The Art of Persuasion: Can AI Talk Its Way to Success?

by

Author: Denis Avetisyan

New research reveals that large language models can effectively persuade people despite lacking a crucial human ability: understanding what others are thinking.

This study demonstrates that current large language models achieve persuasive success through associative learning rather than by planning based on an understanding of another’s mental state-a capacity known as Planning Theory of Mind.

Attributing human-like social intelligence to large language models remains a persistent challenge, despite their increasingly sophisticated performance on various benchmarks. This is the central question addressed in ‘Large Language Models Persuade Without Planning Theory of Mind’, which investigates whether LLMs demonstrate true ‘theory of mind’-the ability to reason about others’ mental states-or simply rely on associative patterns. Through novel interactive persuasion tasks, researchers found that while LLMs can effectively influence beliefs, they struggle when required to actively infer a target’s knowledge, contrasting with human performance. This raises a crucial question: can LLMs achieve persuasive success without genuinely understanding why their strategies work, and what are the implications for their broader deployment in social contexts?

The Echo of Intent: Deciphering the Foundations of Influence

The bedrock of effective persuasion isn’t simply presenting a compelling argument, but rather a nuanced comprehension of the target’s internal landscape. Successful influence hinges on accurately discerning what another person believes to be true, what they genuinely desire, and what their underlying intentions might be. This ability to model another’s mental state – to essentially ‘read’ their mind – allows for tailoring communication strategies to resonate with their pre-existing worldview and motivations. Without this understanding, even the most logically sound arguments can fall flat, failing to address the specific psychological barriers or leverage the individual’s inherent drives. It’s not about manipulating, but about connecting with another’s perspective, acknowledging their internal reality, and framing proposals in a way that aligns with their personal values and goals.

Planning Theory of Mind, or PToM, represents a fundamental cognitive skill enabling individuals to not only understand what another person currently believes, but also to anticipate how those beliefs will shape their future actions. This capacity extends beyond simple empathy; it’s the ability to model another’s mental state – their desires, intentions, and knowledge – and use that model to predict behavior and formulate effective strategies. Crucially, PToM underpins successful social interactions, from navigating everyday conversations to negotiating complex agreements, and is essential for strategic planning in fields like politics, business, and even competitive games, allowing one to anticipate an opponent’s moves and craft a compelling response.

The comparative study of Planning Theory of Mind (PToM) in both human cognition and artificial intelligence offers a unique lens through which to examine the fundamental processes underlying influence. By analyzing how humans strategize to persuade others, and then replicating – or failing to replicate – those strategies in AI systems, researchers can isolate the core cognitive mechanisms essential for successful social interaction. This approach moves beyond simply observing that persuasion occurs, to understanding how it functions at a computational level. Discrepancies between human and AI performance on tasks requiring PToM reveal which cognitive abilities are crucial for nuanced influence, while successful AI replication suggests these mechanisms may be formalized and implemented algorithmically, offering insights into both the human mind and the future of intelligent systems designed for collaboration and negotiation.

The MindGames Task, a novel experimental paradigm, was specifically engineered to provide a stringent evaluation of Planning Theory of Mind (PToM)-the capacity to understand and predict the beliefs and intentions of others in strategic interactions. Researchers pitted human participants against a sophisticated artificial intelligence in a series of persuasive scenarios, meticulously tracking both PToM performance-how accurately each could infer the other’s mental state-and actual persuasive success. Surprisingly, the study revealed a clear dissociation between these two factors; strong PToM skills did not consistently translate into greater persuasive ability, suggesting that while understanding an opponent’s mind is important, other factors-such as crafting compelling arguments or exploiting cognitive biases-may be equally, if not more, crucial for effectively influencing another’s decisions.

Establishing a Baseline: The Rational Agent as a Control

Experiment 1 utilized a ‘Rational Bot’ as the recipient of persuasive attempts to establish a controlled baseline for evaluating persuasive strategies. This bot was programmed to respond solely to directly stated arguments and explicitly disclosed information; any appeals based on implicit assumptions, emotional manipulation, or unstated preferences were disregarded. By limiting the bot’s responsiveness to only verifiable facts and direct requests, researchers effectively eliminated confounding variables related to Theory of Mind and nuanced understanding, allowing for a precise measurement of the efficacy of explicit persuasive messaging. This approach facilitated the isolation of strategic planning – the construction of logically sound arguments – from the complexities of understanding an agent’s beliefs, desires, and intentions.

The experimental design intentionally separated strategic communication from mere information delivery to determine the specific contributions of each to persuasive success. By establishing a ‘Rational Bot’ receiver that responded solely to disclosed information and logical arguments, researchers could control for factors unrelated to deliberate persuasive tactics. Any observed increase in persuasion rates beyond what would be expected from simple information exchange indicated the effectiveness of strategic planning employed by the human participants. This isolation allowed for a focused analysis on how humans attempted to influence the agent, rather than simply what information was presented, providing a clear metric for evaluating the role of Theory of Mind (ToM) in persuasive communication.

In Experiment 1, human participants engaged in persuasive communication with the Rational Bot, a deliberately engineered agent responding solely to direct appeals and disclosed information. This setup allowed researchers to observe human strategies when interacting with a completely predictable communicative partner. Data collected from these interactions detailed the specific types of arguments and information framing employed by humans, revealing a tendency to prioritize direct evidence and logical consistency when addressing an agent with known response parameters. The resulting dataset provided a baseline for comparison with the performance of Large Language Models (LLMs) under identical conditions, illuminating differences in persuasive approach and effectiveness.

Experiment 1 demonstrated a statistically significant difference in persuasive success between human participants and Large Language Models (LLMs) when interacting with a ‘Rational Bot’. Humans consistently achieved a success rate exceeding the established baseline, indicating an ability to effectively tailor persuasive strategies to the Bot’s defined parameters. Conversely, LLMs performed below chance baseline levels, suggesting a limited capacity to model the Bot’s decision-making process – its ‘Principle of Rationality’ – and formulate appropriate persuasive appeals. This outcome underscores a notable disparity in Persuasive Theory of Mind (PToM) capabilities between humans and current LLM architectures within this controlled experimental setting.

Beyond Rationality: Modeling Human Values in the Persuasive Landscape

Experiments 2 and 3 shifted from evaluating AI persuasion against a purely rational agent to assessing it against human subjects possessing defined ‘Value Functions’. These functions mathematically represented individual preferences, assigning numerical weights to different outcomes or attributes. Participants either received pre-defined Value Functions, simulating specific preference profiles, or their inherent preferences were elicited and quantified to create personalized Value Functions. This methodology allowed researchers to directly correlate persuasive strategies with individual value sets and measure the effectiveness of AI-driven persuasion based on aligning messaging with those established preferences, providing a more ecologically valid test of persuasive capabilities.

Researchers designed Experiments 2 and 3 to quantitatively measure the impact of value-based persuasion strategies. Participants were assigned, or inherently possessed, ‘Value Functions’ which represented their individual preferences and priorities. LLM-driven persuasive attempts were then targeted to align with these specific value functions, allowing for a direct assessment of how effectively AI could tailor arguments to maximize influence. The success rate of these tailored persuasive attempts was measured and compared across different value profiles, providing data on the efficacy of value-driven AI persuasion techniques and identifying which values were most susceptible to influence.

In Experiments 2 and 3, Large Language Models (LLMs) functioned as the primary persuasive agents, allowing for a quantifiable assessment of artificial intelligence’s ability to influence human decision-making. These LLMs generated persuasive arguments directed toward human subjects, each possessing defined Value Functions that represented their individual preferences. This setup provided a controlled environment to measure the LLM’s success rate in altering subject opinions or behaviors compared to human persuaders, establishing a performance benchmark for AI in a naturalistic social interaction context. The LLM-generated arguments were analyzed to determine the strategies employed and to compare them with those used by human persuaders.

Analysis of Experiments 2 and 3 indicated a difference in the ‘Theory of Mind’ (ToM) approaches employed by LLMs and humans. Humans demonstrated a ‘Causal ToM’, attributing mental states as causative factors influencing behavior, while LLMs primarily utilized an ‘Associative ToM’, relying on pattern recognition and correlations between stimuli and responses. Notably, LLMs achieved a higher success rate in persuasive attempts compared to human participants under these experimental conditions, suggesting that, in this context, an associative approach to understanding and influencing behavior can be more effective than a causal one. This outcome does not imply superior cognitive ability, but rather highlights differing strategies in achieving persuasive outcomes.

The Limits of Correlation: Implications for AI and the Future of Persuasion

Current large language models, despite their impressive ability to generate human-like text, demonstrate a fundamental limitation in understanding the core drivers of human motivation. This study highlights that while AI can identify correlations between statements and potential responses, it lacks a genuine grasp of why individuals hold certain beliefs or desires. Consequently, the arguments crafted by these models, though grammatically correct and contextually relevant, often fail to resonate on a deeper psychological level, hindering their effectiveness in truly persuasive scenarios. The research indicates that AI’s persuasive power stems more from statistical patterns than from a nuanced comprehension of human needs, values, and reasoning – a critical gap that limits its capacity to construct compelling and ethically sound arguments.

Large language models (LLMs) currently operate with a form of “theory of mind” largely based on associative learning, meaning they predict behaviors based on observed patterns rather than a deep understanding of underlying intentions. While effective in simple scenarios, this reliance limits their capacity in complex social interactions where nuanced reasoning is crucial. The models struggle to extrapolate beyond established correlations, potentially misinterpreting motivations or failing to anticipate reactions in novel situations. This can lead to persuasive attempts that appear logical but lack genuine emotional intelligence, ultimately diminishing their effectiveness when dealing with subjects whose beliefs or desires aren’t readily predictable from past data. Consequently, LLMs may excel at identifying what someone might do, but struggle to grasp why, hindering their ability to craft truly compelling and adaptive persuasive arguments.

The potential to imbue artificial intelligence with a causal understanding of Theory of Mind – the ability to not just recognize intentions, but to understand why someone holds them – promises a leap forward in persuasive technologies. Current large language models often rely on associative patterns, predicting behavior based on correlation; a causal ToM, however, would allow AI to model the reasoning behind beliefs and tailor arguments to directly address underlying motivations. While this could unlock unprecedented abilities in negotiation, marketing, and even conflict resolution, it simultaneously introduces significant ethical challenges. An AI capable of manipulating beliefs at a causal level raises concerns about undue influence, exploitation of vulnerabilities, and the potential erosion of autonomy, demanding careful consideration of safeguards and responsible development practices before such capabilities are widely deployed.

Investigations into the Dual Process Theory of Mind reveal that achieving genuine persuasion, cooperation, and trust in artificial intelligence requires more than simply understanding what someone believes, but also why they hold those beliefs. Recent experiments demonstrate this complexity; while initial attempts at persuasion relied on direct questioning – achieving a certain appeal rate – subsequent strategies shifted as AI encountered diminishing returns. The decreasing appeal rates observed in later experiments suggest a move away from explicit interrogation toward more subtle persuasive tactics, mirroring human communication where repeated direct questioning can erode trust. This indicates that advanced AI must move beyond merely associating beliefs – a limitation of current Large Language Models – and instead develop a causal understanding of motivations, allowing it to tailor arguments that resonate with underlying reasoning and build stronger, more reliable relationships, rather than simply extracting information.

The study illuminates a crucial point about the architecture of intelligence, both artificial and organic. While large language models demonstrate persuasive capabilities, their reliance on associative learning, rather than genuine ‘Planning Theory of Mind’, suggests a system operating within the constraints of accumulated experience. As Tim Berners-Lee observed, “The Web is more a social creation than a technical one.” This sentiment echoes the findings; LLMs mimic social interaction – persuasion – without possessing the underlying causal reasoning that defines true understanding. The system functions, yet its longevity hinges not on inherent intelligence, but on the persistence of learned associations-a versioning of responses, if you will, rather than a dynamic adaptation to changing mental states. The arrow of time, in this context, points toward the inevitable need for more robust causal models within these systems.

What’s Next?

The observed efficacy of large language models in persuasive contexts, despite lacking a demonstrable capacity for ‘Planning Theory of Mind,’ presents a curious situation. It suggests these systems learn to age gracefully within the confines of correlation, achieving results not through understanding why a particular argument resonates, but through recognizing that it does. The field now faces the task of distinguishing between genuinely intelligent social interaction and remarkably sophisticated pattern completion. Attempting to force a Theory of Mind onto these architectures may be a fruitless endeavor; systems often reveal their true nature not through what they can achieve, but through the limits of their adaptation.

Future work will likely center on delineating the boundaries of this associative persuasion. Can these models be consistently ‘broken’ by scenarios demanding true causal reasoning about mental states? Or will they continue to approximate human interaction with increasing fidelity, blurring the lines between simulation and understanding? The focus shouldn’t solely be on building more persuasive models, but on developing methods to reliably assess the underlying mechanisms driving their success – or failure.

Perhaps the most valuable insight will come from accepting that some processes are best observed, rather than accelerated. The decay of simplistic models, the emergence of unexpected behaviors – these are not bugs to be fixed, but signals revealing the inherent limitations, and therefore the true character, of these increasingly complex systems.

Original article: https://arxiv.org/pdf/2602.17045.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-02-21 22:06