AI (with current ML/DL architecture) does not believe nor reason nor know nor think!
posted: 09-Jun-2025 & updated: 26-Aug-2025
Contemporary LLMs are too powerful, too versatile, and too useful for most people to accept that they fundamentally lack the human capacities we so readily attribute to them. Yet understanding what they truly do—and don't do—is crucial for our technological future.
True knowledge, in any meaningful sense, requires the capacity to distinguish truth from falsehood. LLMs lack this fundamental capability because they don't inhabit the world we human language-users share. They cannot participate fully in what philosophers call “the language game of truth.”
The future of AI must be shaped not just by technological advancement, but by deep understanding of both its genuine capabilities and its essential limitations. We need the right philosophy, demonstrable commitment to social good, and rigorous implementation standards—but all grounded in clear-eyed assessment of what these systems actually are and do.
This clarity doesn't diminish the achievement—it grounds it in reality, where all sustainable progress must ultimately reside. Just as the Rhône River in Lyon carried both Roman vessels and the narratives of human civilization, our technological innovations carry forward human values and aspirations. But we must never forget that we, not our creations, remain the authors of that story.
The future belongs not to human-like AI, but to AI that amplifies uniquely human capabilities while remaining clearly and distinctly itself. This distinction isn't a limitation—it's the foundation for a technological future worthy of our highest aspirations.
NotebookLM Podcasts
- Decoding AI - Why Large Language Models Don’t Think, Believe, or Know Like Us (14:29)
- Beyond Belief - Deconstructing AI’s True Capabilities and Humanizing the Machine (13:00)
- Beyond Human - Deconstructing AI’s True Capabilities and Limits (10:47)
- Decoding AI - Why Large Language Models Don’t Think, Believe, or Know Like Us (10:34)
- Decoding AI - Why Your LLM Doesn’t Think, Believe, or Know (and Why It Matters) (12:12)
- Beyond Belief - Why Your AI Doesn’t Think Like You Do (11:10)
The Illusion of Understanding
Standing in Mozart’s birthplace in Salzburg, surrounded by the tangible artifacts of human creativity, I found myself reflecting on what distinguishes authentic human intelligence from the remarkable but fundamentally different capabilities of today’s AI systems. The irony wasn’t lost on me—here I was, fresh from presenting to the audience about AI as a “creative partner,” while simultaneously grappling with the philosophical implications of what these systems actually do versus what we imagine they do.
This tension between perception and reality in AI has become one of the most critical challenges of our time. As someone who has lived through the evolution of AI from academic curiosity to the driving force behind trillion-dollar valuations in Silicon Valley, I’ve witnessed firsthand how easily we slip into anthropomorphizing these systems—and how dangerous that tendency can be.
What LLMs Really Do – The Uncomfortable Truth
Let me be direct about something that might surprise you: when you ask ChatGPT “Who is Tom Cruise’s mother?” and it responds “Mary Lee Pfeiffer,” it doesn’t actually know this fact in any meaningful sense. What it’s really doing is something far more mechanical yet paradoxically more remarkable.
The Conditional Probability Game
Large Language Models (LLMs) are, at their core, sophisticated conditional probability estimators. When you prompt an LLM1 with “The first person to walk on the Moon was,” you’re not really asking who was the first person to walk on the Moon. You’re asking:
Given the statistical distribution of words in the vast public corpus of text, what words are most likely to follow this sequence?
This distinction might seem pedantic, but not at all! – it’s fundamental. Consider this fascinating example I presented to the audience: when an LLM (as of January 2022) was asked “Who is Tom Cruise’s mother?” it confidently replied with detailed information about Mary Lee Pfeiffer. But when the very same system was asked “Who is Mary Lee Pfeiffer’s son?” it responded, “I don’t have specific information about Mary Lee Pfeiffer or her family.”
If the system truly “knew” these facts, such asymmetry would be impossible. What we’re seeing instead is the artifact of pattern matching and statistical likelihood, not genuine knowledge or understanding.
The Anthropomorphism Trap
The tendency to anthropomorphize AI systems isn’t just a quirk of public perception—it’s an occupational hazard even for those of us deep in the field. When an LLM can improve its performance on reasoning tasks simply by being told to “think step by step,” the temptation to see it as having human-like characteristics becomes almost overwhelming.
This is where my experience across different domains—from Samsung’s semiconductor labs to Amazon’s e-commerce algorithms, from Gauss Labs in Palo Alto to the policy discussions in Salzburg—provides crucial perspective. The more technically sophisticated these systems become, the more seductive the illusion of consciousness becomes.
Knowledge, Belief, and the Prerequisites of Understanding
What Knowledge Really Requires
From both laymen’s and philosophical perspectives, the question of whether LLMs possess knowledge reveals profound category errors in our thinking. Could we argue that an LLM “knows” which words typically follow other words? Perhaps, in a very limited sense. But knowing that “Burundi” is likely to succeed “The country to the south of Rwanda is” is categorically different from knowing that Burundi is to the south of Rwanda.
If you doubt this distinction, consider whether knowing that “little” likely follows “Twinkle, twinkle” is the same as knowing that “twinkle twinkle little.” The idea doesn’t even make coherent sense.
True knowledge, in any meaningful sense, requires the capacity to distinguish truth from falsehood. LLMs lack this fundamental capability because they don’t inhabit the world we human language-users share. They cannot participate fully in what philosophers call the language game of truth.
The Belief Prerequisite Problem
For something to count as a belief about the world we share, it must exist against the backdrop of the ability to update beliefs appropriately in light of evidence from that world. This is an essential aspect of the capacity to distinguish truth from falsehood.
LLMs fundamentally lack this ground. Even when embedded in systems that consult external information sources and update their parameters, the underlying model remains a conditional probability estimator operating on text patterns, not a belief-forming entity engaging with reality.
Reasoning
The question of whether LLMs can reason is particularly nuanced because reasoning, grounded in formal logic, is content-neutral. The modus ponens rule holds regardless of whether we’re discussing squirgles and splonky creatures or real-world entities.
When we prompt an LLM with “All humans are mortal and Socrates is human, therefore,” we’re not instructing it to perform deductive inference. We’re asking:
Given the statistical distribution of words in the public corpus, what words are likely to follow this sequence?
The answer
Socrates is mortal
emerges not from logical reasoning but from pattern completion.
Chain-of-thought prompting, while remarkably effective, exemplifies this pattern completion rather than genuine reasoning. The model generates sequences that mimic well-formed arguments because such sequences are statistically likely given the prompt structure, not because it’s engaging in logical inference.
The Silicon Valley Reality Check
Having lived at the epicenter of the AI revolution—from the venture capital conversations on Sand Hill Road to the technical depths of the Transformer architectures—I’ve observed a fascinating disconnect. The very brilliance of these systems creates a compelling illusion that obscures their fundamental limitations.
Why This Matters Beyond Philosophy
This isn’t merely academic philosophizing. The way we characterize AI capabilities has profound implications:
For Policy and Regulation
When lawmakers hear that AI systems “believe” or “know” things, they make fundamentally different regulatory decisions than they would if they understood these systems as sophisticated pattern matching engines.
For Business Strategy
Companies that anthropomorphize their AI systems make different strategic choices about trust, verification, and human oversight than those that maintain clarity about what these systems actually do.
For Public Trust
The gap between perception and reality creates both unwarranted fear and misplaced confidence, neither of which serves society well.
The Emergence Question
Contemporary LLMs exhibit extraordinary emergent capabilities when trained at scale. Could these emergent properties include something analogous to knowledge or belief? This is where the philosophical nuance becomes crucial.
While we might say an LLM “encodes” or “contains” knowledge (as an encyclopedia does), this is fundamentally different from possessing knowledge in the way humans do. The LLM has no access to external reality against which its predictions might be measured, no means to apply external criteria of truth beyond statistical correlation in its training data.
The Path Forward – Precision Without Diminishment
My argument isn’t that LLMs are somehow lesser or that their capabilities should be dismissed. Quite the contrary—their actual capabilities are extraordinary enough without the need for anthropomorphic embellishment.
What We Gain from Precision
Better System Design
Understanding what LLMs actually do enables better architectural choices, more appropriate human-AI collaboration patterns, and more effective safety mechanisms.
Informed Decision-Making
Clear understanding of capabilities and limitations enables better decisions about where and how to deploy these systems.
Sustainable Progress
Avoiding the boom-bust cycles that often accompany overhyped technologies requires realistic assessment of current capabilities alongside vision for future development.
The Technology-Humanity Bridge
This brings me back to the central theme that emerged during my European journey—the imperative to bridge technology and humanity thoughtfully. Just as I argued in Salzburg that we must harness AI to solve the very problems it creates, we must also resist the tendency to project human qualities onto systems that, however sophisticated, operate on fundamentally different principles.
The future of AI must be shaped not just by technological advancement, but by deep understanding of both its genuine capabilities and its essential limitations. We need the right philosophy, demonstrable commitment to social good, and rigorous implementation standards—but all grounded in clear-eyed assessment of what these systems actually are and do.
Consciousness – The Ultimate Category Error
The question of AI consciousness exemplifies the category errors that plague our discourse. We have no agreed definition of consciousness—and likely never will. Yet we proceed to debate whether AI systems possess something we cannot define.
Rather than getting trapped in these circular discussions, we should focus on what we can observe and measure – the remarkable pattern-matching capabilities, the emergent behaviors at scale, the utility and limitations of these systems in specific contexts.
Does this diminish the wonder of what we’ve achieved? No, I don’t think so! Standing in Mozart’s birthplace, I was struck not by the similarity between human creativity and AI generation, but by their fundamental differences. There is an authenticity to human creativity—rooted in lived experience, embodied existence, and genuine understanding—that cannot be simulated or replicated. This difference exists (almost) by definition of human emotions.2
AI cannot touch our souls, and this truth, while not empirically provable, resonates more deeply and compellingly than any proof could. Because proof can prove only so much. At the end of the day, what does proving something mean anyway? Does it not mean only that we have deduced something based on a handful of axioms that we assume are correct? Where the correctness of the axioms can never be proved because… they are axioms.
The Responsibility of Clarity
As AI practitioners—whether researchers, engineers, entrepreneurs, or policymakers—we bear the responsibility of speaking clearly about these systems. The careless use of philosophically loaded terms like “believes,” “thinks,” or “knows” isn’t harmless shorthand; it actively shapes public understanding, policy decisions, and the trajectory of technological development.
Contemporary LLMs are so powerful and convincing that such imprecision can no longer be safely applied. The stakes are too high, the implications too far-reaching, and the potential for both benefit and harm too great.
We stand at a remarkable moment in technological history. The systems we’re building have extraordinary utility and enormous potential. To ensure we can make informed decisions about their trustworthiness, safety, and appropriate deployment, we must maintain clarity about what they actually do while avoiding imputing capacities they lack.
This clarity doesn’t diminish the achievement—it grounds it in reality, where all sustainable progress must ultimately reside. Just as the Rhône River in Lyon carried both Roman vessels and the narratives of human civilization, our technological innovations carry forward human values and aspirations. But we must never forget that we, not our creations, remain the authors of that story.
The future belongs not to human-like AI, but to AI that amplifies uniquely human capabilities while remaining clearly and distinctly itself. This distinction isn’t a limitation—it’s the foundation for a technological future worthy of our highest aspirations.