If you spend even a small amount of time reading about artificial intelligence, you will know that everyone says they want “Ethical AI”.
Governments want it. Tech companies want it. Universities want it. Consultants definitely want it.
In my opinion, however, saying you’re “in favour of ethical AI” is a bit like saying you hope the sun comes up tomorrow and then it keeps us warm. It signals basic sanity, but it is not useful insight. So the real question isn’t whether AI should be ethical.
The real question is: What do people actually mean by it, and what, if anything, changes because of that understanding?
Ethical AI, as Usually Meant
Most of the time, “Ethical AI” seems to mean one of three things:
- A list of values: Transparency. Accountability. Safety. Human-centricity. Everyone nods. Nobody disagrees. Nothing is tested.
- A posture: “We take ethics seriously.” Often accompanied by a PDF, a committee, and a photo of diverse people sitting around a table.
- A reassurance: Something you say so others don’t worry too much about what you’re building.
None of these are bad. They’re just … thin.
They tell us what people hope for, not what happens when things become hard to judge, or when they go wrong. And that’s precisely when ethics is supposed to matter.
The Awkward Truth
Here’s the uncomfortable bit: Most ethical failures don’t happen because people don’t care about ethics. They happen because ethical language outruns structure, discipline, and responsibility. In other words, ethics fails after we agree it is important, not before that moment of enlightenment.
Everyone agrees they want fairness – until fairness pulls in opposite directions. Everyone wants transparency – until it conflicts with speed, profit, or security. And everyone wants accountability – until it’s unclear who, exactly, should carry it and what it might cost.
At that point, ethical talk doesn’t disappear. It just becomes performative.
Ethical AI as Theatre
We are now very good at sounding “ethical “when we talk about AI.
AI systems can explain themselves fluently. Institutions can publish values frameworks elegantly. Reports can acknowledge uncertainty while quietly smoothing the issues away. This is dangerous, not because it’s dishonest, but because it’s convincing. It isn’t just LLMs that are convincing even as they hallucinate. Humans do that, too.
When ethical language becomes fluent enough, it stops being questioned. It creates the feeling of responsibility, without the burden of actually doing something about it. And that’s how ethics turns into theatre.
So, What Would Ethical AI Actually Require?
It is not about identifying and listing better values. It is not another framework diagram. Nor is it about adding a list of socially good practices in some kind of LLM “values alignment” layer. Frankly, that effort tends to be about talking to your customers in a way that they like, rather than talking about the right things.
“Ethical AI”, if it is to mean anything, requires constraints and discipline.
At minimum, it needs
- Clear boundaries on what the system can and cannot claim.
- Explicit handling of uncertainty, rather than optimistic smoothing.
- Retention of responsibility at critical junctures that can’t be delegated.
- Structures that slow things down when clarity is missing, and that highlight tensions.
- Mechanisms that surface disagreement rather than rushing to consensus.
In short: Ethics has to be operational, not aspirational.
The Hard Part Everyone Avoids
The genuinely hard work of ethical AI isn’t deciding what’s right.
It’s deciding:
- Who is responsible when there are no good, easy options.
- How to act when values collide and consensus can’t be reached.
- How to stop persuasive language from standing in for judgment.
- How to admit “we don’t know” without losing face.
That’s slow. It’s uncomfortable. It doesn’t fit neatly into product roadmaps.
Which is why so much “Ethical AI” talk floats safely above it.
Does Anyone Actually Do This?
That is a fair question. Diagnosing what’s broken is easier than building what works.
But yes, there are emerging approaches that take these requirements seriously. These approaches don’t start from values. They start from failure modes, and they tend to share several characteristics:
- Protocol governance: Explicit rules about what can/cannot be claimed, with violations triggering correction rather than being smoothed over.
- Uncertainty quantification: Using frameworks to provide confidence bounds rather than probabilistic “hand-waving”.
- Multi-perspectival interrogation: Deliberate inclusion of conflicting frameworks to surface disagreement rather than manufacture consensus.
- Human responsibility preservation: Ensuring AI systems are very competent tools within bounded domains, not autonomous decision-makers.
- Conceptual vacancy detection: Building mechanisms that flag when key terms (“fairness,” “alignment,” “trust”) are doing unexplained work.
These aren’t theoretical proposals. They’re operational methodologies that are gradually being validated and applied to real-world governance questions.
They’re slower than standard AI workflows. They require epistemic discipline that feels uncomfortable. They produce bounded insights rather than confident predictions.
Which is exactly the point.
The test of “Ethical AI” isn’t whether it sounds ethical. It’s whether the methodology catches itself when it starts to bullshit, and whether humans retain the authority to act on what’s caught.
I’ve spent the last year researching and testing such methodologies. They work. They are awkward. They surface inconvenient disagreements. They refuse to claim certainty where none exists. Such approaches aren’t designed to be popular with people who want AI to make decisions faster, sound more confident, or bypass the hard work of human judgment. But for people doing work where being confidently wrong is worse than being honestly uncertain, it turns out there’s real demand for systems that know how to shut up when they should.
A Better Question
So maybe the question isn’t:
“Is this AI ethical?”
But:
“What stops this system from sounding more ethical than it actually is?”
That’s the question worth answering.
Because wanting “Ethical AI” is easy. Building systems that remain ethical under pressure is not. And if we don’t learn to tell the difference, “Ethical AI” will keep meaning everything, and therefore mean nothing.
Contact me if you’d like to continue this conversation.