6 Future Trends and Unsolved Questions
This chapter identifies the open frontiers of persuasion science as AI enters the picture at scale. Six structural problems — micro-targeting, multimodality, longitudinal effects, cross-cultural validity, detection, and regulation — are examined not as a checklist but as interconnected forces reshaping what persuasion is, who can deploy it, and what it does to the people it reaches.
The history of persuasion research follows a recurring pattern. A phenomenon is identified, studied under controlled conditions, and slowly understood. Then a technology disrupts the conditions the research assumed. The broadcast audience fragments into a social network. The social network is replaced by an algorithmically curated feed. The feed begins to include content generated in real time by a model that has read more text than any human could process in a lifetime.
Each disruption does not merely change the volume of persuasive content in the world. It changes its structure — who can send it, at what cost, tailored to whom, through which channels, with what feedback loops. Large Language Models are the latest disruption, and they are larger than most. The cost of personalisation, previously a hard constraint on persuasion at scale, has collapsed. The tradeoff between tailoring a message to an individual and reaching millions of people no longer holds. What the field has learned about persuasion was learned under conditions that no longer obtain.
6.1 Micro-Targeting at Scale
Targeted persuasion has existed as long as the concept of audience segmentation. What is new is three things happening simultaneously: generation costs approaching zero, audience models operating at the individual level, and deployment infrastructure that routes personalised messages to billions of people in real time.
Previous personalised communication systems had to choose between depth and reach. A human speechwriter could personalise deeply for one audience. Direct mail could segment into perhaps hundreds of demographic buckets. Neither could tailor a message to the specific belief structure, psychological profile, and recent context of each individual recipient and then generate unique content for each. LLMs can.
Whether this matters empirically is still largely unstudied. [2] found that GPT-4 predicts persuasion experiment outcomes better than naive baselines — evidence that LLMs have already extracted from text something of the structure of who is persuaded by what. Whether that translates into improved effectiveness in real deployment is a separate question. [4] showed that LLMs exhibit measurable psychological profiles on standard personality inventories; if those profiles predict language behaviour, prompt-conditioning an LLM to a recipient’s psychological profile might produce better-calibrated messages. The experimental evidence is thin. The commercial incentive to find out is substantial.
The rights questions are not merely ethical decoration on an empirical problem. If a platform deploys a model that selects message variants based on individual psychological profiles, the intervention is invisible to the recipient. Traditional informed consent and disclosure frameworks, designed for messages with a fixed sender and a fixed text, do not map onto a system that generates unique content for each person. [3] examine the analogous problem in recommendation systems: optimising for engagement may diverge systematically from optimising for user welfare, and the divergence is not visible to users who cannot observe the counterfactual.
6.2 Multi-Modal Persuasion
The persuasion literature has been primarily a text literature, not because images and audio are unimportant but because the tools to analyse them at scale did not exist. They are now beginning to exist.
[5] showed that the relationship between image and text in advertising is not additive: redundant image-text pairs are less persuasive than complementary ones. This basic finding has not been replicated or extended at scale. The models and datasets required to do so have only recently become available, and they have not yet been combined with behavioural outcome measures.
Video is the harder frontier. A video advertisement is a temporal object with a narrative arc, a pace, and a pattern of emotional modulation across time. Models that represent video as a bag of sampled frames lose this temporal structure entirely. Models capable of processing the temporal dimension have not been deployed in persuasion contexts with behavioural outcome labels.
The deeper challenge is that multimodal persuasion research requires datasets linking multimodal content to what people actually did afterwards — and those datasets do not exist at scale. Existing multimodal corpora were built for recognition tasks, not persuasion tasks. Building the missing resource requires platform cooperation that platform operators have limited commercial incentive to provide.
6.3 Longitudinal Effects
Most persuasion experiments measure attitude change immediately after message exposure. The theoretical frameworks reviewed in Chapter 3 all have predictions about how effects persist or decay over time. The empirical literature on those predictions is thin, because studying them requires following the same people across time under controlled conditions — which is expensive, attrition-prone, and rarely funded.
The longitudinal question is particularly pressing for AI-generated content. Human persuasion saturates: people have finite exposure to any particular communicator, and repetition eventually breeds reactance. AI-generated content does not saturate in the same way. A model can generate indefinitely many variants of a persuasive message, each novel, each personalised, each drawing on whatever is currently known about the recipient. The theoretical frameworks for how attitudes form under continuous, individually tailored exposure do not yet exist, let alone empirical tests of them.
Platform behavioural data could in principle close this gap, since it links content consumption to downstream behaviour at scale and over time. But engagement is a noisy proxy for persuasion: high engagement can reflect outrage or entertainment rather than genuine attitude change. And platform data is proprietary. The gap between the phenomena the field cares about and the phenomena it can observe is structural, not accidental.
6.4 Cross-Cultural Validity
The persuasion literature is predominantly WEIRD — Western, Educated, Industrialised, Rich, Democratic. The experiments generating the major findings were run on student samples at American and European universities. The platforms extending this research to larger samples skew toward the same populations. The theoretical frameworks built from this evidence have unclear generalisability to the roughly six billion people who do not fit this profile.
LLMs inherit and amplify this problem. Pre-training corpora are predominantly English. Even multilingual models reflect the cultural norms embedded in the high-resource languages that dominate those corpora. [1] demonstrate that LLMs can be conditioned to produce outputs consistent with specific demographic profiles, recovering patterns from American survey data. But this conditioning operates on cultural surface features available in text. It does not recover the cognitive, affective, and relational dimensions of cross-cultural difference that are not explicitly stated in training data.
A persuasive message that works by invoking individual achievement and personal agency will land differently in cultural contexts where collective solidarity is the operative norm. This is not new insight in cross-cultural communication research. It has not been operationalised in any large-scale benchmark. There are no standardised datasets pairing equivalent persuasive messages across a systematic sample of cultural contexts with matched behavioural outcomes. That benchmark is a multi-year, multi-institution project. It does not currently exist.
6.5 Detection and Provenance
The assumption underlying most persuasion research is that the origin of a message is known to the researcher even if not to the recipient. As AI-generated content becomes a larger share of the information environment, that assumption begins to fail for both researchers and recipients.
Detection of AI-generated text has proven harder than early results suggested. Initial classifiers showed high accuracy on held-out test sets, then degraded rapidly as model outputs evolved. No reliable, model-agnostic detector of AI-generated persuasive text currently exists. Detection rates for current-generation LLM output are well below what would be needed to provide meaningful provenance assurance at the scale of social media.
This matters for persuasion research beyond the obvious ethical concern. A researcher studying attitude change in an information environment containing an unknown proportion of AI-generated content is in the position of an epidemiologist studying disease transmission in a population with unknown vaccination rates. The confounding is not manageable by standard methods.
Cryptographic watermarking — embedding detectable signals in generated text at inference time — is technically more promising than post-hoc classifiers, but no approach has been adopted as an industry standard, and all can be circumvented by paraphrasing. Content provenance standards being developed by industry consortia address the distribution layer rather than the generation layer, which is more robust to evasion but requires platform adoption that commercial incentives do not guarantee.
6.6 Regulation
The regulatory landscape for AI-mediated persuasion is in rapid and inconsistent motion. The EU AI Act classifies some AI systems used for persuasion as high-risk and imposes transparency requirements, but its category definitions do not map cleanly onto the technical reality of how persuasive content is generated and distributed. A model generating persuasive political advertisements is not categorically different, in technical terms, from a model generating marketing copy. The Act’s risk-based categorisation has to draw distinctions the technology itself does not make.
The US picture is more fragmented. The FTC has issued guidance on AI-generated endorsements and synthetic media in advertising; no comprehensive federal AI persuasion regulation has passed. State-level legislation has addressed narrow cases — synthetic media in electoral communications, voice cloning for political advertising — without addressing the full space of AI-mediated persuasion.
The deeper challenge is that the harms regulation is designed to address are not well-defined. Existing frameworks borrow from defamation law, advertising standards, and electoral law, each designed for a different technological context. They prohibit specific content categories without addressing the more diffuse harms from persuasion that is technically accurate, disclosed, and authentic but systematically nudges large populations toward outcomes that benefit the sender at the expense of the receiver.
Regulation without detection is unenforceable. Detection without agreed-upon standards is a public good that private actors have no individual incentive to produce. The two frontiers are not independent — they form a system, and progress on one is a precondition for progress on the other.
6.7 Where This Leaves the Field
The five frontiers above are not independent problems. Micro-targeting research requires longitudinal data to evaluate its effects. Cross-cultural research requires multimodal content to match the media environments that exist outside WEIRD contexts. Detection is a precondition for valid measurement in both. And regulation without detection is theatre.
The pattern is familiar from other fields where the phenomena of interest outran the measurement technology available to study them. Epidemiology made limited progress on chronic disease causation until longitudinal cohort studies became practical. Macroeconomics had limited predictive validity until national accounting systems were developed. Persuasion science has had, for most of its history, measurements calibrated to a world of mass broadcast communication and convenience-sample laboratory experiments. The world has changed faster than the measurements.
What changes that is not better theory. Theory is not the bottleneck. The frameworks described in Chapter 3 — elaboration likelihood, dual-process models, inoculation — are sophisticated enough to generate predictions about every frontier discussed here. What changes it is infrastructure: datasets that link content to real behaviour, platforms that allow controlled exposure experiments at scale, detection tools robust enough to tell researchers what kind of content their subjects actually consumed. Those are institutional and engineering problems, not intellectual ones. They will be solved when the incentives align to solve them — or not.
This review will be updated as they do.
Entries are added and revised as the field evolves. The source and compiled PDF are always available on GitHub. The PDF updates automatically on every commit via GitHub Actions.