Whose values are guiding your AI – what’s its worldview?

We often hear that AI systems need to be “aligned with human values”. It sounds reassuring, but it hides a fairly obvious problem. Which humans, and which values?

There’s no single human worldview. Culture, politics, religion, family background, education, professional experience and the communities we belong to all shape the way we interpret the world. This influences concepts such as fairness, responsibility, freedom, harm, loyalty and success. Two people can look at exactly the same situation and arrive at very different conclusions, not because one of them is irrational, but because they’re starting from different assumptions about what matters.

That makes AI alignment more than a technical question. It’s also a philosophical one, and philosophy is something we often leave out of our conversations with AI.

What does “good” actually mean?

I like the work of Spencer Greenberg, founder of Clearer Thinking, who has developed what he calls the “Snow Globe Theory” of worldviews. His argument is that worldviews are stories about how the world works and what matters, and that you can understand a person’s worldview surprisingly well by asking four questions:

• What is good?
• Where do good and bad come from?
• Who deserves good?
• How can you do good?

They look like simple questions, but they reveal a great deal.

Ask different people what constitutes a good society, a good leader, a good career or even a good life and the differences emerge very quickly. Some people will emphasise freedom, others security; some individual responsibility, others collective responsibility; some achievement, others equality.

That doesn’t mean all answers are equally sound, or that evidence no longer matters. It simply means that our interpretation of evidence often sits within a much broader framework of assumptions about what is desirable, important or right.

The problem is that most of us rarely examine that framework.

We live inside our own “snow globe”, to use Greenberg’s metaphor, and what seems obvious from inside it can feel completely different from somewhere else.

This matters when we talk to AI because an LLM can produce an answer with enormous confidence, fluency and apparent reasonableness. Those qualities can easily create the impression that the answer is somehow neutral.

It isn’t necessarily neutral at all.

Question the worldview behind the answer

Imagine asking an AI: “What’s the fairest way to deal with this?”

That looks like a straightforward question, but before answering it we might ask: what definition of fairness is being used?

Fairness could mean treating everybody equally. It could mean rewarding contribution. It could mean compensating for disadvantage. It could mean protecting individual choice. It could mean prioritising the greatest benefit for the greatest number.

All of those positions rest on deeper assumptions about what is good and who deserves what.

The same applies to questions about leadership, reward, education, work, relationships or social responsibility. A response may sound highly rational while relying on assumptions we have never examined.

That is where Greenberg’s four questions become useful with AI.

When an answer feels persuasive, ask what conception of “good” appears to sit behind it. Ask where the answer assumes good or bad outcomes come from. Ask who seems to deserve benefit, protection, reward or responsibility. And ask what kind of action the answer treats as morally desirable.

It is important to recognise here that you are not trying to psychoanalyse the machine. You’re examining the worldview being expressed through the response.
This distinction matters because LLMs don’t possess a personal worldview in the same way a human does. Their outputs emerge from patterns in training data, reinforcement from human feedback, system instructions, safety rules and the immediate context of the conversation. Those influences inevitably contain assumptions and biases, just as human institutions do. US National Institute of Standards and Technology (NIST) explicitly warns that generative AI systems can scale harmful biases and produce skewed or homogenised outputs.

Confidence is not the same as truth

There is another reason this matters now more than ever. LLMs can be wrong while sounding extremely convincing.

OpenAI’s own research describes hallucinations as plausible but false statements and notes that models can confidently generate answers that are not true. The problem has reduced as systems improve, but it has not disappeared.

The greater danger, in my view, is not simply factual error. Most experienced users now know that AI can get facts wrong. The more subtle danger is accepting a confident interpretation without examining the assumptions underneath it.

A fluent answer can create cognitive ease. It feels coherent and thoughtful, and it may even echo our own beliefs back to us. That is exactly when curiosity matters most.

If an answer strongly confirms your existing worldview, it’s important to ask the AI to make the strongest case against it. If it describes something as fair, responsible or ethical, ask what alternative definitions might produce a different conclusion. If it recommends a particular course of action, ask what values appear to be driving that recommendation.

We must have a healthy scepticism about everything AI tells us, and to become less intellectually passive. Remember that AIs are scraping information indiscriminately from potentially dubious sources, and the onus is always on the end user to fact check and verify their results.

Alignment is partly our responsibility

AI developers clearly have an enormous responsibility for how these systems are trained and governed, and the debate about alignment has become more important as AI becomes more capable.

But users have a responsibility too: as said above, if we treat AI as an unquestionable authority, we surrender something important. We stop examining assumptions, testing arguments and noticing where our own worldview may be narrowing what we can see.

That was one of the themes of the first article in this series: before asking AI to help us make decisions, we need to become clearer about our own values. Building on that, we also need to be clearer on our own worldview.

We also need to become more curious about the values embedded in the answers we receive. Greenberg’s four questions above provide a surprisingly simple way to do that.

If we learn to ask those questions of ourselves, other people and increasingly the AI systems we work with, we begin to see something that can otherwise remain invisible: the worldview beneath the argument.

Once you can see the worldview, you’re in a much better position to decide whether you agree with it.

Food for thought

• Once an AI gives you an answer that feels completely reasonable, how often do you ask what values or assumptions make it appear reasonable in the first place?
• If an AI consistently confirms your existing worldview, is that evidence that it understands what you believe, or is it a good reason to become even more sceptical?

Remember when it comes to confident AI responses . . . stay curious.

David Klaasen with Alli Gibbons