We love the exciting AI dangers.

The model escapes the lab. It deceives its creators. It hacks critical infrastructure. It becomes autonomous. Eventually somebody mentions The Terminator.

These risks are not imaginary. But they have one enormous advantage over the AI dangers already surrounding us:

They make fantastic science fiction.

The mundane dangers are harder to dramatize. A benefits system denies the wrong person. A hiring system quietly downgrades applicants. An AI-generated error enters a medical record and gets copied forward. A law firm summarizes thousands of documents and nobody notices what the system consistently leaves out.

Then something goes wrong. The vendor says the organization was responsible for using the system. Management says a human approved the decision. The employee says the machine generated the recommendation.

Everyone was technically responsible. Nobody was responsible.

The Monster Is the Workflow

Much of the AI-safety conversation keeps staring at the model: Can it deceive us? Escape? Become autonomous? Be weaponized?

Those questions command headlines, hearings and research budgets. There is another question that sounds almost painfully dull:

What workflow did we put it into?

AI doesn't have to become autonomous to cause enormous harm. It merely has to become ordinary.

Human institutions already contain outdated rules, bad incentives, fragmented information, institutional biases and decisions nobody quite owns. AI gives those systems something new: scale.

Before AI, bad decisions were constrained by human bandwidth. Now an institution can make the same bad decision a million times before lunch.

The model doesn't need consciousness, intentions or malice. It doesn't even need to malfunction. It can perform its assigned task beautifully while the larger system produces terrible outcomes.

The danger isn't necessarily that AI refuses to follow our instructions. The danger may be that it follows them.

Human in the Loop Isn't Enough

The standard answer is to keep a human in the loop. But putting a human somewhere in a flowchart does not create accountability.

If an employee has twelve seconds to review an AI recommendation, rejecting it requires extra paperwork, and supervisors reward speed, the human is technically reviewing the machine. Functionally, the human is a rubber stamp.

Responsibility then disappears into the architecture: the model generated it, the employee approved it, the vendor built it, management bought it, and the data team chose the inputs.

This is accountability laundering.

Before Automation Comes Archaeology

Organizations should understand a process before automating it. Most don't understand their own workflows nearly as well as they think.

The official procedure says one thing. Employees do another. Exceptions accumulate. Important knowledge lives in someone's head. Workarounds become invisible infrastructure.

Then someone arrives with an AI system and says: Let's automate this.

Automate what, exactly?

Before automation comes workflow archaeology. Who makes the decision now? What information reaches them? Who handles exceptions? Who can challenge the result? What happens when sources disagree? Who can override the machine? Where is uncertainty recorded?

And most importantly: Who owns the outcome?

AI governance cannot merely govern models. It has to govern the human-machine systems into which those models are inserted.

The Boring AI Safety Agenda

The glamorous AI-safety question is: What happens if AI becomes extraordinarily powerful?

We should ask it. But another question deserves at least as much attention:

What happens when AI becomes extraordinarily ordinary?

When it quietly enters insurance, hiring, medicine, banking, policing, education, legal work and government benefits, there may never be a dramatic moment when artificial intelligence 'takes control.'

Instead, millions of tiny decisions will migrate into machine-assisted workflows. Each efficiency gain will look reasonable. Each automation proposal will have a PowerPoint explaining how much time it saves.

And we may gradually build institutions capable of making consequential decisions at unprecedented speed while making responsibility for those decisions harder and harder to locate.

No killer robots. No sentient superintelligence. No glowing red eyes.

Just databases, APIs, dashboards, procurement contracts, performance metrics and a human somewhere in the diagram labeled REVIEWER.

That's not much of a movie.

It may be a much better description of the danger.

The monster isn't necessarily the model.

The monster may be the workflow we built around it.