AI chatbots are becoming experts at changing people's minds
Research suggesting that AI chatbots can be highly effective at changing people’s minds has triggered concern over how they persuade and what that means for truth and democracy. Commenters note that chatbots combine an authoritative tone, endless patience, and tailored fact-heavy rhetoric—often trusted more than humans—yet can also hallucinate, cherry-pick, or reflect the biases of their creators. Many see this as a double-edged sword: powerful for debunking misinformation or explaining complex topics, but equally suited to scalable, targeted propaganda and subtle manipulation.
Effectiveness of Facts and Presentation
- Many are struck that “facts and evidence” seemed effective, but argue it’s really about framing, tone, and perceived legitimacy.
- Same facts can persuade or fail depending on style, authority, and clarity of explanation.
- Some note LLMs can flood users with examples, which feels convincing even when some are wrong or cherry‑picked.
Why Chatbots Can Change Minds
- Key advantage: they are not humans, so debates feel less like zero‑sum status contests. Changing your mind doesn’t feel like “losing” to another person.
- LLMs don’t tire, stay polite, answer every question, and often validate the user’s ego before nudging them.
- People often assume AI is neutral, rational, and authoritative, and may project themselves onto it, lowering defenses.
Training Methods and Model Behavior
- Discussion of RLHF and newer methods (RLAIF, RLVR, DPO): earlier models were overly agreeable; newer ones are more willing to correct or contradict users.
- Some users welcome pedantic correction as protection against sycophancy; others dislike when models override explicit instructions.
Risks: Misinformation, Manipulation, Democracy
- Models tuned for persuasiveness tend to become less truthful and hallucinate plausible “slop.”
- LLMs are described as “gish gallop machines” and a “DDOS on free resources,” able to overwhelm with volume and confidence.
- Concerns about propaganda at scale: hyper‑targeted persuasion, political manipulation, and erosion of democratic processes.
- Worry that people treat LLM outputs as more reliable than experts, yet the systems can confidently repeat outdated or biased consensus.
Study Design, Limits, and Counterpoints
- Skepticism about Prolific/online‑survey samples and whether results generalize; some think real‑world political identities are far less movable.
- Others argue persuasion is self‑selected: people who engage with a “convincing bot” are already somewhat open to changing their minds.
- Debate over whether releasing many diverse chatbots reduces manipulation or simply overwhelms users with more persuasive noise.