🤖 AI

The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans

This paper reveals that while large language models can identify grounding dimensions when explicitly queried, they exhibit a significant "grounding gap" compared to humans by relying excessively on word associations and failing to naturally recruit emotional and internal state concepts during free generation.

Original authors: Odysseas S. Chlapanis, Orfeas Menis Mastromichalakis, Christos H. Papadimitriou

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Odysseas S. Chlapanis, Orfeas Menis Mastromichalakis, Christos H. Papadimitriou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to explain the concept of "Justice" to two different groups: a room full of humans and a room full of advanced AI computers.

How Humans Do It:
When a human thinks of "justice," their mind doesn't just pull up a dictionary definition. Instead, it triggers a rich, messy web of experiences. They might remember the feeling of being treated unfairly (an internal emotion), the sight of a courtroom (a sensory detail), or the social drama of a family argument about who gets the last slice of pizza. Their understanding is "grounded" in their body, their feelings, and their life with other people.

How AI Does It:
When the AI thinks of "justice," it acts more like a super-fast librarian who has read every book ever written but has never lived a single day. It knows that "justice" is often found near words like "law," "court," "judge," and "fair." It builds its understanding almost entirely out of word associations. It knows the company the word keeps, but it doesn't seem to feel the weight of the concept.

The "Grounding Gap"

The researchers in this paper decided to test this difference. They asked 21 different AI models (including giants like GPT-5, Claude, and Gemini) to play a game: "List the first four things that come to your mind when you hear this abstract word." They did this for words like freedom, honor, and theory.

Here is what they found:

The Mismatch: When humans did this game, they naturally mixed in feelings, body sensations, and social situations. The AIs, however, mostly listed verbal associations (other words related to the topic) and almost completely ignored feelings and internal states.
The Score: The researchers measured how similar the AI's answers were to human answers. Even the smartest AI only scored about 0.37 out of 1.0 on similarity. In contrast, if you ask two different humans to play the game, they agree with each other about 0.97 out of 1.0.
- Analogy: It's like asking a human and a robot to describe the taste of an apple. The human says, "Sweet, crunchy, juicy, reminds me of autumn." The robot says, "Red, round, fruit, apple." They are talking about the same object, but they are describing it in completely different languages.
The "Robot Club": Interestingly, the AIs were much more similar to each other than they were to humans. They all seem to share the same "robot way" of thinking, which is very different from the human way.

The Twist: Can They Do It If Asked?

The researchers wondered: Do the AIs actually understand these feelings, or do they just not know how to bring them up?

To test this, they changed the game. Instead of asking the AI to "list things that come to mind," they asked specific questions like: "On a scale of 1 to 7, how much does this word relate to 'human emotion'?" or "How much does it relate to 'touch'?"

The Result: Suddenly, the AIs got much better! Their answers aligned much more closely with humans.

Analogy: It's like asking a person who is shy to "tell me a story about your childhood" (they might freeze or give a generic answer). But if you ask them, "Did you feel happy when you were 5?" (a specific, guided question), they can answer perfectly.

This suggests the AIs do have the "ingredients" for human-like understanding inside them, but when they are left to generate text freely, they don't know how to mix those ingredients. They default to the "word association" recipe because that's what they were trained to do.

Looking Inside the Machine

To see if this "shy" understanding was actually hiding inside the AI's brain, the researchers used a special tool called a Sparse Autoencoder (think of it as an X-ray machine for AI thoughts).

They found that the AI does have specific internal "switches" or "features" that light up when it thinks about emotions, social interactions, or sensory details.

Analogy: Imagine a piano. When the AI plays a song about "justice," it mostly hits the keys for "law" and "court." But if you look under the hood, you can see that the keys for "sadness" and "touch" are actually connected to wires inside the piano. They are there, but the AI isn't pressing them when it plays on its own.

When the researchers manually forced those "emotion" switches to stay on, the AI did start generating more emotional descriptions. This proved the capability exists, but it isn't being used naturally.

The Bottom Line

The paper concludes that current AI models are not grounded in the world the way humans are.

They are excellent at mimicking the patterns of human language.
They can understand concepts like "emotion" or "touch" if you point them directly at them.
But when they are left to think and speak freely, they rely too heavily on word-to-word connections and fail to tap into the feelings, body sensations, and social experiences that make human meaning so rich.

The researchers suggest that simply making the AI bigger or training it on more text won't fix this. To truly close the gap, we might need to fundamentally change how we build and train these models, perhaps giving them a way to "experience" the world rather than just reading about it.

Technical Summary: The Grounding Gap

Problem Statement

The paper investigates whether Large Language Models (LLMs) ground abstract concepts (e.g., justice, theory, availability) in a manner similar to humans. While human cognition posits that abstract concepts derive meaning from a web of sensorimotor experiences, emotions, and social contexts (grounded cognition), LLMs are trained solely on text. The central question is whether LLMs merely rely on statistical word associations or if they internally represent abstract concepts through the same experiential dimensions (sensorimotor, internal state, social) that structure human understanding.

Methodology

The authors employ a multi-faceted approach combining behavioral replication, rating experiments, and mechanistic interpretability:

Property-Generation Replication (Behavioral):
- The authors replicated two established cognitive science experiments (Harpaintner et al., 2018; Kelly et al., 2024) on 21 frontier and open-weight LLMs.
- Task: Models were prompted to freely generate properties, situations, or associations for abstract nouns.
- Analysis: Generated outputs were coded into grounding categories (e.g., Sensorimotor, Internal State, Social, Verbal Association) and compared against human baselines using Pearson correlation ( $r$ ).
- Scale: Experiments covered 293 abstract nouns (Exp 1) and 236 abstract nouns including emotion/non-emotion subsets (Exp 2).
Rating Experiment (Explicit Querying):
- To distinguish between a lack of knowledge of grounding dimensions versus a failure to recruit them, the authors replicated a rating study (Troche et al., 2017).
- Task: Models explicitly rated words on 14 grounding dimensions (e.g., Color, Emotion, Social, Time) using a Likert scale.
- Goal: Determine if models can recognize these dimensions when explicitly prompted, contrasting this with the spontaneous property generation task.
Mechanistic Interpretability (Sparse Autoencoders):
- The authors utilized Sparse Autoencoders (SAEs) on Gemma 3 models (4B and 12B parameters) to inspect internal features.
- Method: They identified latent features whose activations correlated with grounding dimensions (sensorimotor, internal state, social) during noun generation.
- Validation: They performed "feature steering" (clamping activations) to observe if artificially amplifying these features increased the generation of properties in the corresponding categories.

Key Results

1. The Grounding Gap in Property Generation

Low Alignment: Across all 21 models, there is a consistent and large gap between human and model grounding. No model achieved a Pearson correlation ( $r$ ) with human responses exceeding 0.37, whereas the human-to-human ceiling exceeds 0.9.
Systematic Bias: Models consistently over-produce Verbal Associations and under-produce Internal State & Emotion properties compared to humans.
Model Homogeneity: Models correlate more strongly with each other than with humans, suggesting a shared, non-human mode of abstract concept representation that persists across different architectures and providers.
Scaling Limits: Increasing model parameter count (from 4B to 120B) does not significantly close this gap; the mismatch is not a capability bottleneck solvable by scaling alone.

2. Alignment in Rating Tasks

High Explicit Alignment: In contrast to property generation, models perform significantly better when explicitly asked to rate words on grounding dimensions. The best models reach a mean $r \approx 0.76$ against a human ceiling of $\approx 0.89$ .
Implication: This indicates that LLMs possess the knowledge of grounding categories (they can recognize them when queried) but fail to recruit them spontaneously during free generation.

3. Mechanistic Evidence

Internal Features: SAE analysis identified specific internal features in Gemma models that correlate with grounding dimensions (e.g., "sensorimotor," "social," "internal state").
Feature Steering: Steering these features increased the generation of properties in their target categories (e.g., steering "Internal" features increased Internal State properties by +9.4%).
Limitation: These features are not perfectly isolated; they capture broader semantic patterns rather than clean, human-like grounding categories.

Key Contributions

Quantification of the Gap: The paper provides empirical evidence of a robust "grounding gap" where current LLMs fail to ground abstract concepts in a human-like way, regardless of model size or family.
Distinction of Capabilities: It demonstrates that the gap is not due to an inability to recognize grounding dimensions (as shown by the rating experiment) but rather a failure to spontaneously recruit these dimensions during generation.
Mechanistic Discovery: It identifies internal SAE features in Gemma models that track grounding dimensions, suggesting that the structural information exists within the model but is not behaviorally expressed in a human-like manner.

Significance and Claims

The paper concludes that the observed grounding gap is unlikely to be resolved through scaling or improved text-based training alone, as the reliance on verbal associations is a natural consequence of autoregressive training on linguistic context.

Modest Claim: The authors state that current LLMs "understand the basic grounding categories nearly—but not quite—as well as humans do," yet they "do not recruit them in a human-like way when generating freely."
Future Direction: The findings suggest that achieving true cognitive alignment with humans may require fundamental shifts in training techniques or architectures (e.g., embodied learning or novel training objectives) rather than incremental improvements to current paradigms.
Benchmarking: The authors propose that grounding alignment is a novel quantity currently invisible to standard testing benchmarks and warrants future attention.

The paper does not claim that LLMs possess consciousness or true understanding, but rather highlights a specific, systematic divergence in how abstract concepts are represented and utilized compared to human cognition.