From observation to engagement

I recently joined Ray Poynter on the Talking AI in Market Research podcast to discuss the intersection of human behaviour, cognition, and AI. While we covered a lot of ground, a single conversation can only go so far. This post expands on some of the key themes we discussed, particularly how large language models (LLMs) influence and are influenced by human interaction.

If you’re interested, I encourage you to listen to the full episode (available on the Talking AI podcast - links below). Here, though, I want to explore some points that shape my approach to this field and connect the dots to content covered on Artificial Thought so far.

You can't shape a conversation by watching from the sidelines.

About eighteen months ago, I attended two market research industry events in Amsterdam (ESOMAR Congress and ESOMAR AI conference). At that point, my working knowledge of AI was optimistic at best. It hadn’t been a particular focus of mine, and if I’m honest, it hadn’t sparked much curiosity but sitting in those sessions, listening to conversations about how AI was already transforming industries, I had a sharp realisation: this wasn’t something marginal or optional. AI was seeping into everything. Whether we want it or not, it’s here and it’s reshaping the systems we live and work in.

For me, that moment wasn’t about enthusiasm or alarm - I’m not an AI evangelist, and I’m certainly not a doomsayer. I would describe myself, then as now, as a pragmatist. The question wasn’t whether I liked AI. It was whether I was willing to engage with it, and it became clear that standing on the sidelines wasn’t a serious option.

Looking around those conferences, what struck me was just how much this technology was slipping quietly into everything - and the sheer number of men explaining AI to other men. For a technology as far-reaching and consequential as this, the imbalance was hard to ignore. I don’t claim to know exactly how greater inclusion changes AI, but I do know that diverse perspectives change systems. They shape what questions are asked, what blind spots are noticed, what assumptions are challenged. Leadership matters - not just in who holds it, but in how they shape what comes next.

At the time, I wasn’t sure where I would fit into this space because machine learning engineer was not, and will never be, on my bucket list. The more I thought about it, though, the clearer it became: every product, every system, every innovation ultimately intersects with human behaviour. Somewhere, someone’s behaviour determines success or failure whether it’s a user, a customer, a policymaker, or a stakeholder, and understanding human behaviour is what I know.

Over time, I saw where behavioural science could make a meaningful contribution to building, evaluating, and humanising these systems from the ground up. My contribution is just one small piece of a larger conversation, but by participating, I hope to encourage others to step in as well. As I often remind myself, when you notice an absence, you have a choice: complain about it or be the change you want to see. If not me, then who?

The key to understanding LLMs lies not in what they can do, but in how people interact with them and what that reveals about human behaviour.

When I began using LLMs, I approached them through the lens of behavioural science, focusing on how people interact with these systems, the expectations they bring, and how these interactions shape meaning, judgement, and decision-making.

My background equipped me to concentrate on patterns of interpretation, not just output. I often use LLMs as co-intelligence, as Ethan Mollick describes: not for micro-tasks, but as real-time creative or analytical partners. It’s like briefing a junior colleague—sometimes brilliant, sometimes baffling. This dynamic relationship often starts without a clear goal, with the LLM helping to shape the task through iterative exchange.

This mindset is crucial because many still treat LLMs as single-shot task completers, limiting their utility. Effective use requires skills in framing, probing, and refining questions, setting the right cognitive context.

Behavioural frameworks also help diagnose friction when people’s mental models don’t align with system behaviour. A central challenge is social misalignment: AI systems may be technically impressive but lack understanding of human context, norms, and motivations. This isn’t just about missing “common sense”; it’s about a missing model of people.

Behavioural science can make AI systems more human-aware by embedding insights into how we think, feel, and relate. Critically, it can address the perception gap highlighted in a Pew study from April 2025, where 76% of AI experts see these technologies as beneficial, but far fewer in the general public agree. This gap points to issues of trust, usability, and psychological readiness, all areas where behavioural science can contribute.

For me, engaging with AI has been a pragmatic choice sparked by a recognition that, whether we like it or not, these systems are reshaping the environments we live and think within. If we want those environments to work better for more people, we need to be willing to step in.

Whether LLMs count as thinkers depends more on how humans define and interpret reasoning than on technical capacity.

Discussing large language models (LLMs) presents an immediate challenge: language itself. We often hear that LLMs can’t “think” or “reason,” but pressing these claims reveals a deeper question: What do we mean by thinking or reasoning?

The answer depends on the framework we use. Are we referring to cognitive processes like heuristics and biases? Are we using a normative lens, focusing on logical validity or goal achievement? Or are we considering phenomenological terms, where conscious awareness is the test?

Narrow definitions, such as formal deduction, will always exclude LLMs. Broader definitions, like reasoning as inference under uncertainty or simulating consequences, begin to include them. Similarly, if thinking is defined as conscious experience, LLMs do not qualify. If it’s the manipulation of structured representations towards a goal, the picture becomes less clear.

We already apply these terms pragmatically. We say thermostats “think” it’s too hot, chess programs “plan” moves, or dogs “know” they did something wrong. These aren’t metaphysical claims but shorthand for observed behaviour. The reluctance to extend this pragmatic lens to LLMs may reveal more about human psychology than machine capability.

There’s a deeper irony: human thinking relies heavily on pattern recognition and statistical inference. The predictive processing model of the brain suggests that cognition operates through hypothesis generation, sensory input matching, and belief updating to minimise prediction error. LLMs, similarly, predict the next token, simulate coherent sequences, and adjust internally across interactions. Prediction thrives in both evolution and machine learning because it is efficient, robust under uncertainty, and scalable across domains like language, perception, and social inference.

Breaking human thought down into pattern recognition, context-sensitive updating, and statistical inference makes LLMs seem less alien. They are reflective artefacts—trained on us, shaped to approximate us, and now offering a distorted mirror through which to examine our own minds.

People evaluate LLMs through social lenses, projecting intent, coherence, and even morality onto systems that lack these qualities.

The core challenge in human–LLM interaction is not technical but cognitive and social. People approach LLMs as if they are minds, judging them on coherence, intent, and social appropriateness, not just factual accuracy. This creates a mismatch, as LLMs do not “think” in the human sense but are evaluated as if they do.

Several tensions shape this interface:

• Trust and transparency: Users tend to distrust systems they don’t understand but often overestimate their understanding of systems that mimic human-like interactions. LLMs build trust by offering explanations, even if those explanations aren’t entirely accurate.

• Flexibility and predictability: People value adaptability but also need consistency. Excessive flexibility can make an LLM seem unreliable. Consistency in tone, reasoning style, or values anchors user expectations and supports usability.

• Agency and safety: Users want control but can be overwhelmed by full autonomy. Agency within LLM systems should be introduced gradually and bounded by clear constraints.

• Cultural fit: Expectations around emotion, individuality, and politeness vary widely. An LLM that seems helpful in one setting may feel inappropriate in another. Models trained on aggregate data can easily violate local norms.

At a deeper level, cognitive mismatch occurs when engineering goals clash with behavioural goals. Without stable cues, people struggle to form coherent mental models, leading to friction, overtrust, or disengagement.

Social dynamics are unavoidable. People naturally anthropomorphise LLMs, projecting personality, morality, and intention onto them. Designers must consider this tendency, as it shapes how people interpret outputs, assign responsibility, and respond to errors.

Behavioural science helps identify these interpretive mismatches early, avoiding the assumption of rational, universal users. Crucially, it helps design LLMs to behave not just functionally, but in ways that align with how people expect minds to behave, not machines.

Inconsistent or opaque limitations erode trust, pushing users towards workarounds and reshaping their relationship with the system.

As LLMs become more integrated into everyday tools, their limitations and guardrails are becoming increasingly visible and consequential. These constraints are not merely technical; they shape how people experience, trust, and adapt to the system.

• Inconsistency is a persistent issue. The same question might yield a refusal one day and a full answer the next. This unpredictability makes it difficult to develop reliable prompting strategies, especially in sensitive domains. Behaviourally, inconsistency erodes trust and encourages trial-and-error prompting, increasing cognitive load and the risk of misinterpretation.

• Overcautious filtering is another challenge. Topics requiring nuance are often shut down entirely, and the system may refuse even well-intentioned, constructive queries. This limits usefulness, particularly in educational or research contexts.

• Pseudocompliance adds another layer of complexity. LLMs often reframe unsafe queries into neutral-sounding answers while still delivering problematic content. For example, a question about crash diets might return under the banner of “nutrition advice,” with the underlying risk unchanged.

Many refusals appear driven more by reputational risk than genuine safety concerns. Behaviourally, this teaches users that boundaries are arbitrary and shaped by corporate protection, not ethical principle. People respond predictably by learning to “prompt around” the rules. Over time, the system effectively trains users to treat compliance as an obstacle rather than a shared value.

Behavioural science can help by redesigning systems that work with predictable human reactions, fostering a more trusting and cooperative relationship between users and LLMs.

The most effective approach is to treat LLMs as collaborators, experiment deliberately, and focus on usefulness over correctness.

While I hesitate to offer universal advice for working with LLMs, as these systems and their uses are rapidly evolving, a few principles stand out:

• Treat LLMs as partners, not tools: Instead of crafting perfect queries, focus on framing tasks, providing context, and refining through interaction.

• Experiment deliberately: Try different ways of phrasing, sequencing, and structuring requests. View this as a learning process, not a one-time task.

• Develop the skill over time: Think of it as learning to cook without a recipe or working with a fast but occasionally off-target colleague. With practice, you’ll develop a feel for what works.

• Focus on usefulness over correctness: Aim to improve your own thinking, spark ideas, challenge assumptions, and pressure-test reasoning. Don’t assume LLMs are always correct; prioritise their utility.

• Explore multiple models: Different systems have unique strengths and quirks. Avoid settling too quickly, as valuable insights can come from finding a system that complements your thinking style.

The real challenge ahead is not building better systems, but building better relationships between humans and machines.

As artificial intelligence integrates into our environments, its influence on human cognition and agency becomes less visible but no less significant. Eighteen months ago, what struck me at those conferences was not just AI’s technical complexity, but its inevitability. These systems are reshaping our environments, whether we like it or not. My realisation was pragmatic: the choice was never about wanting AI to exist, but about engaging with it, understanding its intersection with human behaviour, and shaping it more thoughtfully.

Over the past year, this decision has deepened into a conviction. The work ahead is about developing better relationships between humans and machines. This means focusing on the interface, not just the engine; on the interaction, not just the output. It means considering what kinds of thinking, collaboration, and judgement we want to cultivate and what we are willing to delegate.

If there is one thread running through all of this, it is the need for more diverse voices, perspectives, and disciplines in the conversation. Leadership and inclusion matter. The work of making these systems more human-aware cannot be accomplished by technical teams alone.

For me, stepping into this space was a way of answering the question that struck me at the start:

If not me, then who? And if not now, then when?

For those curious about a broader discussion, the Talking AI in Market Research episode offers a complementary exploration—and here on Artificial Thought, I’ll continue to examine the human side of these questions.

Podcast episode can be found here:

Spotify | Amazon Music | Apple Podcasts | Castbox | PocketCast | ListenNotes | YouTube