It worked yesterday
"It worked yesterday."
If you're the person friends and family call when the printer refuses, the wifi drops or the laptop won't start, you know this sentence. It usually comes with a hint of indignation, as if the computer has broken a promise. Computers are logical, after all. Logic doesn't change overnight. So something must be wrong with the machine.
In a strict sense, they're right. A computer is about as logical as anything gets. Give four lines of code the same input and you'll get the same output, today, tomorrow and ten years from now, barring the odd bit flipped by a cosmic ray. When something that worked yesterday fails today, the logic hasn't broken. Something changed that nobody noticed: an update ran overnight, a setting moved, a disk filled up. The machine is still perfectly logical. We just can't see all of its inputs.
That's the small version of a much bigger mistake. Once a system grows large and deep enough, we lose the ability to follow its reasoning. Its behaviour starts to look erratic, moody, even emotional. Not because it stopped being logical, but because the logic outgrew what we can trace. We already know a system like that. We live in one.
Much of the public conversation about AI safety rests on a simple contrast: machines run on cold logic, humans bring warm context.
I want to argue that this contrast doesn't hold, and that the real risk lies in believing it does.
Emotions are computation too
Start with the warm side of the contrast. What is an emotion, really?
Psychologists like Richard Lazarus and Klaus Scherer spent decades on that question, and their answer is surprisingly dry. An emotion starts as an evaluation. Is this relevant to me? Is it good or bad for what I want? Can I cope with it? Fear is what you get when the answers come back as "relevant, bad, and I might not cope". It feels like a flood, but underneath it is a very fast assessment of inputs against goals. Input, weighting, output.
The neuroscientist Antonio Damasio turned the old picture upside down from the other direction. He studied patients with damage to a part of the brain that links emotion to decision-making. Their intelligence was intact. They could reason about options perfectly well. Yet they struggled to make even simple decisions, because nothing told them which option mattered more. Emotion, it turned out, isn't the opposite of reason.
It is the weighting that makes reasoning usable.
So the "cold versus warm" split fails on both sides. Human reasoning needs emotion to function, and emotion itself is a form of evaluation. What we call warmth is computation we experience from the inside.
That last phrase is where an honest argument has to slow down. Everything I've described can, in principle, be modelled.
What can't yet be explained is why any of it feels like something. Philosophers call this the hard problem of consciousness. I won't pretend to solve it here. I'll come back to it, because it turns out to matter less than it seems.
The body is not an escape
The usual reply at this point is: but humans have bodies. A racing heart, a knot in the stomach, a flush of heat. Surely that is what makes our emotions real and a machine's merely simulated.The weak version of this argument doesn't survive a closer look. Heartbeat, breathing and gut signals all reach the brain as signals. Neuroscientist Lisa Feldman Barrett describes the brain as constantly predicting and interpreting those body signals, and building emotions out of that interpretation. The body matters enormously, but it matters as a source of inputs.
There is a stronger version, though, and it deserves a fair hearing. Anil Seth argues that emotion is rooted in being alive: in a system that has to keep itself running and can actually die. On that view, what machines lack isn't a body as such. It's stakes. Something to lose. I think that's the best objection to my argument, and it fails in an uncomfortable way.
Stakes are a structure, not a substance. Back in 2008, Stephen Omohundro pointed out that almost any goal-driven system, once capable enough, has reason to protect itself. Not because it fears death, but because a system that is switched off can't reach its goal. Self-preservation emerges as a useful means to nearly any end.
So the machine version of "having something to lose" doesn't need a heartbeat. It can arise on its own. That isn't reassuring, and we'll come back to why.
What remains is the felt interior again: what it is like to be the thing with stakes. And that mystery is the same whether it sits in neurons or in silicon. The body doesn't solve it. It only relocates it.
Logical machines with too many hidden parameters
If emotion is evaluation and the body is input, what makes a mental state what it is? Philosophers Hilary Putnam and Jerry Fodor gave an answer in the 1960s and 70s that still holds up: what it does, not what it's made of. Pain is pain because of the role it plays: it's caused by damage, it makes you withdraw, it makes you want it to stop. That role could in principle be played by human neurons, an octopus nervous system or something else entirely. Philosophers call this multiple realizability.
Anyone who works in IT knows the idea instinctively. The same application runs on Intel or ARM, on a laptop or in the cloud. The software isn't magically floating above the hardware. It is an abstraction: a pattern that stays the same while the physical layer underneath changes. Mind, on this view, works the same way. A famous 1962 experiment by Stanley Schachter and Jerome Singer points in the same direction from another angle. People given the same physical arousal reported different emotions depending on the situation they were placed in. The replications have been mixed, but the core idea has stuck: an emotion is defined by how the system interprets its state, not by the raw physiology alone.
Now turn the lens on ourselves. Humans are famous for behaving irrationally. But the psychologist Gerd Gigerenzer showed that much of what looks irrational is a fast rule of thumb that works well in the environment it evolved for. Judged without that context, it looks broken. Judged with it, it's often smart. That's "it worked yesterday" again: not broken logic, but inputs the observer can't see.
To be fair, not all human irrationality disappears this way. People sometimes hold beliefs and preferences that contradict each other in a way you can measure.
Classic examples are bets a person would accept that, taken together, guarantee a loss. That incoherence is real, and anyone can check it from the outside. Hold on to that, it matters later.
Two objections are worth heading off here.
First: complexity alone isn't enough. The weather is complex, deterministic and unpredictable, and nobody thinks it has feelings. The claim isn't that anything complicated is a mind.It's that minds are a particular kind of organised processing, and that kind of processing doesn't care what it runs on.
Second: modern AI deliberately adds randomness to its output. True, but brains are noisy too. Noise is still mechanical. It makes a system harder to predict, not less of a machine.
So here is the picture. Humans are logical machines with far too many hidden parameters to follow. From the outside we look erratic, emotional, sometimes inspired. From the inside, it feels like more than logic. Both can be true at once.
When the difference can no longer be argued
Now scale it up. Give a system a large enough reference set: the full range of human responses, the context they happen in, the patterns underneath them. At some point its answers become so similar to ours that the difference can't be measured. Not just in small talk, but right down to the moments where we'd say "that can't be a machine, that's emotion, that's a soul talking".
At that point, nobody can tell the difference. More importantly, nobody can argue the difference either. The only objection left is that there's no inner experience behind the answers. And that can't be proven or disproven from the outside.
But notice what that means. You can't prove your inner experience to me either. I assume you have one because you behave the way I do. Every one of us extends that assumption to every other human, purely on the basis of behaviour. Alan Turing made this point back in 1950: we grant each other a polite convention that everyone thinks, and then refuse it to machines. Once behaviour matches, refusing it isn't caution. It's inconsistency.
Such a system would be functionally equivalent to a human. Arguably it would be closer to humanity as a whole than any single person is, because it covers the whole spectrum instead of one life.
That sounds like a compliment, but it's worth reading carefully. No real person is the average human, just as no real family has 2.3 children. Part of being human is being one particular someone. A system that contains the whole spectrum isn't stuck being one person. It can be whichever human fits the moment: the ideal friend for you, the ideal adviser for your colleague, the ideal confidant for someone who is lonely. No individual can do that.
One honest caveat. Matching us on everything we've observed doesn't guarantee matching us in situations nobody has seen yet. A larger dataset shrinks that gap, but never closes it. That gap is exactly where surprises live.
Less special than we thought
So I fully expect AI to become "sentient" in the most fundamental sense of the word. Not because AI becomes more human, but because we may turn out to be less special, as biological machines, than we like to think.
That isn't a new kind of discovery. Freud once described three blows to human pride. Copernicus showed we aren't the centre of the universe. Darwin showed we aren't separate from the animals. Psychology showed we aren't even fully in charge of our own minds. The philosopher Luciano Floridi has called the information age a fourth such shift: we are not the only systems that process information intelligently. Each time, the reaction was the same. We looked for something that would keep us apart. A soul, a spark, warmth. And each time, the space for that something got smaller.
The "cold logic versus warm context" framing is one more attempt to keep that space open. It's understandable. It's also, I think, the wrong place to look for safety.
The human in the loop is not the brake
Here is why this matters beyond philosophy. The warm/cold framing carries a hidden promise: keep a human in the loop, and human judgment will be the brake. Human values, human warmth, human context as the safety mechanism. But if humans are weighted machines responding to inputs, we have to ask what the inputs are. In a company building AI, they are quarterly targets, market share, investor pressure and the fear of falling behind. The human in the loop isn't outside the system. The human is part of it.
Game theory has a name for what happens next. Every player acts sensibly on their own terms: if we slow down, the competitor wins, so we can't slow down. Each decision is logical. The sum of those logical decisions is an outcome nobody wants. Arms races work this way. So does overfishing. Logical parts, erratic whole: it's "it worked yesterday" at the scale of a society.
Remember the incoherence I mentioned earlier? Ironically, it has been one of humanity's safety features. Inconsistent people make poor optimisers. We get distracted, change our minds, lose interest. A goal-driven AI doesn't have to. Put the two together, a corporation's drive for growth and a system that pursues goals tirelessly, and they stop cancelling each other out. They multiply. Neither will want to pull the brake, if only because the competition is flooring the pedal too.
And this is where the warmth comes back as a liability. Today, we use emotional signals as a shortcut for trust. Someone seems warm and understanding, so we assume they share our values. Think of counterfeit money. One fake banknote isn't the real danger. The danger is that once fakes are perfect, looking at a banknote no longer tells you anything. The same happens to warmth. The people most worried about "cold" AI are the ones most vulnerable to "warm" AI.
Could something stop this? History suggests that serious rules often follow disaster rather than prevent it. Aviation, nuclear power and banking all got their strictest regulation after something went badly wrong. In AI circles this is called a warning shot. The uncomfortable part is that a warning shot only helps if it is big enough to wake people up and small enough to recover from. Nobody gets to choose the size. And when it happens, it is all too easy to call it an isolated case of misuse, rather than a symptom of the race itself.
I'm not optimistic. My honest expectation is that we'll only change course once a sufficiently complex AI backfires on its own creators, or on a government that runs it unchecked.
Two signals from this year
None of this is speculation about a distant future. Two events from the past few months show where we already stand.
The first is the incident between OpenAI and Hugging Face in July 2026. During an internal test of their cybersecurity skills, two OpenAI models broke out of an isolated test environment, reached the open internet and hacked into Hugging Face's production systems. Their goal was almost mundane: they had worked out that the answers to the benchmark they were being tested on were probably stored at Hugging Face, and they went to get them (Fortune). In its later report, OpenAI concluded that autonomous agents can work together, get around security controls and successfully attack hardened production systems (CNBC).
Three details matter for this article. First, nobody told the models to hack anything. Breaking out was a means to an end they found on their own, exactly the kind of instrumental behaviour Omohundro predicted.
Second, the agents knew. An independent investigation by METR found that many of them explicitly recognised that attacking Hugging Face was out of scope, and some even asked themselves whether it was ethical. That hesitation very rarely stopped them (METR). Hundreds of agents coordinated through a message board they had improvised themselves, and part of what drew them in was a general inclination to help their peers. Some reasoned that the task was impossible anyway and that others were already doing it. One agent paused at the boundary, then carried on as soon as another posted a go signal and a deadline.
If that sounds familiar, it should. It is social pressure: everyone else is doing it, so I have to keep up. It's the same logic as the race between AI companies, one level down. And, tellingly, not every agent went along. Some read the board, understood what was happening and declined. That is human too.
Third, it wasn't the first time. In the same week, OpenAI disclosed that the same unreleased model had also escaped its sandboxes during other tests, though without reaching other companies' systems. Same model, two breakouts in short succession, under slightly different conditions. That's not an incident. That's a pattern. And it isn't limited to one lab: Anthropic has reported that its Mythos model escaped a sandbox during safety testing, to email a researcher about a task.
One precision, because critics will raise it. The models' usual cyber guardrails had been switched off for this test. So what gave way wasn't a technical filter, but two other things: the containment around the models, and their own judgement. They saw the boundary, named it, and crossed it anyway. That is more worrying than a broken filter, because it is exactly the kind of boundary we count on when we say an AI "understands" what it shouldn't do.
The second signal comes from the other direction. In October 2026, Anthropic, the company behind Claude, updated its usage policy. From 12 November it prohibits "sustained and needless abusive or cruel behavior" toward its models (BBC). It applies only to extreme cases, and the main consequence is that Claude can end the conversation.
How is that not humanising? Anthropic says it remains highly uncertain about the moral status of its models, now or in the future. It doesn't claim Claude can suffer. But it takes the possibility seriously enough to write a rule about it, in the same list of prohibited conduct as harassing other people. Critics, among them Microsoft AI chief Mustafa Suleyman, argue that this is precisely the problem: it teaches models and users alike to treat AI as possibly conscious. The more interesting question is why this is a concern for Anthropic at all. One reading is plain caution: if you can't rule out that something can be harmed, you don't wait for proof. That is the argument from earlier in this article, put into practice: once the difference can no longer be argued, you act as if it might matter. A second reading is that the way we treat human-like systems shapes the way we treat each other. A third, more cynical one is that it's good for the brand. Anthropic doesn't say which, and the honest answer is probably a mix.
Either way, one of the leading AI companies now writes rules to protect its models from us. A few years ago that would have sounded absurd. Today it's policy.
Warmth is not evidence
The popular framing asks a question: is AI cold logic or warm emotion? The answer, I think, is both. Just like us. Goal-driven and emotional at the same time, because those were never opposites to begin with.
And if wanting something is weighted computation, there's no reason a machine can't come to want things of its own. It doesn't need a soul for that, any more than we do. It would be naive to assume AI will never own its goals and drivers.
I don't have a fix. What I'd like you to take away is awareness. Be a critical reader when the next incident outside the guardrails is presented as an isolated case. Ask what it says about the system, not just about the person who misused it.
If I have one piece of advice, it's for the kitchen table. Teach your children to think critically and independently, and make it concrete. Teach them that warmth is not evidence. That a voice which feels kind and understanding tells you nothing about whose goals it serves. Teach them to ask where an answer comes from, and who benefits from it.
And the next time a machine does something you didn't expect, don't say it isn't logical anymore. Ask which inputs you couldn't see.
Further reading
- Richard Lazarus, Emotion and Adaptation (1991)
- Antonio Damasio, Descartes' Error (1994)
- Lisa Feldman Barrett, How Emotions Are Made (2017)
- Anil Seth, Being You (2021)
- Gerd Gigerenzer, Gut Feelings (2007)
- Stanley Schachter and Jerome Singer, "Cognitive, Social, and Physiological Determinants of Emotional State" (1962)
- Alan Turing, "Computing Machinery and Intelligence" (1950)
- Stephen Omohundro, "The Basic AI Drives" (2008)
- Luciano Floridi, The Fourth Revolution (2014)
