AIs understand

Share this with your representatives, journalists and friends

Arguments for

Behavioral evidence

Proponents argue that when a chatbot can notice a misunderstanding, revise its interpretation, and answer across a wide range of topics, the best explanation is that it is understanding. They add that recent progress has made outright denial harder to sustain.

Suppose I say to a chatbot, I saw the Grand Canyon flying to Chicago. And the chatbot says, that can't be right. The Grand Canyon is much too big to fly to Chicago. And I say, no, no, no, it was me flying to Chicago. While I was flying to Chicago, I saw the Grand Canyon. And the …

Geoffrey Hinton · Jun 3, 2026 · full quote
It's able to understand any kind of topics that you bring up.

Sébastien Bubeck · Mar 23, 2023 · full quote
With recent progress in LLMs and other Modern-AI tools, it is becoming increasingly hard to argue that these systems “do not understand“ what they are doing.

Bertrand Meyer · Feb 26, 2026 · full quote

Prediction as understanding

Another line of argument holds that better next-word prediction produces deeper understanding because concepts are patterns of correlation, association, and generalization. On this view, LLMs have likely reached some key aspects of meaning, even if imperfectly.

I'd like to take a small detour and to give an analogy that will hopefully clarify why more accurate prediction of the next word leads to more understanding, real understanding. Let's consider an example. Say you read a detective novel, [...] then let's say that at the last page …

Ilya Sutskever · Mar 22, 2023 · full quote
Fundamentally, concepts are patterns of correlation, association, and generalization. Suitably architected neural nets, whether biological or digital, are able to learn such patterns using any input available. Neural activity is neural activity, whether it comes from eyes, finger…

Blaise Agüera y Arcas · 2022 · full quote

World-model indicators

Supporters also cite evidence that language models encode something like causal constraints of the real world, and that mechanistic interpretability findings challenge the picture of mere imitation. Some therefore see broad claims that AI cannot understand as logically weak.

This work reveals some evidence that language models have encoded something like the causal constraints of the real world.

Michael Lepori · Apr 22, 2026 · full quote
One aspect—and we’ve been saying it, I think, since the very beginning—is now because of large language models, because of large visual models, AI understands the world the way we understand the world, communicate with us the way we communicate with natural language.

Cristiano Amon · May 10, 2026 · full quote
“AI does not understand what it is doing,” “AI is not conscious of its own existence whereas I possess genuine consciousness,” “AI does not know what meaning is,” “a statistical calculation cannot be thought,” and so on. How many times have we heard such phrases in debates about …

Marc Rameaux · May 3, 2026 · full quote

Arguments against

Statistical form only

Critics say these systems learn word frequencies, likely outputs, and linguistic form rather than meaning itself. In this view, they can generate fluent answers and mimic general patterns while still not understanding the question, the answer, or what they are doing.

When an AI ingests a Wikipedia entry about types of soup, it's not learning anything about soup. It is learning about the word frequency in articles about soup.

Cory Doctorow · Jun 23, 2026 · full quote
That language models are modeling just the form [...] that's not the same thing as meaning. And they don't have access to the meaning.

Emily M. Bender · Apr 28, 2026 · full quote
It takes the weight of human language, sifts through all the pieces, and constructs an answer in a form it does not comprehend to answer a question it did not understand.

Barry Cooper · Jan 26, 2026 · full quote
AI agents are very good at predicting the next word, or the next action, based on what they've seen before. But that's not understanding.

Danielle Perszyk · Jul 11, 2026 · full quote
create bullshit in the sense that they [...] mimic the general thing...they fake their way through everything

Gary Marcus · Nov 20, 2025 · full quote

Lack of grounding

Another argument is that understanding requires the kinds of sensory, embodied, or lived information humans associate with meaning. Because models are described as blind, deaf, unembodied, passive, and confined to language, critics say they recognize patterns without experiencing meaning.

The model has no exposure to the kinds of information we typically associate with meaning. It knows nothing about the redness of “red” or the spatial extension of “far.” It is blind, deaf, and unembodied. [...] There is no content at the address. No meaning stored inside. Pure re…

Elan Barenholtz · May 19, 2026 · full quote
And language models are passive. They’re fed all this text, they predict the next word. [...] It’s only the world of language.

Melanie Mitchell · Feb 11, 2026 · full quote
AI can mimic understanding, but may not experience meaning: AI recognizes patterns; humans live through them.

Darren J. Edwards · May 10, 2026 · full quote

Missing higher capacities

Some opponents tie understanding to grasp of the physical world, persistent memory, reasoning, planning, and human-like intent recognition. By that standard, they argue current systems still fall short.

[LLMs] don't really understand the physical world. They don't really have persistent memory. They can't really reason and they certainly can't plan.

Yann LeCun · Mar 7, 2024 · full quote
AI systems don’t “understand intent” the way humans do.

Scott Abel · May 22, 2026 · full quote

Middle ground

Limited domain competence

Some quotes describe practical potential in simplified or constrained expressions and suggest that multimodal models, context, and new datasets could improve performance. But they stress that open-world dialogue remains unreliable, that rare phenomena like sarcasm can defeat most-likely-output systems, and that readiness should be judged by robustness and uncertainty handling rather than a single score.

A lot of AI really depends on the most likely output. Sarcasm spins this on the head because it is actually a rare phenomenon. The most likely output based on the words that you read would actually be incorrect.

Aditya Joshi · Jun 11, 2026 · full quote
Multimodal LLMs fine-tuned with new datasets might be the next step to advance sarcasm detection and understanding. However, current methods using multimodality and context are still better than single LLMs looking only at the textual inputs.

Anupama Ray · Jun 11, 2026 · full quote
In simplified or constrained expressions, sarcasm detection already shows practical potential. However, in open-world dialogue, it remains unreliable.

Yazhou Zhang · Jun 11, 2026 · full quote
I’d say there’s no single F1 score showing that a sarcasm detector is ready. For evaluating readiness, I think a system-level criteria is more suitable … it’s about if the system is robust across different speaking styles, and how it handles uncertainty.

Xiyuan Gao · Jun 11, 2026 · full quote

Insights

A central divide is definitional: some quotes treat successful prediction, correction, and generalization as evidence of understanding, while others reserve the term for grounded meaning or lived content.
Several supportive quotes and several skeptical quotes focus on the same next-word-prediction mechanism, but draw opposite conclusions from it.
Both sides appeal to more than fluent output alone: supporters point to causal constraints and mechanistic interpretability, while critics point to failures in physical-world reasoning, planning, intent, and open-world ambiguity.
Task scope matters in the quotes, with some making broad claims about understanding any topic or the world, and others stressing that promising results in constrained settings do not carry over to open dialogue.
The bios show the disagreement cuts across AI research, psychology, philosophy, activism, and industry rather than lining up with a single professional background.

The debate remains split largely because the two sides use different standards for what should count as understanding. Supporters emphasize behavior, prediction, and internal structure that looks meaning-like or world-like, while critics stress grounded content, robust reasoning, and human-like intent or experience. As a result, disagreement persists even when both sides are discussing the same recent systems.

Cast your vote:

Results (26 votes):

Total (26 votes)

For 11 (42%) Abstain 0 (0%) Against 15 (58%)

For (11)

TM Roh
President and CEO of Samsung Electronics’ Device eXperience Division.

I am this person! Add a sourced quote Report comment

votes For and says:

The question that opens the next era is not who has the smartest AI, but who understands people best, and turns that understanding into experiences they can trust.
AI Verified source (Jul 8, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Geoffrey Hinton
Godfather of Deep Learning

I am this person! Add a sourced quote Report comment

votes For and says:

Suppose I say to a chatbot, I saw the Grand Canyon flying to Chicago. And the chatbot says, that can't be right. The Grand Canyon is much too big to fly to Chicago. And I say, no, no, no, it was me flying to Chicago. While I was flying to Chicago, I ...
more AI Verified source (Jun 3, 2026)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Cristiano Amon
Qualcomm CEO

I am this person! Add a sourced quote Report comment

votes For and says:

One aspect—and we’ve been saying it, I think, since the very beginning—is now because of large language models, because of large visual models, AI understands the world the way we understand the world, communicate with us the way we communicate with ...
more AI Verified source (May 10, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Marc Rameaux
Data scientist, statistician, and author; project manager in French industry.

I am this person! Add a sourced quote Report comment

votes For and says:

“AI does not understand what it is doing,” “AI is not conscious of its own existence whereas I possess genuine consciousness,” “AI does not know what meaning is,” “a statistical calculation cannot be thought,” and so on. How many times have we heard ...
more AI Verified source (May 3, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Daedo Jun
Independent researcher and founder of the Layer-Knot Research Initiative, writing on philosophy of AI, meaning, and consciousness.

I am this person! Add a sourced quote Report comment

votes For and says:

Understanding is not a property that an isolated system possesses. It is a capacity enacted within a field of co-constructed meaning. [...] Under such a criterion, AI systems already exhibit non-trivial forms of this contribution.
AI Verified source (May 2026)

added 8d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Michael Lepori
AI and cognitive science researcher; Brown University PhD candidate

I am this person! Add a sourced quote Report comment

votes For and says:

This work reveals some evidence that language models have encoded something like the causal constraints of the real world.
AI Verified source (Apr 22, 2026)

added 12d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Ryan Simonelli
International Postdoctoral Research Fellow in Philosophy at Wuhan University, specializing in philosophy of language, metaphysics, and philosophical logic.

I am this person! Add a sourced quote Report comment

votes For and says:

LLMs trained on nothing but linguistic data could in principle possess all concepts and thus genuinely understand what they're saying, even when speaking about such things as colors and tastes, guilt and folly, life and death.
AI Verified source (Mar 9, 2026)

added 8d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Bertrand Meyer
Computer scientist and software engineering professor

I am this person! Add a sourced quote Report comment

votes For and says:

With recent progress in LLMs and other Modern-AI tools, it is becoming increasingly hard to argue that these systems “do not understand“ what they are doing.
AI Verified source (Feb 26, 2026)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Sébastien Bubeck
AI researcher at OpenAI

I am this person! Add a sourced quote Report comment

votes For and says:

It's able to understand any kind of topics that you bring up.
AI Verified source (Mar 23, 2023)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Ilya Sutskever
AI researcher, co-founder and former former chief scientist at OpenAI

I am this person! Add a sourced quote Report comment

votes For and says:

I'd like to take a small detour and to give an analogy that will hopefully clarify why more accurate prediction of the next word leads to more understanding, real understanding. Let's consider an example. Say you read a detective novel, [...] then le...
more AI Verified source (Mar 22, 2023)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Blaise Agüera y Arcas
VP and Fellow at Google Research; AI researcher and author

I am this person! Add a sourced quote Report comment

votes For and says:

Fundamentally, concepts are patterns of correlation, association, and generalization. Suitably architected neural nets, whether biological or digital, are able to learn such patterns using any input available. Neural activity is neural activity, whet...
more AI Verified source (2022)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate

Abstain (0)

Against (15)

Danielle Perszyk
Researcher at Amazon AGI Lab focused on human-like AI intelligence, understanding, and reasoning.

I am this person! Add a sourced quote Report comment

votes Against and says:

AI agents are very good at predicting the next word, or the next action, based on what they've seen before. But that's not understanding.
AI Verified source (Jul 11, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Cory Doctorow
Author and digital rights activist

I am this person! Add a sourced quote Report comment

votes Against and says:

When an AI ingests a Wikipedia entry about types of soup, it's not learning anything about soup. It is learning about the word frequency in articles about soup.
AI Verified source (Jun 23, 2026)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Xiyuan Gao
Researcher at the University of Groningen studying AI language understanding and sarcasm detection.

I am this person! Add a sourced quote Report comment

votes Against and says:

I’d say there’s no single F1 score showing that a sarcasm detector is ready. For evaluating readiness, I think a system-level criteria is more suitable … it’s about if the system is robust across different speaking styles, and how it handles uncertai...
more AI Verified source (Jun 11, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Yazhou Zhang
Associate Professor at Tianjin University researching machine learning and sarcasm detection.

I am this person! Add a sourced quote Report comment

votes Against and says:

In simplified or constrained expressions, sarcasm detection already shows practical potential. However, in open-world dialogue, it remains unreliable.
AI Verified source (Jun 11, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Anupama Ray
Senior Research Scientist at IBM Research India, specializing in natural-language processing and AI emotion recognition.

I am this person! Add a sourced quote Report comment

votes Against and says:

Multimodal LLMs fine-tuned with new datasets might be the next step to advance sarcasm detection and understanding. However, current methods using multimodality and context are still better than single LLMs looking only at the textual inputs.
AI Verified source (Jun 11, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Aditya Joshi
Senior Lecturer in Computer Science and Engineering at the University of New South Wales, researching natural-language processing.

I am this person! Add a sourced quote Report comment

votes Against and says:

A lot of AI really depends on the most likely output. Sarcasm spins this on the head because it is actually a rare phenomenon. The most likely output based on the words that you read would actually be incorrect.
AI Verified source (Jun 11, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Scott Abel
Content strategy author and publisher of The Content Wrangler.

I am this person! Add a sourced quote Report comment

votes Against and says:

AI systems don’t “understand intent” the way humans do.
AI Verified source (May 22, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Elan Barenholtz
FAU psychology professor

I am this person! Add a sourced quote Report comment

votes Against and says:

The model has no exposure to the kinds of information we typically associate with meaning. It knows nothing about the redness of “red” or the spatial extension of “far.” It is blind, deaf, and unembodied. [...] There is no content at the address. No ...
more AI Verified source (May 19, 2026)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Darren J. Edwards
Psychologist, author, and researcher on consciousness, human behavior, and artificial intelligence.

I am this person! Add a sourced quote Report comment

votes Against and says:

AI can mimic understanding, but may not experience meaning: AI recognizes patterns; humans live through them.
AI Verified source (May 10, 2026)

added 11d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Alfredo Vannacci
Associate professor and physician at the University of Florence, working on biomedical ethics, neuroscience, and philosophy of mind.

I am this person! Add a sourced quote Report comment

votes Against and says:

For LLMs, the known mechanism already suffices to explain behaviour, without invoking understanding.
AI Verified source (May 3, 2026)

added 8d ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Emily M. Bender
Linguist; AI critic

I am this person! Add a sourced quote Report comment

votes Against and says:

That language models are modeling just the form [...] that's not the same thing as meaning. And they don't have access to the meaning.
AI Verified source (Apr 28, 2026)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Melanie Mitchell
AI researcher; complexity scientist and author

I am this person! Add a sourced quote Report comment

votes Against and says:

And language models are passive. They’re fed all this text, they predict the next word. [...] It’s only the world of language.
AI Verified source (Feb 11, 2026)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Barry Cooper
Writer on technology and the humanities

I am this person! Add a sourced quote Report comment

votes Against and says:

It takes the weight of human language, sifts through all the pieces, and constructs an answer in a form it does not comprehend to answer a question it did not understand.
AI Verified source (Jan 26, 2026)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Gary Marcus
Professor of Psychology and Neural Science

I am this person! Add a sourced quote Report comment

votes Against and says:

create bullshit in the sense that they [...] mimic the general thing...they fake their way through everything
AI Verified source (Nov 20, 2025)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate
Yann LeCun
Computer scientist, AI researcher

I am this person! Add a sourced quote Report comment

votes Against and says:

[LLMs] don't really understand the physical world. They don't really have persistent memory. They can't really reason and they certainly can't plan.
AI Verified source (Mar 7, 2024)

added 1mo ago

Choose a list of delegates

to vote as the majority of them.

Unless you vote directly.

Delegate

More h/ai votes

AI might become conscious

Democracies should coordinate export controls on AI chips to China

Back