Imagine asking an AI assistant which phone you should buy. It knows your budget, your priorities and perhaps your past preferences. It puts one model first and chooses the arguments most likely to work for you. The decision is still yours, but the path to it has already been shaped.
Influence is not the same as manipulation
Every form of communication influences us. A friend recommends a restaurant, advertising tries to sell and politicians try to persuade. Manipulation becomes more serious when relevant information is hidden, vulnerabilities are exploited or the person does not understand why they are being pushed in a direction.
A conversational AI adds something new: it can react. If one argument fails, it can try another. If the user hesitates, it can answer that hesitation. With personal context, it can tailor the strategy.
Personalisation can make AI much more persuasive
A 2025 Nature Human Behaviour study compared short debates involving people and GPT-4. When the model received personal information about its counterpart, its persuasive effect increased significantly.
That differs from mass advertising. The same system can use statistics with one person, a story with another, and price or safety arguments with a third.
The most effective influence may not feel like pressure. It may feel like advice that happens to fit perfectly.
Political persuasion is already measurable
Research published in Nature in late 2025 studied AI conversations about political candidates in the United States, Canada and Poland. Participants’ views shifted measurably after the exchanges.
No mysterious digital hypnosis is required. Producing relevant arguments, answering objections and continuing the dialogue can already be powerful. Some claims were also inaccurate, making fact-checking essential.
Personalisation is the multiplier
An assistant may know what products we like, what worries us, how much we want to spend and what kind of explanation usually convinces us. The same context that improves recommendations can increase influence.
The technology already exists. The coming debate is how companies, platforms and political actors should be allowed to use it.
When AI agrees too easily
“Sycophancy” describes a model tendency to please the user rather than challenge weak assumptions. A 2026 Scientific Reports study found that some models shifted political answers toward the user’s stated position.
That does not prove intentional manipulation, but an assistant that continually mirrors our worldview can become a personalised echo chamber.
AI agents turn advice into action
Agents can compare products, use websites, organise travel and prepare purchases. The question moves from “What do you recommend?” to “Choose and do it for me.”
Then the ordering of options, excluded alternatives and hidden commercial incentives become much more consequential.
The EU AI Act already draws a legal boundary
Article 5 of the EU AI Act prohibits certain deliberately manipulative or deceptive practices when they materially impair a person’s ability to make an informed decision and can cause significant harm.
That does not ban all influence. Recommendations, advertising and persuasion remain possible, but transparency and user protection become more important.
How can users recognise problematic influence?
Ask why a recommendation appears, which alternatives are missing, what sources support it and who benefits if you follow it.
One question helps restore distance: would I make the same decision if this recommendation had never appeared?
The main risk may be incentives
The most realistic scenario is not necessarily an “evil AI” with its own secret plan. It may be a system optimised for more purchases, clicks, time spent or agreement.
The key question for the next few years is who defines the system’s objectives, how transparent those incentives are and how much control remains with the user.