In the era of generative AI, distinguishing truth from fiction is becoming increasingly difficult. Deepfakes — not only videos, but also text-based news created by neural networks — have flooded the information space. To understand who can help users recognize such content, researchers conducted an unusual experiment: they compared how much people trust advice from AI assistants, their peers, and linguistic experts when searching for fakes.
What the experiment showed
In the laboratory assignment, students had to determine what proportion of text in synthetic news was written by a human rather than generated by AI. As hints, they were offered three sources: ChatGPT (based on GPT-4), advice from other participants, and recommendations from linguistic experts. The results showed that participants followed advice from ChatGPT more often than advice from peers when dealing with deepfake news created using GPT-2. At the same time, linguistic experts were trusted more than other people, but the relative trust in experts versus ChatGPT varied from one wave of the experiment to another.
Interestingly, in an additional round conducted in 2025 using the same methodology, students trusted experts more than ChatGPT. The authors attribute this to a shift in perceptions of AI detection: perhaps people have come to better understand the limitations of neural networks and value human expertise more.
Why does AI inspire more trust than people?
At first glance, it seems paradoxical that people are more willing to rely on an algorithm that may itself be a source of fakes than on live interlocutors. One possible explanation is the perception of AI as an impartial analyst devoid of subjective motives. Peers, on the other hand, may raise doubts about competence, and their opinions are easy to question.
Importantly, the improvement in deepfake detection accuracy did not happen by itself, but through a combination of trust in the advice and its quality. In other words, when participants relied on high-quality advice, their results were noticeably better. This highlights that blind faith in any source is not the answer.
Conclusions: generative AI — both a threat and a defense
The main conclusion of the study is that AI can and should be used to combat content created by AI itself, but with a caveat. Relying on neural-network detectors is justified only if the tools themselves are accurate enough. Otherwise, trusting bad advice will only make the situation worse.
The authors of the study remind us of the dual role of generative models: they simultaneously create deepfakes and help detect them. Therefore, the future most likely lies in hybrid approaches, where AI tools work alongside experts, while users maintain healthy skepticism.
