Support for “genuine understanding” comes from mechanistic work suggesting that LLMs can build internal representations grounded in perceptual structure, despite operating on text alone. Related analyses also report introspective-like access to internal states and evidence that multi-step planning can occur before generating final responses—for example, when producing structured outputs such as poems.

Skeptics counter with a different diagnostic: brittleness. They argue that small changes—paraphrases that preserve meaning, or irrelevant additions—can trigger sharp accuracy losses. Moreover, LLMs often struggle with logic tasks that lack training-like templates, and chain-of-thought benefits can appear to mix reasoning with memorized fragments from training data.

The article proposes experientialism as a unifying framework to dissolve the dichotomy. Instead of assuming that understanding must mirror either objective reality or unconstrained imagination, experientialism views cognition as arising from interactions between an agent and its environment. Under an extended version of this idea, meaning is constructed by a system’s internal representational machinery while being constrained by the environment it learns from.

This view aligns with Bayesian cognition: an agent’s internal model is separated from the outside world by an interaction boundary (a Markov blanket). Lacking direct access, it must infer hidden causes of sensory-like inputs. Crucially, the internal model’s structure depends both on environmental structure and on the representational capacity of the agent itself.

Applying this to LLMs suggests a comparative but non-identical picture of cognition. For humans, embodied sensorimotor experience supports embodied cognition. For LLMs, the Transformer’s training corpus plays an analogous role, shaping how internal representations are constructed—an insight that reframes the question from “Does it understand?” to “How does it construct meaning under its constraints?”

A key empirical example comes from recent work on temporal cognition. In a similarity-judgment task spanning years from 1525 to 2524, large models reportedly develop a subjective temporal reference point and follow a Weber–Fechner-like scaling pattern. Neural and representational analyses are described as showing logarithmic temporal coding, hierarchical abstraction relative to the reference, and non-linear temporal structure embedded in the training data.

Taken together, the proposal reframes LLMs as “the other mind”: not a lesser version of human cognition, but a different cognitive construction process. The danger, the article warns, is treating model outputs as direct surrogates for human behavioral data without verifying cognitive equivalence.

Instead, it advocates task-based cognitive comparisons that test convergence and divergence directly. As LLMs enter high-stakes domains, machine experientialism argues for non-human-centric cognitive science—understanding how internal reality is built, predicting where divergence will emerge, and intervening to mitigate risks.

Subject of Research: Distinguishing human projection from machine cognition in large language models

Article Title: Understanding large language models demands distinguishing human projection from machine cognition.

Article References: Li, L., Teng, Y., Wang, Y. et al. Understanding large language models demands distinguishing human projection from machine cognition. Commun Psychol 4, 108 (2026). https://doi.org/10.1038/s44271-026-00508-6

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s44271-026-00508-6

Keywords: anthropomorphism; experientialism; Bayesian cognition; Markov blanket; mechanistic interpretability; temporal cognition; chain-of-thought; cognitive comparison; AI safety alignment; internal representation

Social perception—the ability to intuitively understand others’ emotions, intentions, and behaviors—is a fundamental human skill central to communication and collaboration. Traditionally considered a uniquely human faculty reliant on empathy and contextual understanding, deciphering social cues presented a profound challenge for AI systems. However, the surge in multimodal models, capable of processing both text and visual input, has presented new opportunities to bridge this divide. The researchers at Turku set out to rigorously test whether GPT-4V’s assessments of social features concur with those made by human observers.

The study involved the model evaluating 138 distinct social attributes extracted from a diverse set of images and video clips. These attributes encompassed a broad spectrum: from minute facial microexpressions and nuanced body language, to more abstract social interaction parameters such as cooperation, hostility, and social engagement. To establish a baseline for comparison, human evaluations totaling over 2,000 assessments were gathered from a large pool of participants, providing a robust dataset against which GPT-4V’s performance was measured.

Remarkably, GPT-4V’s analyses demonstrated a degree of consistency and reliability that rivaled—and in many cases surpassed—the evaluations of individual human observers. The AI’s ratings were not only aligned with general human consensus but also exhibited less variance compared to evaluations from single participants. Although collective assessments from multiple humans remained the gold standard in accuracy, the AI’s stable and reproducible judgments underscore its potential as a dependable proxy in social behavioral analysis.

This breakthrough extends far beyond academic curiosity. By integrating AI-powered social perception into neuroscience research, the study illustrates a practical pathway to accelerate the mapping of social cognition networks in the brain. Functional brain imaging experiments often require exhaustive coding of social content in stimuli, a process traditionally dependent on labor-intensive human annotation. GPT-4V’s ability to swiftly and accurately perform these evaluations can dramatically expedite data processing pipelines, enabling high-throughput experiments that were once prohibitively resource-intensive.

Dr. Severi Santavirta, the postdoctoral researcher spearheading the study, highlights the implications: “Where human evaluations demanded more than 10,000 collective work hours from over 2,000 individuals, the AI accomplished equivalent assessments within mere hours. This scalability can revolutionize social neuroscience by reducing costs, limiting human labor, and maintaining high data quality.”

On a neurological level, the study revealed that brain activation patterns correlated strongly with the social feature evaluations provided by both humans and GPT-4V. When participants viewed social scenes, functional MRI scans indicated that the brain networks engaged during these observations were nearly identical whether the underlying social features were annotated by human raters or the AI model. This remarkable convergence lends neuroscientific credence to the fidelity of AI-based social perception.

The broader societal potential of this research is enormous. Automatic, continuous social situation analysis could find immediate utility in healthcare settings, where monitoring patient well-being and behavioral cues is critical yet often limited by staffing constraints. AI systems could serve as tireless assistants that parse complex social environments around the clock, alerting caregivers to subtle changes indicating distress or recovery.

Moreover, commercial sectors such as marketing stand to benefit considerably. Understanding audience reactions and the perceived social nuances of audiovisual content can reshape how campaigns are designed and targeted. AI-driven analytics might predict viewer engagement or detect unfavorable social signals, enabling brands to optimize content with unprecedented precision.

Security applications present another promising frontier. Surveillance systems augmented with AI capable of interpreting social interactions could foresee potential threats by identifying patterns of hostility or abnormal crowd behavior well before human operators might recognize them. This anticipatory capability promises to enhance public safety without infringing on privacy through automated non-intrusive analysis.

Despite these advances, the researchers caution that AI should not be viewed as a wholesale replacement for human judgment but rather as a powerful complement. The fatigue-free consistency and rapid throughput of AI can enable human experts to focus on confirming critical findings, troubleshooting ambiguous cases, and applying contextual expertise not yet replicable by machines.

Technically, this study pushes the boundary of natural language processing fused with computer vision, highlighting how large multimodal models like GPT-4V capture complex semantic and social information embedded in visual media. This paradigm shift reflects an evolution from simple object recognition toward deep social cognition simulation, enabling machines to understand human experiences more profoundly.

The work also raises intriguing questions about how AI’s internal representation of social features compares to the neural encoding in the human brain, inviting future interdisciplinary research teams to explore the computational neuroscience of social perception. Such inquiries could eventually lead to AI systems that not only interpret but also simulate social behaviors in ways beneficial for education, therapy, and human-computer interaction design.

As these models continue to evolve, the fusion of artificial and human intelligence could forge a new era of social insight, combining vast computational resources with uniquely human empathy. The University of Turku’s study exemplifies this transformative potential, setting a benchmark for AI’s role in decoding the intricacies of human social life.

Subject of Research: People
Article Title: GPT-4V shows human-like social perceptual capabilities at phenomenological and neural levels
News Publication Date: 12-Aug-2025
Web References: 10.1162/IMAG.a.134
Keywords: Artificial intelligence, social perception, GPT-4V, multimodal models, neuroscience, brain imaging, social interaction analysis, functional MRI, human behavior, AI evaluation consistency, computational social cognition

In striving to uncover the limitations of AI, researchers focused on multimodal large language models (MLLMs) capable of processing both text and images. These models are often showcased as leading the frontier of AI research, yet their struggle with interpreting simple visual cues raises important questions about the practical applications of such technology in real-world scenarios. The team aimed to assess whether these AI systems could accurately interpret images depicting clocks and calendars, providing insight into the cognitive skills AI systems have not yet achieved.

To conduct their analysis, the researchers presented various clock designs to the AI systems, ranging from simple analog styles to more complex variations featuring Roman numerals, different hand styles, and intricate color schemes. Surprisingly, the results were disappointing. At most, the AI systems could accurately determine the positions of clock hands less than 25% of the time. The frequency of errors particularly soared with the introduction of more complex clock designs incorporating Roman numerals or uniquely styled hands. The research team noted that the difficulties encountered by AI in discerning hand positions and interpreting angles indicate that there are fundamental flaws in their detection algorithms and spatial reasoning capabilities.

Additionally, the researchers expanded their investigation to cover calendar-based questions that require an understanding of temporal relationships, such as identifying holidays and calculating dates across various timeframes. Their findings stressed that even the top-performing AI model miscalculated date-related queries one-fifth of the time. This error rate underscores the necessity for AI models to possess a more intuitive grasp of time and dates, features that are insignificant hurdles for most human beings but remain formidable challenges for AI.

These revelations carry significant implications for the future integration of AI in essential, time-sensitive applications. The study authors suggest that if these capabilities are not addressed, the full potential of AI technologies might never be realized in fields such as scheduling assistants, automation tasks, or assistive technologies designed for individuals with visual impairments. Instead, the ongoing shortcomings could hinder the efficiency and effectiveness expected from AI systems, leading to frustrations in both personal and professional contexts.

The research sheds light on a paradox in AI development: while there is a pronounced emphasis on advancing complex reasoning capabilities, many AI systems continue to falter in executing everyday tasks that humans navigate effortlessly. As robotic solutions and smart assistants become more prevalent in various sectors, the findings indicate an urgent need for focused research on the fundamental skills that remain underdeveloped within these systems. Ultimately, addressing these deficiencies is paramount for the viable deployment of AI in solutions that users can rely on for daily tasks.

The research team underscored the importance of these findings, asserting that the gap in basic cognitive skills reflects a pressing issue within AI development. Rohit Saxena, a lead researcher affiliated with the University of Edinburgh’s School of Informatics, emphasized that most individuals can read a clock and utilize calendars from a young age, thereby indicating the simplicity of the skills AI has yet to master. His sentiments were mirrored by fellow researcher Aryo Gema, who highlighted the irony of advancing sophisticated reasoning capabilities while neglecting essential functions required for simple, everyday interactions with technology.

Given the essential role that time management plays in modern society, this study elucidates the necessity for further exploration into the realms of AI’s cognitive architecture. The researchers indicate that by developing AI systems with enhanced temporal reasoning abilities, new avenues could open up in various applications, including automated scheduling tools and intelligent assistant technologies that cater to the visually impaired. The improvements in these areas could represent a significant leap forward in making AI more accessible and useful for individuals who rely on these technologies to navigate their daily realities effectively.

As the world of AI continues to evolve, the challenges highlighted by this research call for a paradigm shift in both expectations and development priorities. Rather than simply chasing higher degrees of sophistication in reasoning and pattern recognition, there is a profound need for researchers and developers to return to fundamentals. By equipping AI systems with the capabilities to accurately interpret time and manage dates, we may progress towards a future where AI can be seamlessly integrated into the fabric of daily life.

In conclusion, the study serves as a wake-up call to the AI research community. The integrity of AI systems that are deployed in time-sensitive environments heavily relies on their ability to excel not only in complex computational tasks but also in fundamental cognitive functions that most humans take for granted. The need for rectifying these issues is more pressing than ever, especially as AI continues to grow in prevalence and importance across multiple domains. The results of this inquiry remind us that even as we celebrate the advancements in AI, we must not overlook the vital groundwork that must be laid to ensure reliable, effective, and user-friendly systems in the future.

Subject of Research: Limitations of AI in Time and Date Understanding
Article Title: Advanced AI Struggles with Basic Tasks: Telling Time and Understanding Dates
News Publication Date: October 2023
Web References: University of Edinburgh Press Release
References: Peer-reviewed publication presented at ICLR 2025 workshop
Image Credits: University of Edinburgh

Keywords: Artificial intelligence, Multimodal large language models, Time interpretation, Calendar understanding, Cognitive skills, Robotics, Scheduling assistants, Visual impairments.