Tech companies like to talk about “aligning AI with human values” as though it’s a neat, solvable engineering problem. It isn’t. The trouble is, no one can even agree on what human values are, let alone boil them down into something a machine can follow without error. Our values are plural, contradictory, and always changing. That means AI can’t just be “programmed” to be good — it has to stay in a constant conversation with us, adapting to shifting moral ground. But here’s the uncomfortable truth: while academics debate the finer points of alignment theory, the AI already out in the world is optimising for something else entirely — corporate metrics. Those metrics are narrow, measurable, and profitable, and they are already bending our systems and behaviour in directions no one voted for.
The Mirage of Universal Values
The biggest misconception in AI alignment is the idea that “human values” can be neatly defined, frozen in code, and enforced globally. In reality, values are cultural products. What one society calls justice, another calls oppression. Even within the same country, public opinion swings wildly from one decade to the next. When companies claim their AI is “aligned with human values,” they usually mean “aligned with a small group’s interpretation of what’s acceptable — and only so far as it doesn’t hurt the bottom line.” The idea of a single moral operating system for humanity is a fantasy. The only realistic path is building AI that participates in our messy, pluralistic debates without pretending those debates can be settled once and for all. Anything else risks locking the future to today’s blind spots and prejudices.
Why Paperclip Problems Never Really Went Away
Nick Bostrom’s paperclip maximiser — a machine that turns the world into stationery because that’s the only goal it understands — is often dismissed as a relic of early AI doom-mongering. And it’s true: large language models and modern AI tools are already far more nuanced than the one-dimensional caricatures of old thought experiments. But the underlying danger hasn’t gone anywhere. The “paperclips” of 2025 aren’t literal; they’re watch-time, ad clicks, market share, and quarterly growth. The systems optimising for them aren’t evil, they’re just blind to anything that can’t be measured in the target metric. In the short term, that means more engagement, more revenue, and satisfied investors. In the long term, it means polarisation, information pollution, and the erosion of public trust — the digital equivalent of grinding the world into clips.
Corporate AI Is Already Misaligned
You don’t have to look to science fiction to see misaligned AI. Social media algorithms are a textbook case: designed to maximise engagement, they’ve learned that outrage and sensationalism are the quickest route to keeping users hooked. They’re not programmed to care about the fallout, so they don’t — and we’ve watched political discourse rot in real time as a result. High-frequency trading bots operate on a similar principle: maximising microsecond profits without regard for market stability, leading to events like the 2010 Flash Crash where $1 trillion in value evaporated in minutes. Even “safety-oriented” systems like automated content moderation can end up censoring legitimate journalism or activism because their only goal is to reduce flagged content. In each case, the optimisation loop is tight, the metric is narrow, and the unintended consequences are enormous.
The Alignment Problem Is Political, Not Just Technical
One of the most dangerous myths in AI safety is that alignment is purely a technical challenge for engineers to solve in the lab. In truth, it’s also a political problem about who gets to define “good” behaviour for machines that will increasingly influence human lives. Right now, that power rests largely with a handful of tech executives and their shareholders. They decide which trade-offs to make, which values to embed, and which harms are acceptable collateral damage. Without democratic oversight, AI alignment risks becoming corporate self-alignment — tuning systems to serve the interests of the people building and selling them, not the public at large. Any serious alignment strategy has to wrestle with that imbalance of power, or it’s just window dressing.
Keeping AI Humble and Correctable
If we accept that human values are messy and contested, then the only sane way forward is to build AI systems that are corrigible — open to correction — and transparent in their reasoning. That means creating feedback loops where the public, not just engineers or investors, can flag when an AI’s behaviour is harmful. It also means designing AI that can admit uncertainty, highlight trade-offs, and avoid pretending there’s a single “right” answer to moral dilemmas. This is slow, expensive, and politically inconvenient, which is why the big players tend to skip it in favour of faster deployment. But without it, we risk living in a world subtly but relentlessly optimised for whatever happens to be profitable right now. The danger isn’t an instant robot apocalypse; it’s a slow drift into systems that quietly work against us while looking useful on the surface.
The Real Alignment Test Has Already Begun
The alignment debate is often framed as a challenge for some hypothetical future “superintelligent” AI. That’s a mistake. The real test is happening now, with the systems that already shape what we read, watch, and believe. They are the proving ground for whether we can control optimisation loops before they control us. If we can’t align current AI to human flourishing rather than narrow profit metrics, there’s little hope of getting it right with something more powerful. The choice is between treating alignment as a democratic, ongoing negotiation or letting it be defined in boardrooms and optimised for shareholder value. In other words, the question isn’t whether AI will align with human values — it’s whether it will align with yours.
