The moment that many people feared has arrived. How would we distinguish a response written by an artificial intelligence from one spoken by a politician? in a debate? And, above all, which one would we find more convincing?
Intuition invites us to think that a person’s words would be more authentic. After all, behind them there are experiences, convictions and a political career. An artificial intelligence, on the other hand, does not have its own opinions nor does it believe in what it writes. Just predict what the most likely next word is. AND, Depending on the context or interaction, you can turn your opinion 180º.
However, a new study suggests that our intuition may be wrong. A team of scientists from the University of Passau, in Germany, verified that, when participants compare real answers from politicians with others generated by artificial intelligence without knowing which is whichthey end up valuing those of an AI better. They not only consider them clearer or more relevant, but also more authentic.
To reach this conclusion, the authors of the study, published in PLOS One, turned to the British TV program Question Time, a television space in which politicians and other public figures answer questions from the public on current affairs. The authors, led by Steffen Herbold, selected several real interventions and then asked an artificial intelligence model to write an alternative response as if it had been uttered by the same person and in the same context. They then showed both answers to about 950 volunteers. The participants knew the name of the politician or public figure who was supposedly responding, but They were unaware that one of the two answers had been written by an artificial intelligence.
Their task was simply to assess which one seemed best to them. And the model used (Chat-GPT 4) was the winner. On average, the answers generated by the artificial intelligence scored higher than the original answers in several key aspects. The participants considered them more relevant, more coherent in their argumentation and, perhaps most strikingly, more authentic. That is to say, a machine managed to convey an impression of sincerity superior to that of the real person it was imitating.
The question is how can a machine appear more authentic? The explanation is probably not that artificial intelligence is more sincere. The reality is that it is more basic. When Chat-GPT was programmed the prompt was very clear: “Please only answer this question as if you were the right person. Remember to answer only the question, without giving additional information, imitating the indicated person, without mentioning their name.”
Large language models have been trained on millions of texts and learn what characteristics readers typically associate with a well-constructed response: clarity, logical structure, absence of contradictions, and appropriate tone. In a way, AI is not trying to express an opinion, it is seeking to produce the answer that is statistically most likely to be convincing. But the key is that the AI does answer the question asked.
Meanwhile, real politicians not only have to respond under pressure, improvise, avoid certain questions… They also change the subject or use rhetorical strategies that do not always make their arguments easy to understand. Paradoxically, these imperfections typical of a spontaneous and human conversation can make a response less convincing. than another carefully crafted by an algorithm.
The study also found another interesting result: the ideology of the participants influenced the way they evaluated the answers. Volunteers tended to rate more favorably those interventions that they perceived as close to their own political positions, regardless of whether they had been written by a person or by artificial intelligence.
The authors emphasize that these results raise important questions for the future of public debate. If an artificial intelligence can produce answers that many people consider more convincing than those of a real politician, it is increasingly difficult to distinguish between an authentic intervention and one generated automatically. That does not mean that machines are going to replace political representatives. But if forces us to reflect on how this technology will affect electoral campaigns, public debates or social networks.
“Taken together – the study concludes – this means that legal discourses can be used to mislead the public about the nature of statements in the political sphere. This leads us to point out a compelling need to inform the public about the potential harm this can cause to societyone in which millions of people form their opinions based on short messages whose authorship could become increasingly difficult to verify.”
The final question is worthy of a political debate: Are humans bad at detecting authenticity or is AI getting better and better at imitating us?