For years, computational psycholinguistics operated under an intuitive assumption known as the Quality-Power hypothesis: as a language model gets better at predicting the next word in a text (minimizing perplexity), it becomes a more accurate model of human language processing. Early foundational work by Stefan L. Frank et al. in Word surprisal predicts N400 amplitude during reading demonstrated that a model's statistical "surprisal" correlates beautifully with human cognitive signatures, such as electrical brain potentials. More recently, Ethan Wilcox et al. in Language Model Quality Correlates with Psychometric Predictive Power in Multiple Languages validated this trend across 13 diverse languages, showing that higher textual accuracy generally aligns a model's behavior with human gaze durations.
However, as models scale into massive modern architectures, this relationship breaks down. We hit a point of inverse scaling, where further optimization for text prediction actually makes a model less like a human reader. This perplexity-behavior dissociation reveals that the statistical objectives driving modern AI are fundamentally diverging from the biological and cognitive constraints of human language comprehension.
The Peak of Human Likeness
If scaling eventually degrades a model's cognitive plausibility, where is the turning point? Byung-Doh Oh et al. identified a distinct "sweet spot" in training volume in Transformer-Based Language Model Surprisal Predicts Human Reading Times Best with About Two Billion Training Tokens. The authors found that language models predict human reading times most accurately when they are trained on roughly two billion tokens. Beyond this threshold, as perplexity continues to drop, the model's predictive alignment with human reading behavior steadily declines.
This decline is not an artifact of data contamination or memorization. In a subsequent study, Byung-Doh Oh et al. demonstrated in The Inverse Scaling Effect of Pre-Trained Language Model Surprisal Is Not Due to Data Leakage that the inverse scaling effect persists even when training data is strictly filtered to remove overlaps with test stimuli.
Instead, the divergence seems tied to structural changes that occur as a model optimizes. Ethan Wilcox et al. showed in Language Models Grow Less Humanlike beyond Phase Transition that this performance drop coincides with a specific "phase transition" during pretraining—the sudden emergence of highly specialized attention circuits. Once these circuits form, the model begins optimizing for statistical patterns that move its internal probability distributions away from human-like expectations.
Why Superhuman Prediction Diverges from Human Cognition
To understand why a better statistical predictor becomes a worse cognitive model, we have to look at how humans and models handle context and frequency. As Buck Shlegeris et al. demonstrated in Language Models Are Better Than Humans at Next-token Prediction, modern language models achieve "superhuman" performance at next-token prediction by mastering long-range dependencies and obscure statistical nuances that the human brain ignores.
Context-Frequency Decoupling
Human reading speeds are heavily reliant on heuristic shortcuts. When reading real-time text, our brains lean on components of context that correlate strongly with general word frequency. In On the Role of Context in Reading Time Prediction, Andreas Opedal et al. explained that as language models grow larger and more accurate, they become exceptionally good at completely decoupling contextual information from global frequency. By learning to ignore these frequency correlations to achieve near-perfect contextual prediction, the models abandon the very heuristics that dictate human reading speeds.
Statistical vs. Abstract Generalization
Human language processing is constrained by memory and a biological drive for communicative economy. While humans generalize language using abstract categories or type-based analogies, large models remain heavily driven by raw token statistics. For instance, Valentin Hofmann et al. found in Derivational Morphology Reveals Analogical Generalization in Large Language Models that LLMs rely on raw token frequency rather than human-like abstract linguistic categories when generalizing.
Because of this reliance on raw statistics, models fail to capture the sheer "mental whiplash" humans experience when encountering structural anomalies. When models process complex "garden-path" sentences, Suhas Arehalli et al. showed in Syntactic Surprisal From Neural Models Predicts, But Underestimates, Human Processing Difficulty From Syntactic Ambiguities that their probabilistic calculations drastically underestimate the actual human reading slowdown—often by an order of magnitude. The models calculate a smooth statistical probability, missing the fact that a human reader must stop and actively reconstruct a coherent mental representation.
The Cost of Alignment and Optimization
The gap between AI performance and cognitive plausibility widens even further when we try to make models more useful or more human-like through fine-tuning.
Techniques like instruction tuning and reinforcement learning from human feedback (RLHF) are designed to make models helpful assistants. Yet, Tatsuki Kuribayashi et al. discovered in Psychometric Predictive Power of Large Language Models that instruction tuning acts as a cognitive degradant. By shifting the model's probability distributions toward stylized, helpful responses, alignment destroys the naturalistic, word-by-word expectations that mirror human reading behavior.
Attempts to force models to be more human-like also reveal a fundamental trade-off. When Samuel Kiegeland et al. explicitly fine-tuned language models to match human reading times in Reverse-Engineering the Reader, the models' performance on standard NLP tasks and next-word prediction significantly degraded. The internal representations optimized for human cognitive constraints are fundamentally at odds with the representations optimized for raw statistical efficiency.
Where Human Likeness Hides
Does this mean large language models are entirely un-human? Not necessarily. The problem may lie in where we are looking.
[Input Text] ---> [Internal Layers] ---> [Final Output Layer]
| |
Aligns with human brain Diverges due to
activity & reading times superhuman optimization
In Large Language Models Are Human-Like Internally, Tatsuki Kuribayashi et al. discovered that while the final output layers of large models diverge from humans due to over-optimization for next-token prediction, their internal layers remain highly aligned with human reading patterns and brain activity. The core linguistic representations that mirror human cognition are preserved deep within the network, but they are overwritten at the finish line by the model's superhuman predictive drive.