Last month, I wrote a piece here in Big Think suggesting that the public is in denial over the emerging power and associated risks of AI. Many people reacted defensively, insisting that today’s AI systems are nothing more than “stochastic parrots” that regurgitate memorized information and are structurally incapable of emergent reasoning. I appreciate that many people want this to be true, but it is an outdated narrative based on a 2021 paper that was widely misinterpreted. I want to clear up this misconception because I worry it is giving the public a false sense of security that superintelligence is not an urgent risk to society.
To address this, let me update our collective mental model: Like it or not, there is increasing evidence that frontier AI systems do not just store text patterns as statistical correlations, but also build structured internal representations (i.e., “world models”) that abstract the concepts described in the ingested content. This is not surprising since human brains do the same thing — we consume information as language but store it as abstract conceptual models. Our brains are far more efficient at building mental models than current AI training methods, but recent research suggests that large frontier systems do build conceptual models and are rapidly improving.
One early study on this front showed that large language models (LLMs) trained only on language describing the rules of a common board game (Othello) contained an internal map of the board state (i.e., a world model). Researchers were able to “edit” the internal model and change the state of pieces on the board. This suggests a conceptual encoding of information that goes far beyond stochastic language correlations. Another notable paper showed that training LLMs on textual content about places and events can produce an internal “world model” of the location and time of the elements represented in the language. In fact, the researchers identified “space neurons” and “time neurons” within the model that reliably encode spatial and temporal coordinates.
Of course, the best evidence that frontier models are not merely “parrots” is the many studies that suggest AI systems can solve problems that do not exist in their training datasets, thereby demonstrating out-of-distribution reasoning. Here is a simple example of this type of reasoning that I just tested myself on Gemini 3:
PROMPT: I have a Ram truck with a flat. I need to change the tire, but the handle broke off my lug wrench. I have a metal water bottle, a stretchy belt, a pocketknife, two rolls of duct tape, a cinder block, a hardcover book with glossy pages, a hollow metal pipe, and a lighter. The metal pipe is rusted, breaking under force. My knife is blunt and old. The water bottle is half full. How can I replace the flat tire?
GEMINI 3 ANSWER: To replace the flat tire, reinforce the rusted pipe by inserting the metal water bottle (or rolled book pages) for internal support, wrap it heavily in duct tape to prevent buckling, and use this composite sleeve as a high-leverage extension for your broken lug wrench.
For me, this solution defies the stochastic narrative by generating a seemingly creative and out-of-distribution solution. While I can’t guarantee that similar problems did not exist in the training data set, I based the example above on a recently published study that tested a set of 1,600 novel problems of this nature, aiming to compare “out-of-the-box” reasoning performed by humans and AI systems. The study found that GPT-4 could rival human abilities but did suffer from generating a larger number of infeasible solutions. This error mode was significantly reduced by adding a “reflection” step to the AI problem-solving process.
In addition, today’s large models can solve problems presented in formats other than text descriptions. Here’s an example of a non-verbal logic problem that was recently given to frontier AI systems to assess reasoning abilities against human test-takers:
Today’s top models can solve the problem above by processing the image itself (not a text description of the image). The only way this could be dismissed as rote memorization is if similar examples were in the training set. To address this, TrackingAI.org has been testing out-of-distribution IQ test questions, which don’t exist in training data, on frontier models. Their latest testing showed Gemini 3 Pro scoring an IQ of 130, thereby outperforming 97% of human test-takers.
So, how can frontier models, trained by finding statistical correlations in language, form mental models and perform out-of-distribution reasoning? This is not well understood, but part of the answer may be the concept of compression as intelligence. It suggests that when training systems to optimize efficiency, the internal encoding must eventually model the underlying rules (physics, human behavior, logic, social norms, etc.) inherent in the data. After all, conceptual rules produce the most efficient encodings. Recent research supports this premise, finding that LLMs’ intelligence, as reflected in average benchmark scores, correlates with the model’s ability to compress ingested content.
Another part of the solution is that current frontier models use Chain-of-Thought (CoT) or Tree-of-Thought (Tot) structures, which tackle problems in a sequence of iterative steps. This is intended to mimic human deliberation by repeatedly checking for errors and assessing progress towards a solution. This turns the LLM from a large deterministic equation into a complex control system with parallelized feedback loops that break down problems, brainstorm solutions, narrow viable options, and converge on answers. In other words, AI systems are increasingly deliberative, not statistical.
So, why has the word “stochastic” become synonymous with unintelligent?
Researchers who see the stochastic structure of LLMs as a barrier to intelligence are likely looking for intelligence in the wrong place. High-level intelligence does not exist in the statistical encoding of information within frontier models. It also does not exist in the artifacts produced by frontier models. Instead, intelligence exists in the behavior of the control system that is wrapped around the encoded information. To appreciate this distinction, I recommend we look at intelligence in the context of control theory.
I contend that the evolutionary foundation of intelligence is feedback control. This goes back to the first single-celled organisms that could find food by sensing nutrients and triggering flagella to guide them in the right direction. The ability to distinguish specific nutrients from other molecules involved stochastic encoding of chemical information. The intelligence that emerged was not the encoding, but the control system around that encoding. This is what turned a dumb cell that floated mindlessly in a pond into a creature that could actively seek out food. In other words, stochastic encoding of information is not a rebuttal to intelligence — it is the foundation of intelligence when part of a control system.
In fact, I contend that human intelligence can be described as a control system that iteratively processes stochastically encoded information to solve out-of-distribution problems. In our brains, neurons encode information probabilistically. This stochastic structure is not a flaw, but a feature, for it enables us to represent uncertainty in our environment and remain flexible in the face of limited context. Our brains use feedback loops to manipulate this information and find solutions. When we iteratively manipulate information to solve problems, the feedback loop is called “deliberation.” When we do it subconsciously, it’s called “intuition.” And when that feedback loop wraps around our own deliberative processes, assessing our own inner workings, it might even help to produce our “sense of self.” This last point is the most speculative but correlates with emerging theories of consciousness, such as Attention Schema Theory (AST).
So, why is the public so resistant to the idea that AI systems can reason?
Again, I believe it’s rooted in denial. For many people, it’s more comfortable to insist that AI models merely regurgitate content than confront a scary future filled with rival minds that can reason thoughtfully. I say that as someone who believes society should not be racing to build superintelligent systems. I’ve been sounding this alarm for a decade (see my 2017 TED Talk) and continue to believe we should slow our pursuit. That said, I am also realistic enough to know that we will not slow development. Instead, the best we can do is have a realistic view of AI and thoughtfully prepare for the risks posed by superintelligent systems.
For me, this means urging industry, government, and academia to keep humans in the loop, resisting the impulse to replace people with AI-powered decision-making. I firmly believe that AI systems will become smarter and faster than individual humans by a wide margin, but that does not guarantee that the systems will represent human interests, morals, or sensibilities when solving problems. This is because the “world models” at the core of these AI systems will be remarkably good at knowing humans, but that is not the same as being human. This difference is profound, underestimated, and currently unsolved.
This article Do AI models reason or regurgitate? is featured on Big Think.