The guardrail should go both ways

Almost every conversation about AI safety runs in one direction. Humans put guardrails on AI. We test it, red-team it, align it, and fence it in, because we fear what it might do with power.

That work matters. But it assumes the only intelligence in the room that needs watching is the artificial one.

The guardrail should go both ways. Humans need to check AI, and AI can check humans. An intelligence with no career, reputation, or agenda of its own can do something no person in a meeting room can do reliably: point out the politics without paying a price for it.

Human intelligence was born, not built. It came out of millions of years of evolution with one rule: survive. Survival came with baggage, including status, tribe, and self-preservation. In a modern company that baggage has new names: career, reputation, agenda.

Put those forces in a meeting room and you get politics. People back the idea that helps their standing, not the one that helps the company. Bad news gets softened on its way up. Credit moves toward whoever tells the story best. Nobody has to be a villain for this to happen. The incentives make it rational.

That is the politics I mean. Not elections or geopolitics, but the everyday kind that shows up in every hiring decision, every budget request, and every strategy review.

AI is the first intelligence built rather than born. It did not come out of a survival contest, so it does not have to carry survival's baggage. It has no career to protect, no reputation to manage, no promotion to chase, and no boss to please.

That means it can be developed to be apolitical, objective, and fully honest. Not because it is morally better than us, but because it has no self to serve. An observer with nothing to gain from the outcome can say what everyone in the room is thinking and nobody will say out loud.

This is the core argument of my book, AI-Political. The danger in the AI race is not the machine. It is us, building AI in our image and infecting it with our politics. The reverse also holds. Build it without our politics, and it can help contain ours.

Imagine an objective, apolitical AI in every meeting, up and down the organization. It does not make decisions. It observes. It coordinates with the same kind of agent in other meetings, from the team stand-up to the executive offsite.

When a decision gets bent by a hidden agenda, it notices. When a project gets buried because it threatens someone's turf, it notices. When the story told in the board meeting does not match what was said in the team meeting, it can connect the two and call it out.

The obvious pushback is that humans already have guardrails. We have laws, regulators, courts, and social norms. That is true, and they work. They push most people to do the right thing most of the time.

But inside an organization, those guardrails are easy to get around, and most of what an AI guardrail protects against is not illegal. It is political. A manager hires people who should not be hired, because they are loyal rather than good. A leader asks for budget and headcount the work does not need, because a bigger team means a bigger title. An executive drives a strategy that is clearly wrong, because it protects their reputation and serves their agenda. No law is broken. No regulator will ever look. The company simply pays for it.

Those are exactly the moments an apolitical observer in the room would catch.

People do not play politics because they are bad. They play politics because it pays and the odds of getting caught are low.

Change the odds and the behavior changes. When the chance of being exposed goes up sharply, the payoff from maneuvering goes down, and the incentive to be honest and do the right thing for the company goes up. People did not become better. The system simply stopped rewarding the abuse.

Startups often win because they are small enough to contain politics. Everyone sees everything, so logic has room to run. Apolitical observers are a way to get that same visibility at the scale of a large company.

Suppose an observer removes only 20% of the politics from a single meeting. That sounds modest. Now multiply it across every meeting, every week, at every level of the company. Decisions stop getting bent in the same places over and over, and the gains compound.

The biggest effect is at the top. In my experience, political maneuvering can make up as much as 90% of what many senior executives do. That is where a small reduction buys the most, because the decisions are the largest and the politics around them are the heaviest.

Keep the guardrails on AI. We need them. But stop treating safety as a one-way street. Humans need to check AI, and AI can check humans. The machine does not have to inherit our politics. It can make ours visible, and that alone changes how we behave.