TLDR:
Large Language Models lack deep psychological awareness, impacting their effectiveness and limiting their potential in mental health, education, and customer service.
They miss out on understanding implicit psychological and cognitive signals in language, like intent or cognitive load.
Simply enlarging models won't address this; LLMs need an interpretative layer for extracting implicit psychological and cognitive signals.
This enhancement could lead to more empathetic AI, better mental health support, personalized education, and improved customer interactions.
Current Large Language Models (LLMs) lack a crucial component needed to reach their full potential in high-impact areas like mental health support, education, and human-quality customer service: Psychological Awareness. Despite their impressive capabilities, LLMs cannot interpret the deeper cognitive and psychological signals in language. Without this awareness, they fall short of the nuanced understanding, empathy, and insight required to be truly effective in these sectors.
You might be wondering: Is this really the case? And if so, how do we fix it? Let’s take a closer look:
People can phrase the same idea in different ways. The choice of how to phrase something is sometimes made consciously, and sometimes unconsciously, but those choices are a result of their psychological state and thought processes at the time they are expressing themselves. For example, someone might say:
“I’m sure this is the right decision.”
or
“This seems like the right decision.”
or
“I think this is probably the best option.”
All three express a similar idea, but the different phrasing reflects the speaker's varying levels of certainty, and underlying cognitive processes like doubt, openness to new information, and social positioning. Traditional LLMs will treat these sentences as near-equivalent, but an LLM with a psychological awareness layer would recognize the deeper psychological meaning that makes each statement unique.
Current LLMs excel at identifying statistical relationships within text, allowing them to process language and generate coherent responses, but their psychological awareness remains largely at the surface level, focused on what is being said rather than on the underlying psychological and cognitive processes that determine why people say things the specific way they do.
The language people use contains more than just content. It contains signals about why they think the way they do, their intent, how they reason, and how they make decisions. These cognitive factors influence how people communicate, but they exist outside of what LLMs can understand. This isn’t just theoretical – it’s a major limitation for AI in critical areas that need psychological awareness most, including mental health support, education, sales enablement, and even customer service.
The companies that win in the LLM space won’t be the ones with the most parameters; they’ll be the ones that develop the deepest, most accurate models of how people's underlying psychology is reflected in the language they use. This limitation can't be solved with fine-tuning models or adding more parameters, it's a core limitation of current transformer architectures that impacts their ability to decode the deeper psychological signals that are implicitly embedded in language, establish conceptual connections, and infer more than surface-level associations.
To achieve LLMs that are psychologically aware, AI companies must augment their transformer models with interpretation capabilities that can understand the cognitive and psychological processes that dictate why humans say things they way they do, not just what they say. This requires embedding a contextual layer of human psychology.
Why Transformer Models Don’t Recognize Intrinsic Psychological Signals
Transformers are the backbone of modern LLMs; they use self-attention mechanisms to weigh relationships between tokens within a given context window. This design allows for powerful understanding of word meanings (lexical features) and sentence structure (syntactic features), but it has limitations.
Here are some examples of the limitations that restrict LLMs from truly understanding people beyond the face value of the words they use:
1. LLMs Don’t Understand Underlying Intent:
LLMs are incredibly effective at recognizing patterns and correlations in text, they can predict the next word in a sentence, generate coherent responses, and mimic human-like writing. While LLMs can approximate intent based on surface patterns and fine-tuning, they lack the ability to consistently infer the deeper psychological processes that influence how they speak (cognitive intent). For example, a transformer model might recognize explicit sentiment (surface-level understanding), but it will not recognize the psychological processes that influence why they chose a particular phrasing to express that sentiment.
Here's an example that illustrates this shortcoming:
"That’s an interesting idea."
A literal interpretation of this statement is that it's positive.
In reality, this phrase can mean genuine interest or polite dismissal.
A LLM won’t pick up on this difference without a more nuanced understanding of how social cognition impacts a person's use of language.
This gap impacts applications where understanding a person’s real intent is critical, for example, in high-value business negotiations or in urgent customer support scenarios.
It's important to note that humans' cognition also relies on pattern recognition, but humans have the benefit of incorporating contextual information, background knowledge, and social cues to understand the deeper meaning behind communication. Current LLMs typically don't have the benefit of access to this information.
2. LLMs Miss the Psychological Reasoning Behind Shifts in Focus:
Humans don’t just state ideas, they move between them, they shift their focus, and they introduce and connect ideas in ways that reflect their cognitive and psychological states. But, traditional LLMs process language as discrete token sequences, missing the deeper psychological cues behind these shifts.
Here's an example that illustrates why they miss shifts in focus:
"The product launch is on track. We had a few delays, but the team is making progress."
The word “but” signals a subtle cognitive shift.
Depending on what follows, the speaker may be downplaying an issue or reframing uncertainty.
3. LLMs Lack Cognitive Trajectory Awareness:
Current LLMs can determine if a sentence expresses certainty or hesitation, but human certainty and hesitation don’t always exist in isolated utterances, they often emerge across longer narratives or sequences of thoughts.
Example:
"I think this approach might work." (indicates tentativeness)
"It aligns with previous results, so I’m confident." (the speaker's certainty is building over time)
While LLMs can process multiple sentences within a context window, they do not inherently track how a speaker’s confidence evolves over time or recognize 'implicit uncertainty" unless explicitly trained or prompted to do so. They also struggle with detecting subtle shifts in focus and intent over extended interactions, as their understanding is limited to pattern recognition rather than true psychological inference.
Why Bigger Models Won’t Solve The Problem:
Research, including findings presented at the NeurIPS conference, shows that beyond a certain scale, LLMs see minimal performance improvements, even while their computational costs continue to multiply. While increasing the size of models has led to improvements in explicit reasoning, it doesn't improve their ability to understand implicit cues like cognitive shifts over time. This limitation won’t be solved by bigger models; it requires a fundamentally new approach.
A Contextual Layer of Psychology Transforms Text into Psychological Data
To enable LLMs with greater psychological inference abilities, a new interpretation capability is needed. By embedding a layer that recognizes the deep cognitive and psychological signals implicit in language, LLMs can derive insights like shifts in focus, markers of intent, cognitive load, hesitation, certainty, and reasoning—insights that are not explicitly present in their training data. This contextual layer of human psychology goes beyond the capabilities of current transformer models, enabling a deeper understanding of the cognitve and psychological signals embedded in language, which is critical in high-stakes interactions.
One such approach involves utilizing tools like Receptiviti’s language psychology models to assist in extracting the implicit cognitive and psychological signals that are embedded in language, including hundreds of psychological signals that traditional sentiment analysis and syntactic parsing overlook, and which can then be incorporated into the LLM's understanding. This will enable LLMs to incorporate a metadata layer of the underlying cognitive and psychological processes at the root of communication, to complement their existing surface-level understanding of textual data.
This added layer of psychological inference has the potential to unlock new opportunities and applications for LLMs:
Next-Generation Conversational AI: AI that detects shifts in cognitive states, and adapts responses dynamically.
Mental Health Applications: AI that detects subtle cognitive signals, offering early indications of mental health concerns. Currently, LLM’s lack of cognitive understanding could lead to misinterpretation.
Highly Empathetic Customer Support: AI-powered customer support agents that can understand and respond to customer frustration and confusion, and underlying needs like reassurance will lead to considerable improvements in customer satisfaction and loyalty.
Enhanced Educational Support Tools: AI that adapts to students' cognitive progress and learning patterns, identifying gaps in understanding and offering highly personalized learning pathways.
Enhanced AI-Augmented Decision Making: Models that understand shifts in reasoning, hesitation, and certainty, improving strategic insights and decision-support for businesses and researchers.
Enhanced AI-Powered Coaching & Guidance: Systems that understand cognitive roadblocks, helping users navigate complex thought processes and decision-making.
The Competitive Differentiator: Psychologically Aware Models
LLMs have become commodities, and the race for cheaper, faster, and more efficient models is a zero-sum game. AI leaders now need to shift their focus to differentiating their models and finding profitable real-world applications to generate a return on the billions of dollars they’ve invested in them.
Differentiation can happen at the application layer, but a transformative opportunity exists for AI companies that differentiate at the interpretation layer.
Marc Benioff recently stated, "The real treasure of AI isn’t the UI or the model…what breathes life into AI is the data and metadata that describes the data to the model.”
Integrating a contextual layer of psychology will define the next generation of AI that understands people at a deeper, and far more human level. This begins with psychologically aware large language models.
