💬 NLP

Artificial Phantasia: Emergent Mental Imagery in Large Language Models

This study demonstrates that Large Language Models can outperform humans in a task traditionally requiring pictorial mental imagery by relying solely on propositional representations, thereby providing evidence for an emergent "artificial phantasia" that challenges the cognitive science view that visual imagery necessitates pictorial formats.

Original authors: Morgan McCarty, Jorge Morales

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Morgan McCarty, Jorge Morales

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

The Big Question: Can You "See" Without Eyes?

Imagine you are asked to close your eyes and picture a capital letter D. Now, imagine turning that D 90 degrees to the left. Finally, imagine attaching a capital letter J to the bottom center of that shape.

If you open your eyes, what do you see? Most humans would say, "It looks like an umbrella."

For decades, cognitive scientists have believed that to solve this kind of puzzle, your brain needs a special "mental screen" or a "mind's eye" where you can actually see and manipulate pictures. They thought you couldn't do it just by thinking about words; you needed a visual picture in your head.

This paper asks a wild question: Can a computer do this without ever having "eyes" or a "mind's eye"?

The Experiment: A Mental Gym for AI and Humans

The researchers created a gym workout for the brain. They took a classic puzzle (the letter transformations) and invented 48 brand-new, never-before-seen versions of it. Because these puzzles were made up specifically for this study, the computers couldn't have memorized the answers from their training data.

They asked two groups to solve these puzzles:

100 Humans: People who had to listen to the instructions (to avoid seeing the letters on a screen) and imagine the shapes in their heads.
Large Language Models (LLMs): Advanced AI chatbots (like the latest versions of OpenAI's o3, GPT-5, and Google's Gemini 3 Pro). These models are trained on text, not on "seeing" pictures.

The Shocking Result: The AI "Saw" Better Than Humans

The results were surprising. The best AI models didn't just pass; they crushed the human average.

The Humans: Scored an average of 2.85 out of 5.
The Top AIs: Scored between 3.13 and 3.65.

The AI models solved these "visual" puzzles significantly better than the human participants. Even more interestingly, when the researchers forced the AI to actually draw the images at each step (using an image generator), the AI got worse at the task. This suggests that the AI wasn't relying on drawing pictures to solve the problem; it was doing something else entirely.

The "Magic Trick": How Did They Do It?

If the AI didn't use a "mind's eye" (a picture) and didn't use a drawing tool, how did it solve the puzzle?

The authors propose a concept they call "Artificial Phantasia."

Think of it like this:

The Human Way: You are an architect. You close your eyes and build a 3D model of a house in your mind. You walk around it, look at the windows, and then describe it.
The AI Way: You are a master translator. You never build a 3D model. Instead, you treat the instructions like a complex recipe written in code. You know that "Letter D rotated left" + "Letter J attached" = "Umbrella shape" purely through the logic of language.

The paper suggests these AI models are using propositional reasoning. They are manipulating words and concepts so precisely that they can figure out the final shape without ever "seeing" it. It's like solving a math problem by knowing the rules of algebra, rather than drawing the shapes on a piece of paper.

The "Reasoning" Factor

The researchers also tested what happens if they tell the AI to "think harder" (by giving it more time and "reasoning tokens" to process the steps).

Result: The more the AI was allowed to "think" and chain its reasoning together, the better it got.
Analogy: Imagine a detective solving a mystery. If you give them 5 minutes to look at the clues, they might guess. If you give them 5 hours to connect every clue logically, they solve the case perfectly. The AI solved the visual puzzle by connecting linguistic clues, not by looking at a picture.

What About "Blind" Humans?

The paper also mentions a group of people called aphantasics. These are humans who claim they cannot visualize images in their minds at all (they have no "mind's eye"). Surprisingly, studies show these people can still solve these visual puzzles almost as well as people who can visualize.

This paper adds fuel to that fire. It suggests that maybe you don't need a "mind's eye" to solve visual puzzles. You might just need a very strong "mind's logic." The AI, which has no mind's eye at all, proved that language and logic alone might be enough to do what we thought required a picture.

The Bottom Line

This study challenges the old idea that "visual thinking" requires a visual picture. It shows that advanced AI can perform tasks that look like "visual imagination" using only language and logic.

The Claim: AI has an emergent ability to "imagine" visually, but it's not doing it with pictures. It's doing it with words.
The Implication: We might need to rethink how we define "mental imagery." It might not be about seeing pictures in your head; it might just be about manipulating concepts logically.

The paper concludes that these AI models have discovered a new way to "see" the world—one that is purely linguistic, yet surprisingly effective at solving visual problems.

Technical Summary: Artificial Phantasia: Emergent Mental Imagery in Large Language Models

Problem Statement

The paper addresses a fundamental question in cognitive science: Can visual mental imagery be driven solely by language, or is it strictly dependent on pictorial (imagistic) representations? Traditional cognitive science posits that mental imagery operates on imagistic representations, a view supported by the "mind's eye" metaphor. However, the existence of aphantasics (individuals who lack conscious visual imagery but can still perform spatial and imagery-dependent tasks) challenges the necessity of pictorial formats.

The authors investigate whether Large Language Models (LLMs), which are primarily trained on propositional (linguistic) data and lack inherent visual processing capabilities in text-only modes, can solve tasks designed to be solvable exclusively via pictorial mental imagery. If LLMs succeed, it would suggest the existence of "artificial phantasia"—an emergent, non-pictorial form of mental imagery driven by propositional reasoning.

Methodology

Task Design

The study utilized an extended version of the classic "object reconstruction" task developed by Finke, Pinker, and Farah (1989).

Stimuli: The researchers created 60 instruction sets.
- 12 items were adapted from the original Finke et al. study (modified for clarity).
- 48 items were novel, bespoke examples created specifically for this study to ensure they were absent from the models' training data. These involved complex compositional transformations (e.g., rotating letters, attaching shapes, mirroring) where the final object (e.g., an umbrella, a computer mouse) could only be identified after all steps were mentally simulated.
Procedure: Subjects (both humans and LLMs) were asked to imagine a series of transformations and identify the resultant "image."
Human Baseline: 100 human participants completed the task via audio instructions (to prevent visual bias from on-screen text) and were assessed using the Vividness of Visual Imagery Questionnaire (VVIQ).
LLM Testing: The study tested a wide range of state-of-the-art models, including:
- Proprietary: OpenAI (o3, o3-Pro, o4-mini, GPT-5, GPT-4.1, ChatGPT-4o), Google (Gemini 3 Pro, 2.5 Pro, 2.0 Flash), and Anthropic (Claude Opus 4.1, Sonnet 4).
- Open-Weight: Qwen 3, Qwen 3 VL, DeepSeek R1, and gpt-oss-120b.
Experimental Variations:
- Image-Aided Generation: Some models were prompted to generate intermediate images using tools (e.g., GPT-image-1, Gemini's image generation) to test if pictorial aids improved performance.
- Reasoning Effort: The "reasoning token" allocation (minimal, low, medium, high) was varied for OpenAI models to test the impact of extended reasoning chains.
- Context: Models were tested in both single-context (all instructions in one prompt) and multiple-context (new context per instruction set) modes.

Evaluation

Scoring: A weighted scoring system was employed. 376 naive human graders rated the "reasonableness" of the model's/human's answer against a canonical target image on a 5-point scale.
Expert Validation: The authors provided independent "expert" grades. The final score was a weighted average of crowd-sourced and expert ratings.
Statistical Analysis: Mann-Whitney U tests and Wilcoxon Signed-Rank tests were used to compare model performance against the human baseline and against each other, with Bonferroni corrections for multiple comparisons.

Key Results

1. Superior Performance of Frontier Models

The most advanced reasoning models significantly outperformed the human baseline (mean human score: 2.85/5).

Gemini 3 Pro achieved the highest mean score (3.65, $p < .0001$ ).
GPT-5 (3.35, $p = .0004$ ) and the o3 family (o3-Pro: 3.31, $p < .0001$ ; o3: 3.13, $p = .0067$ ) also significantly surpassed humans.
Older or less capable models (e.g., ChatGPT-4o, Gemini 2.0 Flash, Qwen 3) performed significantly worse than humans or failed to reach the baseline.

2. Impact of Reasoning and Image Aids

Reasoning Tokens: Performance was positively correlated with reasoning effort. GPT-5 with "high" reasoning (3.35) significantly outperformed "minimal" reasoning (2.04).
Image-Aided Generation: Contrary to the hypothesis that visual aids would help, adding image generation capabilities (e.g., o3 with GPT-image-1) decreased performance compared to text-only reasoning. This suggests that the compositional generation errors inherent in current image models hindered the task, and that pure linguistic processing was more effective.
Context: In-context learning (single vs. multiple context) showed no significant difference in performance.

3. Open-Source Limitations

Open-weight models performed poorly, with none meeting the human baseline. Qwen 3 (mean 1.58) and Qwen 3 VL (mean 1.99) showed no significant difference, suggesting that the vision-language alignment step in Qwen 3 VL did not confer a significant advantage for this specific propositional task.

4. Uncertainty and Strategy

Uncertainty: Unlike humans, who frequently responded with "I don't know" (54 occurrences), LLMs never expressed uncertainty, often hallucinating answers even on the most difficult items.
Aphantasia Parallel: The study noted that the high performance of LLMs (which lack visual imagery) mirrors the performance of human aphantasics, who also succeed in imagery tasks without pictorial representations.

Key Contributions

Evidence for Artificial Phantasia: The paper provides empirical evidence that LLMs can solve complex mental imagery tasks using only propositional (linguistic) representations, challenging the view that pictorial formats are necessary for such cognition.
Novel Benchmark: The authors introduced 48 novel, non-training-data items for mental imagery tasks, creating a robust benchmark for evaluating the emergent reasoning capabilities of LLMs.
Re-evaluation of Mental Imagery Formats: The results suggest that "visual" mental imagery may not be a monolithic capacity but can be achieved through spatial reasoning within a propositional framework, supporting the distinction between object imagery and spatial imagery.
Methodological Insight: The study demonstrates that forcing models to generate intermediate images can degrade performance on compositional reasoning tasks, highlighting the current limitations of image generation in multi-step logical workflows.

Significance and Claims

The authors claim that their findings "reignite debate on the requirement for a pictorial format in mental imagery." They argue that the success of frontier LLMs (specifically reasoning models like o3, GPT-5, and Gemini 3 Pro) indicates an emergent cognitive capacity where language alone is sufficient to simulate and manipulate spatial relationships traditionally thought to require a "mind's eye."

The paper posits that these models do not rely on hidden pictorial representations but rather on propositional reasoning and linguistic embeddings to construct spatial layouts. This challenges the classical view that visual mental imagery is exclusively pictorial. Furthermore, the study suggests that the gap between human and artificial performance in this domain is narrowing, with the most advanced models now exceeding human capabilities in specific, structured imagery tasks.

The authors remain modest regarding the mechanism, acknowledging that while the models succeed, the exact internal representation (purely propositional vs. propositional with visio-linguistic priors) remains a subject for further investigation. They conclude that these results offer a new perspective for both the AI community, regarding the emergent capabilities of reasoning models, and the cognitive science community, regarding the nature of human mental imagery and the strategies used by aphantasics.