AI Doesn’t Understand How We Think: Fixing It Is a Massive Opportunity
We trust AI to understand us. But its depth of human understanding is barely more advanced than that of a calculator. Let me explain:
AI systems’ understanding of the world comes from its training data, the majority of which is language data created by humans. The Large Language Models (LLMs) that conduct the literate functions within AI systems have a remarkable command of facts, semantic relationships, and patterns within the massive language datasets they are trained on, but their understanding of humans is significantly more simplistic than their ability to generate human-sounding language makes it seem.
True understanding of humans doesn’t come from facts, semantic relationships, or pattern matching. It requires a contextual layer that current AI systems lack, one that unlocks a deep understanding of humans, how they think, how they perceive, their motivations, perspectives, and values. Without this, AI cannot truly understand us.
If we can enable AI systems with a deep understanding of how humans think, they could offer far more genuine empathy and human-level support, anticipate mental health needs with incredible acuity, personalize our learning more effectively, and predict human behavior with far more accuracy than current systems. AI systems with a deep understanding of humans could also accelerate scientific discovery by revealing previously invisible relationships within human language data.
As the development and our reliance on AI systems accelerates, as a matter of safety and for our benefit, we must ensure that we are imbuing AI with the deepest possible understanding of the humans that it’s been developed to serve.
Discovering Signals in the Noise
In 2013 I met James W. Pennebaker, then Chair of Psychology at the University of Texas at Austin. Professor Pennebaker had spent much of his career sharpening his pencil on a deeper level of abstraction into the human mind though the lens of language. He was honing a science that could quantify how variations in word usage reflect shifts in mindset, cognitive processes, emotional states, and even physical health. His work identified the subtle ways human psychology dictates word choice, often unconsciously. He identified hundreds of linguistic signals in the form of patterns in the rates and frequencies of different word categories that correlate strongly with cognitive processes and psychological states like confidence, cognitive load, analytical thinking, and markers associated with mental states like depression and deception. These deeply human signals are significantly deeper, more meaningful, and more consequential than sentiment. They are hidden in language but are invisible without the right analytical lens.
Intriguingly, much of these cognitive and psychological signals are encoded in the language that is discarded by Natural Language Processing (NLP) as 'noise'. Why does NLP consider this language noise? Because humans are noisy communicators: Our communications don’t just reflect what we are talking about, they reflect our mindsets, our cognitive processes, and the psychological processes that are at play when we communicate, independent of what we are talking about.
For computational tasks that focus on semantic content or topic extraction, the noise in our communications doesn’t have much meaning. But if you’re looking to deeply understand humans, decoding the signal hidden in the noise is absolutely critical.
By dismissing the “noise”, traditional NLP techniques and the AI architectures that are heavily influenced by them, overlook these important signals. This has profound implications for AI. If AI systems are intended to assist humans, interface naturally with us, and learn from huge troves of human-generated language data, using techniques that only skim the semantic surface is insufficient. To ensure AI benefits humanity, it needs to understand the deep layer of human cognition and psychology encoded in our language.
Turning Signals into Advancements
In 2015, long before the broad rollout of large language models, Professor Pennebaker and I established Receptiviti. Our mission was to transform our ability to decode these cognitive and psychological signals into scalable technology that Receptiviti customers could use to solve real-world problems. We wanted to help organizations understand people more deeply, so they could better help the people they serve. We secured early funding, pulled together a team that includes the preeminent experts in the domain, and began building what would become the Receptiviti API. Receptiviti holds the exclusive commercial rights to this scientific framework, which has been validated across thousands of peer-reviewed studies. In the 10 years since, Receptiviti has collaborated with leading universities and researchers as we’ve continued to advance the science, all while partnering with customers across healthcare, technology, marketing, finance, government, and even professional sports teams.
Receptiviti’s science is backed by thousands of independent studies and cited in tens of thousands of academic publications. This body of evidence demonstrates the power of decoding deeply human signals in language to address challenges and drive advancements across mental health, hiring and management practices, marketing science, human performance, and more.
When LLMs started becoming widely available, it naturally raised questions about their ability to perceive these deeper signals. Could LLMs, that are so capable of generating human-like text, be effective at understanding cognitive and psychological signals embedded in language?
Signals that Shift AI Paradigms
Through our research, benchmarking, and feedback from the increasing number of Receptiviti customers who use LLMs, it became increasingly clear that LLMs are not nearly as effective as Receptiviti at decoding cognitive and psychological signals.
While LLMs excel at generating coherent text based on learned statistical patterns, their architectures are not designed to interpret the nuanced markers that reflect true human thought processes.
In contrast, Receptiviti employs a quantitative, empirically grounded methodology, that when used in conjunction with LLMs, reduces inaccuracies and common pitfalls of probabilistic inferences made by LLMs. Additionally, fine-tuning LLMs for psychological inference has proven prohibitively expensive for AI companies, and is prone to delivering shallow, unvalidated results, whereas integrating Receptiviti provides deep, reliable, empirically grounded insights at a fraction of the cost.
By revealing over 200 quantifiable signals that measure cognitive function, psychological disposition, and social dynamics, Receptiviti creates a robust foundational contextual layer of cognitive and psychological signals that provides AI with the ability to abstract a deeper dimension of understanding from language. As a foundational layer, Receptiviti can not only complement LLMs but can also overcome their inherent limitations in achieving deep human understanding.
LLMs can augment vector embeddings with these additional quantifiable signals, integrated during pre-processing to enhance input data or post-processing to refine outputs. This enables every new language input to be evaluated against empirically validated cognitive and psychological measures, providing a deeper contextual understanding that standard models lack, at every stage.
The resulting benefits are twofold: uncovering hidden linguistic relationships and establishing a consistent framework for measuring cognitive and psychological processes. The potential these signals hold for new discoveries cannot be overstated. They offer AI systems new capabilities to model and anticipate human behaviors with far greater accuracy, and present the opportunity for a paradigm shift in how AI understands and interacts with humans.
Consider just a few of a countless number of practical, real-world implications when this foundational layer is applied to AI systems:
Predicting and Preventing Crises: In mental health, the ability to detect subtle linguistic markers predictive of depression, anxiety, or suicidal ideation enables a shift from reactive treatment to proactive intervention. Integrated into large-scale health platforms, AI equipped with Receptiviti's foundational contextual layer of cognitive and psychological signals could monitor anonymized communications and identify individuals exhibiting early warning signs often missed by traditional methods or surface-level AI analysis. This isn't pattern matching; it's the ability to understand the cognitive shifts that precede a crisis, and it has the potential to transform AI into a lifesaving tool, and one that also advances our scientific understanding of mental health and well-being.
Decoding Unspoken Needs and Intent: In domains requiring nuanced human interaction, such as customer service or negotiation, Receptiviti allows AI to interpret unspoken needs, values, and cognitive states. An AI-enabled customer support system can move beyond keyword spotting, sentiment and simple emotion detect to a more sophisticated understanding of a customers's underlying psychology, cognitive load, or specific decision-making styles, and prompt responses with greater situational awareness, empathy, clarity, and effectiveness. This capability is crucial for developing AI systems that can collaborate, reason, and interact more naturally and effectively with humans, and it represents a significant step towards artificial general intelligence.
Breakthrough Opportunities for Science and Society
The power of Receptiviti’s contextual layer of cognitive and psychological signals extends far beyond these two examples. Integrating Receptiviti's layer with AI can unlock entirely new opportunities for scientific inquiry and advancement:
Revolutionizing Psychological and Cognitive Science: Receptiviti provides empirical, scalable data for testing theories about the mind, mapping linguistic features to cognitive processes, and potentially uncovering new models of human thought and emotion.
Improving Mental Health Understanding and Treatment: Receptiviti enables the ability to identify at-risk populations, personalize therapeutic interventions based on individuals’ cognitive styles, and objectively and quantitively measure treatment efficacy through cognitive and psychological linguistic markers.
Enabling More Effective and Adaptive Education: Receptiviti enables AI-powered learning platforms to adapt not just to knowledge gaps, but to a student's motivation or cognitive load, enabling deeper engagement and enhancing cognitive development.
Making Human Capital Decisions More Scientific: Receptiviti makes it possible to move beyond resumes and interviews to assess core cognitive and collaborative traits through language analysis, enabling data-driven decisions in hiring, team composition, and leadership development based on predictive indicators of people’s performance and potential.
Enhancing Leadership and Organizational Science: Receptiviti enables the identification of linguistic markers of effective leadership, team cohesion, or psychological safety, informing strategies to improve workforce effectiveness, reduce burnout, and boost productivity based on fundamental human dynamics.
Driving Discoveries in Behavioral Finance and Economics: Receptiviti enables analysis of corporate communications, market commentary, or earnings call transcripts for signals of deception, risk tolerance, or predictive behavioral patterns that have been shown to influence economic outcomes.
Making Persuasion and Communication More Scientific: Receptiviti enables marketing and public health campaigns to go beyond demographics to target underlying psychological needs, values, and cognitive frameworks, increasing the effectiveness of communications and more effectively encouraging positive behavioral change.
Advancing Cognitive Computing and Neuroscience: Receptiviti provides rich valid data on how cognitive and emotional states are encoded in language, which can power the development of AI systems that more closely mimic human cognitive architectures and offer new insights into neuroscience.
Why Understanding Cognition and Psychology is Essential for AI's Future
As a foundational layer, Receptiviti doesn't just enhance existing models; it presents a fundamental improvement to AI's ability to become more insightful, predictive, and aligned with the complexities of human cognition. It provides AI the signals it needs to anticipate human needs and behaviors with greater accuracy and nuance, and the potential to drive transformative scientific discoveries that improve lives and advance humanity.
As we increasingly rely on AI across all facets of our lives, we must ensure that AI has the foundation to truly understand us. Language is the primary fuel for LLMs, yet a critical component in understanding humans has been missing. The foundational contextual layer of cognitive and psychological signals required to unlock that understanding from language presents an opportunity to transform AI into truly human-centric tools capable of advancing scientific frontiers and solving complex societal challenges.
For organizations aiming to lead the next generation of AI and potentially AGI, embracing this deeper layer of human understanding is key to unlocking AI's true potential and to ensuring its beneficial integration into society.
If you're working on foundational models or building applications or agents that depend on a deeper understanding of human thinking or behavior, I'd love to connect.