This is a boring ass essay I wrote for the university of TUM a while ago. Wanted to share it with my big crowd of readers, including you… yes you before it will go completely forgotten in a folder of my laptop (even if boring at least acknowledge my music choice).
Can Computers Really Think?
Introduction
Transformers have completely revolutionized AI. For decades, it was an obscure research topic that only people in the field deeply understood, while now it is on everyone’s lips: even my grandma a few weeks ago, during a family lunch, asked me what “LLM” meant. This paradigm shift is so strong that, in Silicon Valley, the only companies receiving large investments often seem to be those that can attach the buzzword “AI” to their products, ranging from smart fridges to electric toothbrushes, even when there is no genuine need for it. There are people convinced that this is a temporary financial bubble about to burst, all while a growing number of solo developers quit their jobs to vibe-code apps for a living, CEOs confidently predicting that AGI will arrive in the next five years, and critics insisting that progress will soon hit a plateau because of fundamental limits in energy production.
In this wild landscape of opinions and overall instability, we can boil down people’s fundamental beliefs into two groups: those who believe that AI can think and those who do not. It is hard to give a strong, definitive answer: a recent survey of NLP researchers found an almost even split, with 51% believing that large language models can understand language and 49% disagreeing (Mitchell et al., 2023). However, when we ask ourselves “Can computers think?”, we are usually not only taking into consideration current models but also what advanced systems might be able to do in the future; and from this broader perspective, I strongly lean towards a positive answer. Before explaining why, though, we need to clarify what is meant by “thinking” (or intelligence), since much of the disagreement arises from a fundamental conflict over this very definition.
What Does It Mean to “Think”?
Thinking and intelligence, being human constructs, are hard to define within strict boundaries even if we follow science; for this reason I will propose three views, with increasingly demanding requirements, of what it could mean for a system to “think”.
At the most basic level, cognitive science often treats thinking as information processing: a system takes in inputs and manipulates them to choose actions or predict outcomes. Following this definition, even systems from over ten years ago were already “thinking”. DeepMind’s AlphaGo, for example, learned a statistical model of the game of Go by training on millions of positions through self-play. In 2016, in the second game of its match against former world champion Lee Sedol, it came up with its now-famous move 37, a move that no human had played before, described by many Go players as an undeniable example of creative play.
A second, stricter view insists that thinking requires not just the ability to analyse data, but the presence of an internal embedding of a world model, meaning that the system forms a set of concepts and relations that it can reuse in different situations. Using this view, we are getting into AGI (Artificial General Intelligence) territory: systems that are not just good at one narrow task, but can flexibly transfer what they have learned across domains through a coherent worldview, reaching an understanding at least comparable to, if not better than, that of humans. As of today, we are clearly not there yet, but we live in times where a system like that is not hard to imagine. A recent trend in the space is multimodal LLMs that can take not only text as input but also images, video, and audio, and can output these different formats by calling specialised tools such as web search or image generation, effectively merging different models under the hood to cover far more functionality than only a few years ago. In this sense, one can argue that they possess a rudimentary world model and can reuse their internal knowledge across very different tasks, routing requests to different specialised components based on their own decisions.
Finally, a third and most demanding view ties thinking to conscious, subjective experience: there is genuine thought only when there is a “someone” experiencing doubts, expectations, and mental imagery while they process information. This level of intelligence has not been reached by any AI yet, since it builds on top of the previous ones; a model at this level would typically be labelled ASI (Artificial Superintelligence). Whether ASI is even reachable is itself controversial: some researchers argue that continued scaling will eventually produce systems with richer self-models and long-term autonomy, while others are more sceptical, pointing to limits in data, energy, and current architectures that lack stable world models or persistent agency.
A Brief History of Machines That “Think”
Long before modern deep learning, Alan Turing was already pondering whether machines could think. In his 1950 paper he proposed what is now called the Turing Test as a way to test machine intelligence: if a machine can carry on a text conversation that is indistinguishable from that of a human counterparty, then we should treat it as thinking. This deliberately sidesteps inner mechanisms and focuses on behaviour. One of the first famous attempts to pass the test was ELIZA, Joseph Weizenbaum’s chatbot from the 1960s. Technically, ELIZA was simple: it scanned user input for keywords and applied hand-written transformation rules to reflect the sentence back in the style of a psychotherapist. Yet some users became emotionally attached to it and even asked to be left alone with the program. This already illustrates perceived intelligence, still very far from any genuine self-awareness.
A few decades later, we saw the rise of game-playing programs that fully cover the first definition of thinking. Deep Blue’s victory over Garry Kasparov in 1997 was the first widely publicised event showing a machine beating a world-class human in a specific task. In the 1990s, TD-Gammon went further by using a neural network trained via self-play, allowing the system to discover internal features of good and bad positions and develop a non-trivial intuition for the game. Over time, reinforcement learning methods created internal representations of finite worlds through reward signals, roughly analogous to biological learning mechanisms.
All of this leads to the current generation of large language models. Trained to predict the next token on internet-scale text, they now pass professional exams, write code, and follow complex instructions. Under certain conditions, these systems can display behaviours that resemble deception and self-preservation. From the outside, their behaviour is now close enough to that of intelligent agents—planning, scheming, sometimes lying, and resisting shutdown—that it becomes increasingly difficult to maintain that they are “just calculators”.
So, Can Computers Think?
Under the first, minimal view of thinking as non-trivial information processing and problem solving, computers clearly already think and have been doing so for decades. Under the second, stronger view based on internal world models and reusable concepts, today’s systems are not yet at the human level, but they are beating us in narrow, continuously expanding domains. The third view, which ties thinking to conscious experience, remains genuinely open. In practice, Turing’s stance remains compelling: when a system reliably talks, plans, and adapts like a thoughtful human, treating it as if it can think is the only reasonable strategy. In that sense, computers can already think; the more interesting question is what kind of thinkers we are building and how we choose to live with them.