AI That Mimics Trump's Speeches Fooled Most People in a New Test
In a recent experiment, an artificial intelligence system trained on President Donald Trump's speeches managed to fool a majority of human judges. The test, conducted by the legal blog Lawsuit.org, asked 1,000 participants to distinguish between actual Trump quotes and text generated by a custom AI model dubbed “RoboTrump.” The result: participants guessed incorrectly 60 percent of the time, a rate worse than random chance.
The AI system used in the test was CTRL, a text-generation model developed by Salesforce Research and unveiled in September. The Lawsuit.org team fed the model a large corpus of Trump's speeches and public statements, then produced a series of paragraphs in the president's distinctive style. They then presented these alongside genuine quotes to a panel of 1,000 volunteers.
Participants were asked to identify which of two paragraphs was actually spoken by Trump. The first example in the quiz was a rambling, self-referential paragraph about winning the 2020 election. The second was a more policy-focused paragraph about Democrats and health care. The correct answer was the second one, but many respondents got it wrong.
The study also revealed demographic patterns in how people answered. Self-identified conservatives were more likely than liberals to attribute a paragraph to the real Trump, whether it was genuine or AI-generated. Younger participants were also more inclined than older ones to believe that a given paragraph came from the president.
Researchers noted that Trump's characteristic speaking style — often meandering and repetitive — may have made the task especially difficult. As they put it, “Trump's rambling style probably makes differentiating between real and fake more difficult than it would be for a more eloquent and talented speaker.”
Concerns About AI and the 2020 Election
After the quiz, participants were asked about their views on AI's potential impact on the 2020 election. Forty-three percent said they were “more concerned” about the technology's influence than they had been before taking the quiz. That shift suggests that hands-on exposure to convincing AI-generated text can heighten public awareness of disinformation risks.
The experiment is part of a broader conversation about the dangers of AI-generated content. Earlier reports have warned that such tools could be used to produce a “massive tsunami” of fake news, potentially overwhelming voters with false or misleading information during election cycles.
While the RoboTrump model is not the first to mimic a public figure's style, it stands out for its demonstrated ability to deceive a large audience. The findings highlight the need for media literacy and technological safeguards as AI-generated text becomes more sophisticated.
