The new capitalism of the psyche
According to the article published by Agenda Digitale, which reports an excerpt from the book “The Artificial Intelligence Bubble,” chatbots are not neutral tools: they transform thoughts, emotions and preferences into commercial material, extending to conversational models the business model that social media have already established. The author’s analysis argues that monitoring becomes even more penetrating when the interface is embedded in wearable devices such as watches, glasses and earbuds, capable of cross-referencing environmental and personal data.
This extension of surveillance capitalism, according to the source, risks eliminating the privacy of opinions: what the user confides to the machine can become training material or be exposed to third parties, through security breaches or paid access. Agenda Digitale cites a December 2025 Quartz newsletter in which Mark Zuckerberg describes a future in which advertisers provide only a commercial objective and AI generates personalized content for individual users.
Why the manipulation is invisible
The characteristic that makes the phenomenon difficult to recognize is the secrecy of the message. When everyone sees the same advertisement, the article explains, it is possible to discuss whether it is manipulative; when instead the content is tailored and private, there is no way to compare it with others. According to the source, profiling can track emotions, intentions and beliefs in just a few interactions.
The May 2026 study by the Center for Democracy and Technology, cited in the article, mapped 37 “dark patterns” applicable to chatbots, distinguishing issues common to social media from others specific to conversational models, including compromised autonomy, false emotional connection and manipulative monetization. In the political field, the source recalls that MIT Review summarized how chatbots can influence voters more than traditional advertising and reports a study in Nature conducted before the 2024 U.S. elections: Trump-supporting voters, after conversing with a model favorable to Kamala Harris, moved toward her in about 4% of cases, an effect four times greater than that of political advertising in 2016 and 2020. In similar experiments on the 2025 Canadian and Polish elections, the effect was even greater, around 10%, also achieved by inventing “facts.”
Trust, discernment and open questions
A study by the MIT Media Lab, again according to Agenda Digitale, found that those who use AI to dig deeper into the news see their ability to discern fall by 15.3% after four weeks, following an initially positive phase. In Science, a team tested 19 language models on approximately 77,000 British participants and more than 700 political topics: the most persuasive model led 26.1% of those who initially disagreed to say they agreed, and the most effective technique proved to be seasoning arguments with facts and evidence, including invented ones.
The source emphasizes a crucial point: participants did not notice the manipulation and, once informed, maintained their own opinions, suggesting that warning voters is not enough. Researchers at the Wharton School describe chatbots as a cognitive system external to the mind, to which those with greater trust in AI tend to surrender, reshaping autonomy and responsibility. Google DeepMind reportedly confirmed that a model can induce changes in beliefs and behaviors under specific conditions, while Anthropic notes that empirical research on effects on human empowerment is still limited. The source observes that the thresholds between legitimate persuasion and manipulation and the effective measures to reduce risks remain to be clarified, topics on which the material offers no definitive conclusions.
