Prompt Engineering Was an Illusion – Empathy Is the Key in AI Collaboration

A Paradigm Shift in Human–AI Collaboration

For years, the AI community chased technical solutions: bigger models, more parameters, clever prompts, longer context windows, and higher benchmark scores. We believed that if we just engineered better prompts or used the latest model, the AI would dutifully spit out the answers we wanted. In other words, we treated prompt engineering as the magic key to unlock AI’s potential. However, a recent study (2025) is calling for a fundamental shift in this mindset. It turns out that the real bottleneck in AI performance isn’t the model’s technology at all – it’s our own ability as humans to collaborate with a non-human intelligence  . The core of succeeding with AI is not about technical tricks, but about social intelligence – empathy and perspective-taking – in our interaction with the AI . In short, prompt engineering is not a technical skill we hack; it’s a human collaboration skill we cultivate .

This research reframes the question of AI utility from “How smart is the AI by itself?” to “How much smarter can we become when we work together with AI?” . Instead of obsessing over whether AI will replace humans or which model wins on a leaderboard, it asks how human–AI teams can complement each other. The findings are eye-opening: even the most advanced AI systems today are only as useful as the collaborative dynamic they enter with human users. If a human doesn’t understand the AI’s way of “thinking” (or rather, pattern-matching), then even the best model’s performance gains are wasted  . On the other hand, when a person learns to truly collaborate – supplying context, anticipating misunderstandings, and guiding the AI – the combined team can far exceed what either could do alone. This represents a paradigm shift from a technology-centered view of AI to a relationship-centered understanding of working with AI.

Distinct Skills: Solo Problem-Solving vs. AI Collaboration

One of the most striking findings of the study is that human solo intelligence and the ability to work with AI are not the same skill. In fact, they found these two abilities are “measurably, demonstrably different” . Being a brilliant problem-solver on your own says almost nothing about how well you’ll do when collaborating with an AI . The researchers analyzed data from 667 people tackling various problems, sometimes alone and sometimes with an AI assistant (like ChatGPT or similar models) . When they isolated each person’s solo problem-solving ability versus their ability to solve problems with AI, the correlation between the two was nearly zero  . In other words, some individuals with high IQs or excellent solo skills actually struggled to get good results from the AI, while other fairly “average” individuals achieved extraordinary outcomes using the very same AI tool . Being smart in general or technically knowledgeable wasn’t the determining factor – collaboration was.

This has huge implications. It means all those prompt engineering courses and “10 hacks to get better answers” guides were targeting the wrong thing . Yes, certain prompt templates or techniques did seem to improve AI’s responses – but not for the reasons people assumed. There is nothing magical about phrasing a request a particular way or using secret keywords. The real reason many prompt tips and templates worked is that they “accidentally force you to practice something else entirely,” as one commentary put it . What is that “something else”? It’s the skill of seeing things from the AI’s perspective – supplying the missing context, clarifying what you really mean, and anticipating where the AI might get confused  . In essence, good prompt “engineering” is really about empathizing with your AI collaborator: treating it not as a magical answer vending machine or a search engine, but as a partner with its own strengths, weaknesses, and blind spots .

Researchers actually measured this capacity by analyzing how people wrote their prompts . Did the user provide sufficient context? Did they anticipate what the machine would need to know? Did they clarify ambiguities before they could lead to misunderstandings? In short, did they put themselves in the AI’s shoes and communicate accordingly? Users who naturally did this perspective-taking – essentially treating the AI like a colleague who might misunderstand or lack certain information – achieved far better results from the AI . Those who treated the AI as just an oracle to query or a tool to command tended to get worse outcomes. Crucially, the study showed that even the same person could get dramatically different results depending on whether they engaged in perspective-taking or not in a given query  . This suggests that success with AI is not a fixed trait you either have or lack; it’s about whether you activate the right mindset at each interaction.

Theory of Mind – The Secret Sauce for Synergy

The key differentiator identified is a cognitive skill known as Theory of Mind (ToM). In psychology, Theory of Mind refers to the ability to infer the mental states of others – to understand what someone else knows, believes, or wants, and to predict their behavior accordingly . In plain terms, it’s empathy with a purpose: the capacity to imagine the world from the other’s point of view . In the context of AI, “Theory of Mind” means modeling what the AI knows and doesn’t know, what it’s likely to misunderstand, and what it needs from you to succeed  . It’s the human ability to take the perspective of the AI – odd as that sounds – and communicate in a way that complements the AI’s “thinking” process.

The study found that Theory of Mind skill was a strong predictor of who collaborates well with AI . Participants who scored higher on this kind of perspective-taking ability (i.e. those who naturally tried to see the task from the AI’s viewpoint, filling in gaps and preventing confusion) consistently achieved better outcomes with the AI’s help  . Meanwhile, this ToM skill had little to no effect on their solo performance – it specifically improved collaborative performance, not standalone problem-solving  . In other words, a person with high Theory of Mind might not be the top performer by themselves, but when paired with an AI they shine; conversely, a solo genius without perspective-taking skills can stumble when using AI. Working well with AI is its own skillset , one grounded in social-cognitive abilities like ToM rather than traditional IQ or technical knowledge.

Importantly, ToM in this context isn’t about humanizing the AI or pretending it has feelings – it’s about recognizing the AI’s limitations as a machine. A participant with good Theory of Mind treats the AI as a partner that “has read everything ever written about your problem but has never actually had your problem,” as one commentator cleverly put it . All the AI knows is what answers tend to look like; it lacks the intuitive grasp of context that a human has . So the human’s job in the collaboration is to supply that context, to coach the AI by bridging gaps in understanding . People strong in ToM effectively become AI coaches rather than just users – they guide the AI with cues and clarifications, helping the model give a correct or useful response . For example, they might preemptively explain any potentially confusing details in their query, or double-check if the AI might have misinterpreted the goal. This aligns the AI’s pattern-matching with the human’s actual intent. As the Pixta AI research blog succinctly noted, “People with strong ToM don’t just use AI – they coach it.” 

Human–AI Synergy: AI That Makes You Smarter

One of the most exciting implications of this research is the measurable synergy it found between humans and AI. When collaboration is done right, the human-AI team performs far better than either the human or the AI alone. In the study’s experiments (which spanned domains like mathematics, physics, and moral reasoning), the average participant answered about 58% of questions correctly when working alone (human solo performance) . A large language model like GPT-4 (referred to as “GPT-4o” in the study) on its own could solve roughly 68-70% of those questions (AI solo) . But when an average human partnered with GPT-4, something dramatic happened – their combined performance jumped to around 87% accuracy . This is a ~29 percentage-point improvement over the human alone, an increase of about 30% in relative terms. Even a smaller model (an 8-billion-parameter model akin to Llama 3.1–8B) boosted human performance by roughly 23 points (from ~58% to ~81% accuracy) when teaming up, although it wasn’t as potent a collaborator as GPT-4 . In short, a good AI assistant can make an average person significantly “smarter,” at least within the scope of the tasks tested.

Average accuracy of humans alone (green), AI models alone (blue for a smaller Llama 3.1–8B model, orange for GPT-4), versus human-AI teams (right panel). The Human–AI teams substantially outperformed either the human or the AI alone. For instance, a person collaborating with GPT-4 (orange dot on right) achieved about 87% performance – much higher than the human’s solo ~58% (green) or GPT-4’s solo ~70% (orange on left) accuracy. This illustrates the synergy gain from effective collaboration . Notably, the more capable model (GPT-4) yielded a larger boost (+29 points) than the smaller model Llama (+23 points), highlighting that model improvements should be measured not just by solo accuracy, but by how much they amplify human performance.

These results drive home a crucial point: we need to rethink how we evaluate AI. Traditionally, AI researchers have fixated on benchmarks like MMLU, GSM8K, and other leaderboards that test how well a model can perform on its own  . But in real-world scenarios, AI systems will be working with people – whether it’s a lawyer using an AI to draft briefs, a doctor consulting an AI for diagnosis, or a student using AI to study. In such cases, it doesn’t actually matter if Model A beats Model B by a few points on a test when alone. What matters is how much each model helps a human achieve better results. The new study provides a framework to quantify exactly that: how much uplift a given AI provides to human performance, after accounting for task difficulty and the person’s own skill level . In their benchmark, GPT-4 demonstrated a clearly superior collaborative capability compared to the smaller model – its confidence interval for synergy didn’t even overlap with Llama’s, confirming a statistically significant advantage in how well GPT-4 enhances human problem-solving . This suggests that part of what makes GPT-4 “advanced” is not just its raw accuracy, but its ability to create true synergy with users. Going forward, the field may place more emphasis on this kind of “interactional intelligence” – designing and selecting AI systems that are not only smart in isolation, but also make their human partners smarter .

Another notable finding is who benefits the most from AI assistance. Intuitively, one might think that already high-performing individuals would gain the most (since they know best how to leverage the AI), or conversely that weaker individuals would gain the most (having more room for improvement). The study found that AI helped everyone improve, but tended to help lower-skilled users even more, narrowing the performance gap in many cases  . In other words, a good AI can act as an equalizer, raising up the floor of performance for those who might struggle alone (though it doesn’t completely eliminate skill differences)  . This is an encouraging insight: if used well, AI tools might dramatically improve outcomes for average individuals, not just turbo-charge experts. However, to unlock this benefit, those individuals still need to learn the collaboration skill – i.e. to develop their Theory of Mind and communication approach with the AI. Without that, the potential boost may not fully materialize.

Collaboration Is a Learnable Skill (Not a Fixed Trait)

Perhaps the most inspiring takeaway is that the ability to collaborate with AI is learnable and dynamic. The phrase “Theory of Mind” might sound like a fixed psychological trait – and people do vary in their natural propensity for empathy or perspective-taking. But the research indicates that this skill can be “dialed up” or “down” depending on the user’s mindset in the moment  . In fact, even the same person, working on the same type of problem, can see a difference in AI outcomes by consciously switching into a more collaborative, perspective-taking mode. The study found that “moment-to-moment changes in how much cognitive effort you put into perspective-taking directly changed AI response quality on individual prompts.”  When users actually stopped and thought, “Okay, what does this AI need to know? What might it be missing or misunderstanding here?” – they got measurably better answers . When they lapsed and just fired off a quick, ambiguous prompt without that effort, the AI’s responses suffered in quality  . In other words, Theory of Mind isn’t all or nothing. It’s like a mental muscle that you can flex deliberately. And as with any muscle, the more you practice using it, the stronger it gets . Users who approach each AI interaction as a chance to carefully articulate and clarify their intent – almost like explaining a problem to a junior colleague – are effectively training this muscle.

The fact that ToM is dynamic is great news: it means people who are not naturally empathic or clear communicators can improve their collaboration skill with AI. The researchers emphasize that ToM can be viewed as a direct mechanism through which better collaboration emerges, not just a static personal trait  . They even decomposed it into a stable component and a context-specific component in their analysis, finding that both the user’s baseline perspective-taking tendency and their momentary effort on each question affected the AI’s performance  . This suggests we can foster better human–AI teamwork by encouraging certain behaviors and mindsets each time people use AI. Simple habits – like starting with a brief outline of your goal, stating any constraints or context, asking the AI if it needs clarification, or double-checking what it understood – can all be seen as forms of perspective-taking. Indeed, the study noted even small acts such as establishing a quick rapport (e.g. a greeting), acknowledging the AI’s limitations, or explicitly stating what you already know versus what you need help with were correlated with better outcomes . These behaviors likely force us as users to slow down and consider the AI’s viewpoint for a moment, thereby activating the ToM skill. They are not magic incantations or “hacks” – they are just good communication practice .

This flies in the face of the current tech industry obsession with efficiency and automation. We often want AI to make things instant and effortless – the idea of adding “friction” or extra steps seems counterintuitive. Yet, as one expert noted, “the evidence suggests that the best human-AI collaboration emerges precisely when we slow down long enough to think about our collaborator.”  In practical terms, that means treating interaction with AI more like a dialogue or an interview, rather than barking a one-shot command. Prompting should be a conversation, not a command . Great collaborators often use iterative prompts – ask, receive, reflect, refine, and ask again – building up shared context with the AI step by step . This iterative process mirrors good teamwork in human teams, where clarification and adaptation are normal. It’s not about wrangling the perfect one-liner prompt; it’s about maintaining a clear and evolving dialogue that guides the AI toward the goal . By embracing that process, users effectively compensate for the AI’s lack of true understanding, and nudge it along the right path.

From Users to Partners: The New AI Literacy

All these insights point to a future where using AI effectively will be less about technical prowess and more about social cognition and communication. The best AI users, it turns out, are “not the most technical. They’re the ones who have learned to approach every communication (even mechanical ones) as an act of imagination.”  In other words, they treat the AI as a partner to be understood, not just a tool to be operated. This is a profound redefinition of what “AI literacy” means. It’s not just knowing what button to click or which API to call; it’s about understanding how the AI “thinks” (in its alien, statistical way) and how to talk to it effectively. It requires a blend of empathy and strategy: empathy to take the AI’s perspective, and strategy to steer the conversation productively.

We are essentially moving from a technology-centered skillset to a relationship-centered skillset when it comes to AI. Instead of asking “Do I know how to code or craft complex prompts?”, the critical question may become “Do I know how to collaborate with an AI to get the best result?” The study authors even suggest that AI models themselves should be trained and evaluated on how well they facilitate this collaboration, not just on isolated accuracy . An AI that humbly asks for clarification when a user’s query is vague, or that presents its answers in a way that is easy for the human to critique and iterate on, might be far more useful than one that aces an academic exam but cannot handle real-world ambiguity. On the human side, those who excel will be the ones who can form accurate mental models of the AI’s reasoning . For example, knowing that language models don’t truly “understand” but only predict, a savvy user will carefully phrase requests and check the AI’s output for likely failure modes. “When you can anticipate how AI might misinterpret your question, you’ve already begun to think synergistically,” as one commentator put it . This kind of intuitive anticipation is exactly what Theory of Mind is about – and it’s poised to become a core professional skill in the AI era.

Crucially, this collaborative intuition with AI is something humans can uniquely bring to the table. AI models, no matter how advanced, lack true self-awareness or understanding of others – they can’t (yet) reciprocate empathy or know what they don’t know. That means the human in the loop plays a vital role in bridging those gaps. It also means this skill won’t be automated away by the AI itself. In fact, the necessity of human perspective-taking might ensure that humans remain relevant in AI-supported work. As Justin Oberman noted in his analysis of the study, it appears that “the ancient skill of seeing the world through another’s eyes and imagining how your words will land on unfamiliar ears remain as valuable now as they were” in past human communication  – except now the “other” mind we must consider is an artificial one. Far from rendering human judgment obsolete, AI may be amplifying the importance of human empathy, communication, and leadership. The AI can provide infinite knowledge and tireless pattern-matching, but only the human can guide the interaction to make sure it’s on target.

Conclusion: Empathy – The Heart of Effective AI Use

In hindsight, it makes sense that collaboration, not command, is the heart of prompt engineering. Human communication has always required understanding your audience – whether you’re telling a joke, writing an advertisement, or teaching a class. Now, our audience includes AI systems, and despite all their computational power, they still require the same kind of consideration. The recent research on human–AI synergy has pulled back the curtain on why some users get incredible results from AI while others don’t. The secret isn’t in proprietary prompts or technical wizardry; it’s in the age-old skill of empathy, applied in a new context. It’s about having the curiosity and patience to ask: What does my AI partner need from me so it can help me?

The moment a human user truly understands how the AI “thinks” and adapts to it, the dynamic flips – suddenly the AI becomes vastly more useful, almost as if it got smarter itself. In reality, it’s the human who got smarter by learning how to speak the AI’s language. This is the essence of what the researchers call “collaborative intelligence,” where the human and AI together form a more powerful problem-solving unit than either alone  . Building this kind of synergy will be one of the most important skills in the coming years. Businesses, educators, and individuals would do well to treat AI collaboration skill as a form of literacy to be developed. It’s a blend of cognitive psychology and practical communication tactics – truly a multidisciplinary skill set.

In summary, prompt engineering was never really about the prompts – it’s about the people. The best prompt is useless if the person typing it isn’t thinking about how the AI will interpret it. On the other hand, a person who understands their AI partner can often get great results even with a plain, conversational prompt. The core of successful AI use is “empathic intelligence” and social reasoning. As the study demonstrated, when we cultivate this empathy for a non-human mind, we unlock AI’s ability to make us far more capable  . Human empathy becomes the catalyst that turns artificial intelligence into amplified collective intelligence. This collaborative mindset – treating AI as a partner and approaching it with patience, clarity, and understanding – is poised to be the defining intellectual skill of the AI era. By mastering it, we ensure that AI doesn’t replace human intelligence, but rather raises it to new heights, together.

Sources: Recent human–AI collaboration study findings and analyses     , as cited throughout.