AI and Democracy: Who Shapes Public Opinion?

Before generative artificial intelligence had any notable effect, researchers at the Massachusetts Institute of Technology traced about 126,000 news cascades on Twitter between 2006 and 2017, shared by roughly three million people. They found that false stories were about 70% more likely to be retweeted, and that a true story took about six times as long to reach 1,500 people. Most importantly, automated accounts were not the main cause: the gap remained when the researchers removed them (Vosoughi et al., 2018). People do the sharing, and platforms amplify it.

When manufacture becomes cheap

Misinformation once needed material to excerpt or distort. Generative systems produce text, voice and image at little cost. In its Global Risks Report 2026, published on 14 January 2026, the World Economic Forum ranked misinformation and disinformation second among the risks expected over a two-year horizon. That ranking reflects the perceptions of those surveyed by the Forum, not a direct measurement of harm, and should be read accordingly.

Perhaps the most lasting effect is what the legal scholars Robert Chesney and Danielle Citron called the "liar's dividend": the existence of fakes becomes an excuse to deny a genuine recording as fabricated, so that a suspicious public is primed to doubt honest evidence (2019). We discussed this in the article "Deepfakes: When Visual Evidence Becomes Doubtful".

Is the message tailored to each person stronger?

There is a fear that machines will address each voter with what they know about them, and the evidence here is mixed. In an experiment with 820 participants in short debates, the odds of a participant's agreement increasing were about 82% higher when GPT-4, given basic demographic information about them, debated them, compared with a human opponent (Salvi et al., 2025). By contrast, a study by Hackenburg and Margetts in PNAS in 2024 found that GPT-4 messages were persuasive in general, but that tailoring them to the recipient's political and demographic traits did not increase their effect by a statistically significant amount over general messages. Differences in design and topics prevent a firm conclusion, and the safest statement is that machines' ability to persuade is consistent across experiments, while the advantage of individual targeting over a general message has not been settled.

The machine as mediator, not adversary

The same technology can be used in the other direction. Researchers at Google DeepMind and elsewhere tested a system called the "Habermas Machine" on about 5,700 British adults: it drafts a statement reflecting the views of a small group, then revises it in light of their objections. Participants preferred the machine's statements to those of human mediators 56% of the time, and agreement within groups rose by about 8 percentage points (Tessler et al., 2024). The authors point to limits: the system does not check facts or moderate the discussion, it was tested in Britain and on small groups, and endorsement may reflect trust in a fair procedure rather than genuine agreement. So there is hope, and it depends on good design.

What should be done

• Transparency of ranking and origin: people should know why content appeared to them, and synthetic content should be labelled.
• Verification before sharing: care over news is a well-known principle of Islamic ethics and a habit to be raised at home and at school (see the article "The End of the Age of Rote Instruction?").
• Sound governance: one that protects public processes from fraud without restricting legitimate expression, and suits the nature of each society.
• Adjusting incentives: any system that rewards sensation will produce sensation, and changing what it rewards changes its behaviour.

Societies differ in the forms that public participation takes, but their need is the same: truthful news, debate that preserves respect, and confidence that the person you are speaking to is a human being and that your word is heard. What manufactured voices weaken most is that trust; if each of us takes his opponent for a machine or a hireling, dialogue is lost. Keeping speech truthful is part of the trust of the word, and God is the one whose help is sought.