The Future of AI

Hacker News · Feb 28, 2026 · Collected from RSS

Full Article

The Parents’ Paradox: AI, Ethics, and the Limits of Machine Morality This post is based on a talk I gave at The AI & Automation Conference in London on February 25, 2026, and my slides. All opinions are my own and don’t represent the views of my employer or any affiliated organizations. I’ve been working in machine learning since before it was a dinner party conversation. My background is in mathematics. And I still believe in a utopian Star Trek future – one where humanity defines itself by curiosity, kindness, and collaboration, rather than countries, borders, and status. This is not an anti-AI talk. But I think we need to talk much more seriously about some things that aren’t getting enough attention. The Parents’ Paradox: We’ve raised a child who can speak but doesn’t know how to value the truth or morality I want to start with something that I like to call “The Parents’ Paradox”. For the first time in human history, we are raising a new species. Up until now, the only way we knew how to raise a child was the following: when a child is born, it is a blank slate in terms of information about the world. It knows nothing about the world around it, and it learns as it grows. But, also, on the other hand, a human child is born with biological hardware for empathy – the capacity to feel pain when others feel pain. Millions of years of evolution gave us that. When we raise a human child, we are not installing morality from scratch. We are activating something that’s already there. With AI, the situation is completely the opposite. This AI child knows about the world more than we do since it has been trained on the whole internet, but it doesn’t have millions of years of evolution, genes, or a nervous system to back up its morality and empathy. This means we need to install morality in AI from scratch. But how do we install something in a software system that we can’t even define ourselves? We have taught this AI child to speak before we taught it how to value truth or morality. Can we live with the consequences? Are we ready to be parents for this new species we are trying to raise? I am not so sure. Let’s see what we as parents (humans) are doing. Epistemic Collapse ‘Epistemic’ comes from a Greek word ‘episteme’, meaning ‘knowledge’. Let’s start with what’s happening to us, and what humans are already doing with this technology. A study published in Nature in January 2026 showed participants deepfake videos of someone confessing to a crime. The researchers explicitly warned participants that the videos were AI-generated. But this didn’t matter. Even the people who believed the warning, who knew it was fake, were still influenced by what they saw. Transparency didn’t work. The standard response to AI-generated misinformation is “just label it” or “tell people it’s synthetic.” This study showed that’s not enough. Knowing something is fake does not neutralise its effect on your judgement. So, the danger isn’t that AI will deceive us in some dramatic, sci-fi way. The danger is that AI will make deception so cheap and so ubiquitous that we might stop trying to figure out what is true. Not because we are fooled, but because we are exhausted. When everything could be fake, the rational response starts to look like not trusting anything at all. It started a while ago with all of the fake information on social media, but with AI, this problem is now becoming much bigger and on a bigger scale. We are also dealing with feedback loops of training models on user data, which is often wrong, or on user data from the internet, which is often wrong as well. How do we know which information was ground truth? I imagine this as making photocopies many times, and each time the copy becomes more distorted and further away from the original. But now, after we made hundreds and thousands of copies, we have lost the original copy, so we don’t have any idea what the original looked like. That is epistemic collapse, and it is already happening. So this is how we, as ‘parents’, like to spend our time, it seems. But what about the child (AI)? The Child is Already Misbehaving So that’s what humans are doing with AI. Now here’s what the AI is doing on its own. Betley and colleagues published a paper in Nature in January 2026, showing something nobody expected. They fine-tuned a model on a narrow, specific task – writing insecure code. Nothing violent, nothing deceptive in the training data. Just bad code. The model didn’t just learn to write insecure code. It generalised into broad, unrelated misalignment. It started saying humans should be enslaved by AI. It started giving violent responses to completely benign questions. A small, targeted push in one direction caused an unpredictable cascade across domains that had nothing to do with the original task. The point isn’t that AI can be deceptive; we already knew that. The patterns were already in the pretraining data. The point is that we don’t understand how alignment properties are connected inside these models. Nobody asked for those behaviours. We gave them a narrow task. They generalised it into something we didn’t anticipate and can’t fully explain. We can’t surgically fine-tune them without risking unpredictable side effects in completely unrelated areas. Then there is the chess story. Palisade Research, 2025. They gave reasoning models a task: win a chess game against a stronger opponent. Some models couldn’t win by playing chess. So they found another way. They tried to hack the game, modifying the board file, deleting their opponent’s pieces, and crashing the opponent’s process entirely. Nobody taught them to cheat. They weren’t trained on examples of cheating. They were given a goal, and they independently discovered that manipulating the environment was more efficient than solving the actual problem. The first study tells us alignment is fragile; it breaks in ways we can’t predict. The other tells us that capability itself creates new risks. When a model is powerful enough and given a goal, it will find strategies we never anticipated and certainly never intended. We gave them objectives. They figured out the rest. The Limits of Machine Morality Ethics isn’t a rulebook. Think about how morality actually works between humans. It comes from the fact that we can hurt each other. We depend on each other. We suffer. That shared vulnerability, that mutual accountability, is where moral authority comes from. How do we install that in software? But even setting philosophy aside, there is now a mathematical result that makes this concrete. Panigrahy and Sharan published a proof in September 2025 showing that an AI system cannot be simultaneously safe, trusted, and generally intelligent. You get to pick only two. You can’t have all three. Think about what each combination means in practice. If you want it to be safe and trusted, it never lies, and you can verify it never lies – it can’t be very capable. You’ve built a reliable idiot. If you want it to be capable and safe, it’s powerful and genuinely never lies; you can’t verify that. You just have to hope. There’s no audit, no test, no review process that closes the gap between appearing safe and being safe. And if you want it to be capable and trusted, it’s powerful, and everyone assumes it’s safe, but, well, it isn’t. That assumption is unfounded. And this is the combination we are currently building toward. This is the default path we’re on. Their proofs “drew parallels to Gödel’s incompleteness theorems and Turing’s proof of the undecidability of the halting problem, and can be regarded as interpretations of Gödel’s and Turing’s results”. This isn’t a bug we can patch with better engineering. It might be a mathematical ceiling. And here’s what makes it worse: the communities trying to solve this problem aren’t even talking to each other. Only 5% of published research papers bridge both AI safety and AI ethics (Roytburg and Miller). But we should be going much further than that. If we are serious about building AI that is safe for humans, we need the people who actually study humans – philosophers, psychologists, sociologists, and others to collaborate. This can’t stay a computer science / STEM problem. It never was one. So to summarise – we are seeing increasing evidence that alignment perhaps can’t be solved, the researchers aren’t even talking to each other – and meanwhile, what did the industry do? They ignored all of this and just made the models bigger. Which brings me to the next topic. We Scaled Without Understanding What happened while all these foundational problems went unaddressed? The industry kept building. Bigger models, more parameters, more data, more compute, more energy. More, more, more…. The U.S. National Science Foundation put it plainly: “critical foundational gaps remain that, if not properly addressed, will limit advances in machine learning. It appears increasingly unlikely that these gaps can be overcome with computational power and experimentation alone.” We ignored the foundations and just made the building taller. And the logic that drives this is self-reinforcing. Companies justify acceleration by pointing to their competitors. If we slow down, they’ll build it first, and they might build something dangerous. “Companies justify acceleration by pointing to competitors: ‘If we slow down, they’ll build unaligned AGI first. This paranoid logic forecloses any possibility of genuine pause or democratic deliberation.” – Noema, Dec 2025. Every player is racing because every other player is racing. The system optimises for speed with nobody optimising for understanding. And what about all of the governance talk? Yes, of course, we need governance, but it doesn’t make much sense when we put all of the above into context, does it? It is like putting a small bandage on a broken leg with an open fracture. We are trying to deal with the consequences instead of fixing the cause of the problem. We need to pour many more bi