text than during the short 5-minute test.

Since the previous version of the neural network was outdated and no longer supported by OpenAI developers, in the long test, scientists used the new GPT-5 model, as well as the LLaMa-3.1-405B model. They were again instructed to respond in the style of a young person, and both systems successfully passed the test.

GPT-5 convinced the judges that it was a living human in 59% of cases. The LLaMa model was believed slightly less, but still an excellent result — 56% of games. This means that in most cases, the machine again outperformed the real human participant.

Researchers meticulously checked all demographic data of 205 participants. It turned out that the ability to recognize artificial intelligence does not correlate with life background. Neither age, nor gender, nor higher education, nor even whether you consider yourself an expert in neural networks and how often you use chatbots in daily life — none of this made people more accurate in their decisions.

The only thing people recognized flawlessly was ELIZA — a very old and primitive algorithm from the 1960s, which works according to pre-written rules. It was added to control the experiment, and in the 15-minute test, people successfully identified this program in 83% of cases.

Scientists are sounding the alarm and introducing the term "counterfeit people." Just as fake money devalues real currency, the simulation of human communication can gradually undermine the value of real contacts and destroy trust in any person on the internet.

Systems capable of imitating humans so accurately can be an ideal weapon for manipulation, spreading disinformation, and social engineering, because we won't even realize that an algorithm is controlling us.

We are used to thinking that pure and fast calculations are characteristic of computers, while empathy, humor, spontaneity, and imperfection are our human monopoly. But now machines brilliantly imitate these too.

Perhaps the most unpleasant and disturbing detail of this experiment is not even that a robot is capable of passing the famous Turing test. The main problem is that a real living human, who was simply himself and tried to seem natural, lost this test.

• 4
• 0
• 2
• 1
• 6
• 2
• 0
• 0