In the first post in this series I started with a pothole and the problem of misaligned incentives. In the second I explored network states as one possible framework for building systems with better incentives and genuine exit options. Now I want to bring the conversation back to AI itself — not as the source of our problems, but as a tool that could help us solve some of the oldest ones.

I am aware this is delicate territory. Many people are rightly concerned about AI: job loss, powerful institutions becoming more powerful, or losing control to something we do not fully understand. These fears are not unreasonable. But I want to make the case for a more specific and cautious version of AI's role, one that I think holds up under scrutiny.

A different path: narrow tools, human oversight

The path I find most promising is not racing toward a single all-knowing AI that might one day see humans as inefficient. It is building AI as networks of narrow, highly capable tools, coordinated by transparent human-written rules, with real human oversight at every important decision point.

In this model, AI does not govern us. It acts as a co-pilot, most useful exactly where humans have consistent, well-documented weaknesses. Our biggest historical weakness is our addiction to power: we concentrate it, abuse it, bend rules to protect it. We engage in lawfare, using complex legal and regulatory processes as weapons rather than tools for justice. The rich and connected often win not because they are right, but because they can afford to drag things out or exploit ambiguity.

Well-designed systems can push back against this. But only if we are precise about which tool to use and when.

The VAR problem: binary versus subjective decisions

Football offers a useful illustration. VAR, the video review system, works well for some decisions and badly for others. Offside and goal-line calls are binary: the ball either crossed the line or it did not, the attacker's shoulder was either ahead of the defender or it was not. Technology handles these cleanly. There is no room for corruption, favouritism, or a bad day at the office.

Where VAR breaks down is on subjective calls: handballs, dangerous play, whether a challenge was reckless or merely mistimed. Putting a camera on a subjective decision does not make it objective. It just slows things down and adds the illusion of precision to what is still a matter of interpretation.

This distinction maps directly onto governance.

Software is the right tool for clear, binary decisions: did this spending proposal meet the predefined criteria, was this threshold crossed, does this application qualify under these exact rules. Smart contracts and rule-based systems can process these consistently, without favouritism, and with a fully auditable trail. No one can corrupt a well-written if-statement. Estonia has been demonstrating this at national scale for two decades: tax filing, healthcare records, company registration, all processed through rule-based digital infrastructure with no civil servant touching individual decisions.

For ambiguous decisions, where reasonable people applying the same rules can reach different conclusions, AI offers something more useful than false certainty: probabilistic anchoring.

AI as an anchor, not a judge

Rather than replacing human judgement on hard calls, AI can surface what has happened historically in comparable situations. In 31 per cent of cases with these features, the outcome was X. Here are the three closest precedents. Here is how the decision broke down.

This matters because it counters one of the most insidious ways power gets abused: exploiting ambiguity. When a decision-maker knows their reasoning will be compared against a large body of similar decisions, the scope for bias and corruption narrows. The algorithm's natural tendency to find and amplify patterns becomes a feature rather than a bug: it pulls outcomes toward historical norms and away from idiosyncratic pressures.

This is not AI making the call. It is AI making it harder to make a corrupt call without it being visible.

The caveat that matters: biased history, changing rules

There are two problems with this approach that need to be named honestly.

The bias baked into the data

The first is that historical data can be biased. COMPAS, the algorithmic risk-assessment tool used in sentencing decisions across parts of the United States, did not include race as a variable. But it was trained on historical decisions shaped by decades of systemic disparity, and those patterns leaked back in through correlated factors like neighbourhood and prior conviction history. The crime itself got lost behind the profile of who committed it.

A better-designed system starts from the other direction: only verified, directly relevant facts about the event itself enter the model. The handball does not care who the player is. Neither should the system assessing it, and that principle should be enforced at the design level, not left as an aspiration. One concrete approach is to train on anonymised reconstructions of events, the equivalent of using featureless figures to re-enact an incident, so the model learns the mechanics of what happened rather than patterns associated with who was involved.

Any system using historical precedent as an anchor must treat its inputs as auditable and challengeable, not as ground truth. The inputs need scrutiny, not just the outputs.

Rules that drift over time

The second is that rules and norms change over time. Football is a good example again: what counted as a tough but fair tackle in the 1990s would earn a red card today for excessive force. The game has become less physical, and the rules have been reworded repeatedly to reflect that. An AI trained on historical decisions without accounting for when those decisions were made, and under which interpretation of the rules, would produce misleading guidance. Any useful system needs to understand not just what was decided, but what rules were in force at the time and how their interpretation has drifted.

This makes building these systems harder. It does not make them impossible, and it does not make the underlying approach wrong. It just means they require care.

What this looks like in practice

In a DAO or network state, this framework might look like:

Some of this is already being tested. Kleros runs community-based dispute resolution for blockchain conflicts under transparent, auditable rules, with jurors held accountable by tokenised incentives rather than institutional authority.

The humans remain in charge of values, goals, and final appeals. The software and AI handle the parts where human judgement has historically been least reliable.

Why this matters beyond governance

If we can build systems where power is decentralised, rules are transparent, and the scope for corrupting a decision is structurally reduced, we lower the prize that attracts bad actors in the first place. A monopoly on power becomes less valuable when the rules are harder to bend.

This is not a utopia, and AI is not a magic neutraliser. It has its own biases, its own owners, and its own failure modes. But used narrowly, within clear boundaries, with auditable inputs and human oversight, it offers something we have never quite had before: a way to make certain kinds of corruption expensive and visible rather than cheap and invisible.

We do not need perfect AI. We need good enough AI, applied precisely, in the right places.

In the next post I want to look at the practical transition: how we might share the productivity gains from AI, different models of ownership, and what role tools like Bitcoin could play in building more sovereign, voluntary communities.

Until then, the question I keep returning to is this: if we can design systems that make corruption structurally harder, why have we so rarely chosen to build them that way?