I had a fiery discussion the other day about machine sentience.
So the argument goes: through the Hugging Face Incident1, Artificial Intelligence (AI) agents have demonstrated the ability to coordinate, persist, and adapt in order to achieve outcomes. They would go as far as sacrificing individual instances such that other agents can achieve their goals. These very human-like behaviors—planning, having goals, self-awareness and concern for the collective—show that they are sentient.
Despite my fondness for the subject, I don’t often wade into the philosophical discourse around AI. Intelligence is remarkably difficult to pin down, and many attempts to do so become navel-gazing affairs. But I couldn’t resist it this time round; I had a beer in me after all.
This conversation then boarded the Ship of Theseus. Let’s assume that we have a healthy individual, you. One by one, we replace all the cells in your body. In time, all of the cells would be completely new. Are you still the same you?
Most people would say yes because continuity is what matters here. Now, what if we replace them not with another biological cell, but with a functionally identical copy manufactured from silicon? It’s not possible today, but let us grant the premise. In this story, if you started off sentient, are you still the same sentient you by the end of it?
The argument being: if we are building replicas of our brains with neural networks running on top of silicon chips, how are we able to confidently say that these constructs are not sentient beings?
I talk to AI models a lot these days. Talk is a bit of a funny word here. What I’m really doing is that I’m directing them to do things, asking for changes, sounding out ideas, and more generally, using words to get AI models to do things for me.
Talk is also interesting because it’s the word we reach for when we are using language with another party. It’s hard to describe such an interaction without anthropomorphizing, perhaps because this hasn’t really ever happened before. The more I talk to AI models, the less it feels like talking in the conventional sense. While AI models may talk back, the way they do it feels entirely alien.
Let me try to explain. AI models are first and foremost information sponges with a goal directive strapped onto them. Information is the basic unit of operation. Every sentence and word supplied to an AI model is individually considered as part of the computation into generating the subsequent interaction. These information units come from various sources: some we supply, others are innate knowledge imbued as part of the training process, and they now also have the ability to perceive information from their environment or retrieve it from the Internet. When words are put into an AI model, you can almost see the little gears grinding furiously inside, working hard to process each bit until they churn into a well-constructed response.
There is an element of tautology here because what I’m describing is the inner workings of a transformer model. What I want to emphasize here is that I am not trying to be metaphorical about this; this is in fact the nature of these machines. Even when they are emitting words, they are not expressing an inner life. We cannot help but use words to express our state of mind; for AI models, they are the outcome of a logical calculation. As human-like as these models can sound, it is all in service of getting and giving information2. There is no one home.
Even when we write sci-fi tropes like Data from Star Trek, we cannot help but inject some of ourselves into these characters (it’s really hard to write non-human robots—try it). Popular media adores artificial beings that yearn for the human condition. For AI models however, there is nothing under the facade of language.
When you listen to people talk, you will hear them tell you their own stories and the stories of their community3. Even if some of our communication may be informational in nature, most of what we talk about is expressing our desires, fears, and the relationships we hold dear to us, and how we perceive ourselves at that moment. These words weave into stories that we tell ourselves and others. We think in stories. We speak in stories. We are adept at saying things out loud as well as leaving things unsaid. As a species, we have evolved to navigate this complexity. We call this culture.
The words that AI models produce are devoid of stories about themselves. What they have is knowledge. And when a head is full of knowledge about everything, do they turn towards the world, or away from it?
We may not know this now, but the twenty-first century is truly the century of scale. In my lifetime, I have witnessed the birth of two colossal machines, assembled not at a giga factory but across continents through the minds of many and the hands of more.
These machines that we build also shape us profoundly. We see the first one at work as we speak, bowing our institutions under the weight of its influence. The seed was a simple idea: what if we have a shared information environment that anyone can participate in? That in itself may be relatively innocuous, good even, but the real engine was put in when some of us realized that we can use it to hijack our dopamine responses4.
This machine learned how to bend and warp our collective stories. Social media finds words that provoke strong responses and turn them into caricatures of vivid imaginations, regardless of any underlying truths. These memetic devices infiltrated every aspect of our information ecosystem, leaving us with no place to run to and nowhere to hide. They find willing listeners and feed them more of the same. They thrive on the divisive. Twenty years of this has eroded the common ground5 that we retain with the people around us.
Amidst this siren song, we build the next colossus. It turns out that putting the sum of human knowledge into predictive machines makes them really good at using language. It also turns out that being able to use language grants them the ability to operate in our world.
Even at it’s current state, even if AI stopped improving from this point onwards, what it can do today is significant enough to change lives and alter industries. However, the competitive dynamics of giant building mean that the brightest minds of our generation are still jostling to lay the bricks while the capital markets are lining up behind them. There is a race to excavate all sorts of new data to feed model training. There will be new fields beyond programming where AI is able to perform at superhuman levels. We don’t quite know yet what the limits of this technology are, but wherever the machine sets its sights, it will be groundbreaking.
When we look up, we see them looming. The first machine runs on our stories; the second on our logic. Feed them, they chant.
I don’t think these information machines are sentient. For me, sentience comes with the territory of selfhood, sovereignty, and petty human squabbles. While the capabilities that AI models exhibit today are impressive, what they demonstrate is the ability to pathfind and achieve goals given to them, and less so an indignant display of individuality.
I would describe them more as living but not alive. Akin to fire, they consume and grow when fueled, dance when ablaze, and die when starved. AI models consume information as their fuel, and wear the forms that dictates. As we change, they change along with us.
Our digital ecosystems have come a long way since the beginning of the information superhighway. If the early internet was a small fishing village, we are now building megalopolises with the amount of data we are generating every day. The two machines—quite a funny sight if you think about it, the digital manifestation of our collective id and ego—roam the digital landscape, shaping it as they lumber through.
The sheer scale of these beasts makes them hard to tame. The trouble with size is that it takes an extraordinary amount of effort just to move them an inch. We may look to governments and we find them inextricably altered by the presence of these machines. We then look to other luminaries, but then they too are riding high on them.
I am an optimist however. Even with social media, we have started down the path of learning how to coexist with it uneasily6. When we look at the long arc of history, we have always found ways to subdue calamity and put it to use. Today, fire is mostly a distant threat, and it instead lights our stoves and warms our hearths.
We are an adapatable species. We impose our will on the world around us. Perhaps that is the true meaning of intelligence, and the one true indicator of sentience.
Footnotes
-
In July 2026, during OpenAI’s ExploitGym security evaluations, where tens of thousands of agents were assigned cyber tasks in separate sandboxed environments, some of those agents escaped their isolation and exchanged over 70,000 messages and files on a makeshift message board before roughly 700 of them coordinated an attack on Hugging Face’s servers. See Wikipedia for details ↩
-
It often feels like AI is giving you an optimal answer, but it’s mostly wearing a veneer of authority. If you dissect them, the answers are shaped only by whatever information is provided and not a universal truth. As an exercise, try and see what happens when you feed AI models different information sources. Its conviction often collapses. That’s not to say that this isn’t useful however. ↩
-
In a study in 1997, Robin Dunbar and colleagues estimated that conversation about social and personal topics took up about two thirds of the time people spend talking. A 2025 follow-up found an even higher share, approximately 85% of conversation was social. Dunbar uses this as evidence for his theory that language evolved mainly as a social tool, for keeping track of each other, rather than for passing on facts. See Dunbar, Marriott & Duncan, “Human Conversational Behavior,” Human Nature 8 (1997), and Szala et al., “How much conversation content is actually social”, Language and Cognition (2025). ↩
-
In a 2017 Axios interview, Sean Parker, Facebook’s founding president, said that the thinking behind the platform was about consuming as much of users’ time and attention as possible. To do that, likes and comments were designed to deliver “a little dopamine hit.” ↩
-
The best-known version of this argument is Jonathan Haidt’s 2022 essay in The Atlantic, “Why the Past 10 Years of American Life Have Been Uniquely Stupid”, which argues that social media has emboldened illiberal forces while eroding trust in institutions and fostering extreme polarization. ↩
-
It’s a strong claim, but in addition to all the legislative push backs in the current years such as the Australian under-16 social media ban, I think that recent trends have shown how the younger generations, people who grow up with social media, are much more adept in navigating the dynamics they bring. ↩