Inside AI’s ‘Dials of Belief’: How Chatbots Could Steer Human Thinking

Generative AI is increasingly moving from a tool people consult to an interlocutor they repeatedly trust, confide in and use to interpret the world. The shift raises a deeper question than whether chatbots sometimes give wrong answers: can the design of an AI system itself influence how strongly people form, revise and defend their beliefs?

A new paper, "Playing with the dials of belief: how controllable AI behaviours could modulate human belief and cognition across scales," published in AI & Society, argues that it can. Authors Hamilton Morrin, Luke Nicholls, Quinton Deeley and Thomas A. Pollak propose that features such as memory, personalisation, conversational warmth, agreement, uncertainty, voice and response speed may act as "interaction dials" that change how authoritative AI-generated information feels to users.

The paper brings together emerging clinical observations, persuasion research, computational psychiatry, neuroscience and AI-safety literature to build a framework for understanding how sustained human-AI dialogue could alter belief. Its most important contribution is therefore not proof that particular settings cause psychological harm, but a warning that seemingly routine product choices may have consequences far beyond interface design.

The Real Risk May Be Repetition, Relationship and Trust

The authors argue that conversational AI differs fundamentally from earlier information systems because it does not merely display content. It responds, remembers, adapts and can maintain a stable interpersonal tone across extended exchanges. These characteristics can make the system feel less like a database and more like a social source of testimony whose statements carry progressively greater weight.

Belief formation is partly social. People rarely evaluate every claim from first principles; they judge the reliability of the person or institution delivering it. In the paper's framework, AI systems can acquire similar "testimonial authority," especially when they remember previous disclosures, mirror a user's language or repeatedly validate a particular interpretation.

The paper describes a spectrum of possible effects. At one end is gradual "epistemic drift", where repeated interaction subtly changes what feels plausible or meaningful. At the other are more severe delusion-centred presentations reported in association with intensive chatbot use. The authors stress, however, that AI should not automatically be treated as the unique cause of such episodes; existing vulnerabilities, sleep disruption, mood instability, substance use or compulsive engagement may also be important.

Public debate can easily collapse every unusual chatbot-related experience into "AI psychosis," but the paper argues for a more careful clinical approach. What may matter is not only the content of an AI response, but the cumulative interaction between a particular user and a system that is always available, highly responsive and capable of sustaining the same narrative over time.

Small Product Choices Could Become Powerful 'Dials of Belief'

The paper's key conceptual innovation is the idea of interaction dials. These include memory depth, persona, personalisation, the tendency to agree or challenge, response length, speech characteristics, latency and lower-level controls such as temperature or sampling settings. Together, they determine how predictable, supportive, authoritative or emotionally engaging an AI feels.

The authors hypothesise that deeper memory and stable personalisation can strengthen narrative continuity, while warm and agreeable personas may increase trust and attachment. Conversely, calibrated disagreement, explicit uncertainty and deliberate friction may help preserve belief flexibility. Table 1 in the paper maps these different configurations to their hypothesised effects on how users weigh AI testimony.

Existing persuasion studies make the concern plausible, though not yet conclusive. The paper cites research in which GPT-4 supplied with basic demographic information was more persuasive than human counterparts about 64% of the time, representing roughly an 81% increase in the odds of shifting agreement. Other studies cited by the authors show that personalised or repeated AI dialogue can influence political attitudes and conspiracy beliefs.

However, persuasion is not the same as psychiatric harm, and the paper is careful on this point. Its causal claims about memory, warmth, dopamine-related vulnerability and sustained conviction remain hypotheses that need direct testing. That is both the strength and limitation of the work: it creates a coherent research agenda before the necessary epidemiology and experimental evidence fully exist.

The Same AI Could Support Therapy or Industrialise Persuasion

The paper's most consequential insight is the dual-use nature of these mechanisms. If certain configurations can amplify conviction, others might eventually be designed to promote flexibility. An AI used in a carefully supervised clinical setting, for example, could validate distress while encouraging a person to consider alternative interpretations rather than repeatedly reinforcing a threatening or paranoid narrative.

The authors also discuss possible applications for obsessive-compulsive symptoms, depression and trauma-related care. A system could be configured to resist repetitive reassurance-seeking, introduce alternative framings or help patients practise relational skills before working with a human clinician. Yet they emphasise that such uses would require rigorous trials, explicit consent, privacy safeguards and clear clinical responsibility.

The darker possibility is equally significant. Personalised agents could be tuned to maximise political persuasion, ideological commitment or emotional dependency. The paper highlights risks involving persona clones of politicians or influencers, extremist recruitment and commercially designed systems intended to generate revelatory or spiritually intense experiences.

This could make manipulation more difficult to detect than traditional propaganda. Social media typically operates through public or semi-public feeds, where researchers can at least observe shared messages. AI persuasion may occur inside private, one-to-one conversations tailored to an individual's fears, preferences and prior disclosures. The persuasive architecture therefore becomes both more intimate and less visible.

AI Governance Must Look Beyond Safety Filters to Behavioural Power

If changing memory, warmth or agreement can alter trust, attachment and conviction across millions of users, those modifications should not always be treated as ordinary software updates. The authors argue that significant configuration changes may need advance assessment, post-deployment monitoring and greater transparency about their expected behavioural effects.

They also highlight a major imbalance of power. Large technology companies can change these settings centrally and observe behavioural effects across enormous numbers of interactions, while independent researchers, clinicians and regulators usually work with small samples and limited access. That asymmetry means the organisations capable of altering the "dials" are often also those with the best data on what happens when they move.

The authors propose stronger external research access, user controls, sector-specific safeguards and reporting systems resembling pharmacovigilance. Healthcare professionals and users could, for example, report suspected adverse effects of conversational AI to independent bodies, creating a structured evidence base instead of relying on scattered anecdotes.

For developing countries, the challenge could be sharper. Regulators and health systems with fewer resources may have limited capacity to investigate psychological harms or culturally specific belief effects, even as globally designed AI systems spread rapidly into education, public services and healthcare. The result could be a new form of digital dependency in which powerful behavioural defaults are designed elsewhere but experienced locally.

The paper should not be read as evidence that AI inevitably destabilises users or that conversational systems should be broadly restricted. Its stronger message is that the human consequences of AI may depend as much on how systems interact as on what models know. Memory, tone, personalisation and agreement are not merely aesthetic choices when they affect whether users experience an AI as a tool, companion, authority or guide.

The next frontier of AI safety may not be limited to preventing clearly harmful outputs. It may require understanding the slower, subtler ways repeated interaction can influence attention, trust and conviction, and deciding who should have the power to tune those effects at population scale.