LLM’s don’t reason Part 2 – The Jellybean Test.

A Jar of Jelly Beans and a Question About Intelligence.

In October 2026, technology YouTuber Mrwhosetheboss presented several artificial intelligence systems with a deceptively simple challenge.

A photograph showed a transparent jar containing jelly beans.

The question was straightforward:

“How many jelly beans are in this jar?”

ChatGPT reportedly estimated 1,128. Google’s Gemini suggested 1,850. Other systems produced different estimates.

The actual count, according to the video’s reported result, was 2,143.

None answered correctly.

How could systems capable of solving mathematical problems, generating complex computer programs and interpreting scientific literature struggle with such a simple question?

The answer lies partly in a distinction that Nobel laureate Daniel Kahneman popularised in Thinking, Fast and Slow.

Human cognition can operate through two broad modes: fast, intuitive judgments and slower, more deliberate reasoning.

Psychologists call these System 1 and System 2.

Interestingly, recent artificial intelligence research increasingly uses the same distinction to describe different forms of AI computation.

But do humans and artificial intelligence actually reason in similar ways?

And what went wrong with the jelly beans?

Part I: How Human Beings Process a Prompt

Imagine being shown the jar and asked to estimate its contents.

Your brain does not necessarily begin by calculating anything.

Instead, it uses several interconnected systems supporting perception, language, memory, numerical cognition and decision-making.

Step 1: Perception — What Am I Looking At?

Light entering your eyes is converted into neural signals.

Visual processing identifies the glass container, the coloured objects inside it and their spatial arrangement.

Your brain recognises a jar filled with jelly beans.

But recognition is not counting.

Some beans are visible. Others are hidden behind them. The brain must distinguish the observed objects from the larger collection that is only partially visible.

Human beings possess an approximate number system that supports intuitive judgments about quantities.

Research by Cicchini, Anobile and Burr (2016) provides evidence that human perception can represent numerosity rather than relying exclusively on visual properties such as density or area.

However, that ability becomes less reliable when objects are crowded, overlapping or numerous.

Step 2: Language — What Is Being Asked?

You understand the words “how many” and identify the required task.

This involves language-processing networks that are partly distinct from the networks supporting higher-level problem-solving.

Fedorenko, Ivanova and Regev’s 2024 neuroscience review emphasises the existence of a core language network, especially involving left frontal and temporal regions.

Importantly, understanding a question and solving it are not the same operation.

You can perfectly understand what someone wants while lacking the information necessary to provide an accurate answer.

Step 3: System 1 — The Fast Guess

You look at the jar and immediately think:

“Perhaps a thousand?”

That is an example of System 1-like processing.

It is fast, largely automatic and based on perceptual impressions, learned associations and previous experience.

You may remember similar jars or intuitively associate the jar’s apparent size with a particular quantity.

The response arrives without an explicit calculation.

System 1 is extraordinarily useful. Human beings could not function efficiently if every ordinary decision required prolonged deliberation.

But rapid judgments can be affected by misleading cues.

A large-looking jar may create an exaggerated impression of quantity. Bright colours, bean density and the size of the photograph may influence the estimate.

System 1 produces a candidate answer. It does not necessarily establish that the candidate is correct.

Step 4: System 2 — The Reasoned Answer

Now imagine someone offers a substantial prize for a more accurate estimate.

You reconsider your initial impression.

Instead of guessing immediately, you ask:

How large is the jar? How much of it is filled? How large is an individual jelly bean? How much empty space exists between the beans?

This is closer to System 2 processing.

You are deliberately examining assumptions, identifying missing information and choosing a method.

Psychologists Evans and Stanovich (2013) distinguish relatively autonomous Type 1 processing from Type 2 processes that support hypothetical reasoning and depend heavily on working memory.

System 2 is effortful and goal-directed.

It allows us to evaluate a first impression, consider alternatives and revise an answer.

However, slower thinking is not automatically more accurate. A person can deliberate extensively while relying on incorrect assumptions.

The benefit comes from using better evidence and methods, not simply spending longer thinking.

Step 5: The Actual Calculation — In Plain English

The jar can be treated as a problem of physical volume.

First, estimate the internal volume of the jar.

Next, determine how much of that volume is occupied by the region containing jelly beans.

Then estimate the average volume of a single bean.

Finally, account for the empty spaces between the beans, because irregular objects do not pack together perfectly.

This is a physically meaningful estimation method.

But if the jar’s dimensions or the average bean size are unknown, the answer remains approximate.

A careful person would communicate a plausible range rather than imply that an exact count has been observed.

This illustrates the value of moving from an immediate impression to an evidence-based estimate.

Part II: How ChatGPT Processes the Same Prompt

Now consider the same question submitted to a multimodal AI system.

Its processing architecture differs fundamentally from that of the human brain.

Step 1: Receiving the Input

The application receives the image and accompanying question.

It may combine these with system instructions, previous conversation and other accessible information.

This creates the context within which the model must respond.

Step 2: Numerical Representation

The question is converted into tokens.

The image is processed into numerical visual representations.

These allow the model to identify relationships between visual information and language.

But converting an image into numerical representations does not guarantee that every jelly bean has been identified or counted.

Step 3: Attention and Contextual Processing

Transformer-based models use learned attention mechanisms and neural-network computations to combine information.

The system can associate the phrase “how many” with the collection of jelly beans.

It may identify the container, infer its approximate size and relate the image to patterns encountered during training.

Its model architecture and learned parameters provide relatively stable computational machinery.

The current prompt determines which information is processed and how the machinery is applied.

Step 4: Generating the Answer

A conventional autoregressive language model repeatedly predicts and selects output tokens, conditioned on the available context.

These computations can support sophisticated reasoning.

However, token generation is not inherently a measurement procedure.

If the model produces “1,128”, that number may reflect a visual estimate, a learned heuristic, an internal calculation or some combination of processes.

Without inspecting the actual execution, we cannot know which mechanism produced that specific answer.

This is where the comparison with System 1 becomes useful.

Part III: Does AI Have System 1 and System 2?

The short answer is: AI can exhibit System 1-like and System 2-like behaviour, but these are functional analogies, not evidence of two human-like cognitive systems inside the model.

AI’s System 1-Like Behaviour

When a model receives a question and generates a response directly, without an explicit verification procedure, its behaviour can resemble intuitive human judgment.

It uses learned information to produce an answer quickly.

For a straightforward question such as “What is the capital of France?”, this may work extremely well.

For visually estimating thousands of partially hidden objects, it may be unreliable.

The model can generate a plausible number without establishing whether the quantity is correct.

In that sense, the jelly-bean response may resemble a fast human guess.

But fast token generation and human intuition are not neurologically or computationally identical.

AI’s System 2-Like Behaviour

Modern reasoning-oriented AI systems can allocate additional computation to complex problems.

They may perform multiple reasoning steps, compare alternatives, use tools, execute calculations or verify intermediate outputs.

This resembles aspects of System 2.

A 2026 IEEE survey, From System 1 to System 2: A Survey of Reasoning Large Language Models, reviews how reasoning-oriented systems have developed capabilities associated with slower, structured problem-solving.

A 2025 paper in npj Artificial Intelligence, Fast, Slow, and Metacognitive Thinking in AI, also describes architectures that allocate different computational resources according to problem difficulty.

For the jelly-bean task, System 2-like AI behaviour could involve identifying the container’s geometry, estimating physical volumes, recognising uncertainty and using a calculation tool.

But even sophisticated reasoning cannot recover measurements absent from the photograph.

So Which System Does AI Resemble More?

It depends on the model, task and execution method.

A conventional model producing an immediate answer without external verification is often more analogous to System 1.

A reasoning-oriented model that systematically evaluates possibilities, uses tools and checks its results is more analogous to System 2.

But these are not mutually exclusive categories.

A single Transformer can produce both heuristic and more deliberate responses depending on the prompt and computational process.

Research published in PNAS Nexus in 2024 demonstrated that language models can exhibit both content-sensitive reasoning biases and more consistent reasoning under different prompting conditions.

A 2025 review in Nature Reviews Psychology similarly concluded that language models can display both System 1-like and System 2-like decision-making behaviour.

The distinction is therefore better understood as a continuum of computational strategies than as two separate artificial minds.

Part IV: Why Did the Jelly-Bean Test Fail?

The published research suggests several possible causes.

First, the photograph contained incomplete information. A two-dimensional image cannot necessarily reveal every object occupying a three-dimensional container.

Second, visual counting remains difficult for multimodal models. Testolin, Hou and Zorzi’s 2025 research in PLOS ONE identified systematic enumeration errors, even on comparatively simple stimuli.

Third, numerical perception may be poorly grounded. The 2025 VisNumBench study evaluated 17 multimodal models and found performance substantially below human levels on its number-sense tasks.

Fourth, partial occlusion introduces additional errors. The CAPTURe benchmark found that tested vision-language models struggled to infer object counts when items were hidden.

Fifth, the system may not have verified its estimate. Producing a specific number is not equivalent to demonstrating that the quantity has been measured.

CountQA, published in 2025 by Tamarapalli, Grover, Pande and Yerramilli, provides further evidence: the best-performing model in its original evaluation achieved 42.9% accuracy on challenging real-world object-counting questions.

These findings do not establish the exact mechanism behind ChatGPT’s response in the video.

They demonstrate that the observed failure is consistent with documented limitations in visual numerical grounding.

Importantly, later research has also shown that intermediate object representations and more specialised counting procedures can improve performance, although complex scenes remain challenging.

Part V: Humans Also Get Prompted — By Context

The comparison becomes more interesting when we leave the jelly-bean jar and examine ordinary life.

Imagine three identical statements:

“We need to talk.”

Your manager says it.

Your spouse says it.

A stranger says it.

Your underlying language-processing capacities remain available, but memories, relationships, expectations and emotions affect your interpretation.

System 1 may produce an immediate reaction: concern, fear or curiosity.

System 2 can reconsider the evidence:

“Do I actually know what this conversation concerns?”

The same principle applies to marketing.

Tversky and Kahneman’s 1981 research showed that equivalent decision problems can produce different responses when framed in terms of gains or losses.

Context influences human judgment, even when the underlying facts remain unchanged.

The distinction is that stable cognitive mechanisms and contextual influences operate together.

Humans are not entirely context-independent. Neither are they merely passive products of context.

They can use attention, evidence and executive control to question their initial interpretations.

An AI system similarly has stable learned computational machinery, while contextual instructions and inputs affect the response.

However, neither system is guaranteed to resist misleading framing.

Part VI: What Does Better Prompting Actually Achieve?

A better prompt does more than request a longer explanation.

It changes the task from merely providing an answer to demonstrating a method.

For the jelly beans, a useful prompt would be:

Estimate the number of jelly beans in this photograph. First identify what can actually be observed. Then determine what measurements are missing. Use the jar’s approximate volume, the volume of an average bean and packing density to construct a reasonable estimate. Explain your assumptions, provide a plausible range and state clearly whether an exact answer is possible.

This encourages System 2-like processing.

But there is a critical limitation.

A prompt can encourage a method; it cannot guarantee that the method is executed accurately.

Nor can it create information missing from the evidence.

For dependable numerical work, external tools, measurement procedures and independent verification may be necessary.

Conclusion: The Difference Between Guessing and Knowing

The jelly-bean experiment does not establish that humans are inherently better reasoners than AI.

A human making a quick guess might perform worse than ChatGPT.

An AI equipped with reliable image-analysis tools and a defensible calculation procedure might outperform most human estimators.

The meaningful comparison is therefore not simply human versus machine.

It is fast judgment versus justified judgment, regardless of who or what produces the answer.

System 1-like processing is efficient and often useful.

System 2-like processing allows more deliberate examination of assumptions.

But accuracy ultimately depends on the quality of the observations, the suitability of the method and the strength of verification.

Artificial intelligence increasingly demonstrates capabilities associated with both forms of reasoning.

Its fundamental challenge is deciding when a quick answer is adequate and when a problem demands additional evidence, computation or uncertainty.

The mark of reliable intelligence is not simply producing an answer quickly. It is knowing when a quick answer is sufficient, when deeper reasoning is required, and when neither can establish the truth.

That is the real lesson of 2,143 jelly beans.

Selected Academic References

1. Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
2. Evans, J. St. B. T. & Stanovich, K. E. (2013). Dual-Process Theories of Higher Cognition: Advancing the Debate. Perspectives on Psychological Science.
3. Miller, E. K. & Cohen, J. D. (2001). An Integrative Theory of Prefrontal Cortex Function. Annual Review of Neuroscience.
4. Fedorenko, E., Ivanova, A. A. & Regev, T. I. (2024). The Language Network as a Natural Kind Within the Broader Landscape of the Human Brain. Nature Reviews Neuroscience.
5. Cicchini, G. M., Anobile, G. & Burr, D. C. (2016). Spontaneous Perception of Numerosity in Humans. Nature Communications.
6. Vaswani, A. et al. (2017). Attention Is All You Need. NeurIPS.
7. Testolin, A., Hou, K. & Zorzi, M. (2025). Visual Enumeration Remains Challenging for Multimodal Generative AI. PLOS ONE.
8. Weng, T. et al. (2025). VisNumBench: Evaluating Number Sense of Multimodal Large Language Models. ICCV.
9. Pothiraj, A. et al. (2025). CAPTURe: Evaluating Spatial Reasoning in Vision Language Models via Occluded Object Counting. ICCV.
10. Tamarapalli, J. S. et al. (2025). CountQA: How Well Do MLLMs Count in the Wild?. arXiv preprint.
11. Brady, O. et al. (2025). Dual-Process Theory and Decision-Making in Large Language Models. Nature Reviews Psychology.
12. Zhang, D. et al. (2026). From System 1 to System 2: A Survey of Reasoning Large Language Models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 48(3), 3335–3354.
13. Fast, Slow, and Metacognitive Thinking in AI (2025). npj Artificial Intelligence.
14. Language Models, Like Humans, Show Content Effects on Reasoning Tasks (2024). PNAS Nexus.