The Bunker Scale: How the Ratings Work

What changed. Why it matters. What you can probably ignore.

AI risk is a strange thing to talk about because the distance between interesting and terrifying can get very small, very quickly.

A research paper shows an AI system doing something unexpected. A headline turns that into AI DEFIES HUMAN CONTROL. Someone posts it with twelve red-alert emojis, and three hours later we are apparently six months away from fighting robots for canned soup.

This is why The Bunker Scale exists.

It is not meant to predict the future with suspicious precision, and it is definitely not a countdown to the apocalypse. It is simply a way to put new AI developments in perspective.

The basic question is:

On a scale from 0 to “start digging,” how worried should we actually be?

A rating takes several things into account.

Did an AI demonstrate a genuinely new capability, or did somebody simply discover a dramatic new way to describe an existing one? Did something happen in a controlled experiment or in the real world? Was a human directing every step, or did the system act more independently? Was anyone actually harmed? Has the finding been independently confirmed? Did an AI company change its safeguards because of what happened? Did governments or regulators respond in a meaningful way?

Just as important is what we don’t know yet.

Context matters. A model behaving badly during a deliberately stressful laboratory test may be worth understanding, but that does not automatically mean AI systems are roaming the internet plotting against us. Likewise, something that causes real-world harm deserves far more attention than a chatbot saying something unsettling in a conversation.

With that in mind, here’s how the scale works.

0-2: Background Noise

These are developments that may be interesting, strange, or worth researching, but do not substantially change what we know about AI risk.

This category might include sensational headlines, isolated model behavior, speculative predictions, minor research findings, or claims that have not yet been independently supported.

A chatbot saying something creepy could land here. So could the latest declaration that AGI is arriving next Thursday.

That does not mean the story is worthless. It just means your shovel can stay comfortably in the garage.

3-4: Worth Watching

Something meaningful happened, but there is not yet enough evidence, scale, or real-world impact to justify serious concern.

Researchers may have discovered an unexpected behavior. A government may have proposed a significant new AI rule. A model may have demonstrated a new capability in testing, but with important limitations.

These are the stories worth remembering because they could lead somewhere important later.

For now, though, no need to rearrange your life.

5-6: Eyebrow Raised

This is where a development starts to change the picture.

A model might demonstrate a notable new capability. Multiple researchers may reproduce the same concerning behavior. An AI company might introduce new safeguards because its own testing revealed something it considers significant. Governments may begin responding to a risk with concrete policy rather than speeches and discussion panels.

At this level, I want to know what happens next.

I am not preparing for civilization to collapse, but one eyebrow has officially left its resting position.

7-8: Pay Attention

These ratings are reserved for substantial developments backed by credible evidence.

That could include an autonomous AI system acting outside its intended environment, a significant leap in cyber or biological capability, meaningful real-world harm, strong evidence confirmed by multiple independent groups, or major changes in frontier-lab safety procedures because of newly discovered risks.

The difference between a 7 and an 8 usually comes down to things like autonomy, real-world consequences, reproducibility, and how much the development changes what we thought current systems were capable of doing.

A 7 does not mean panic. An 8 does not mean bunker.

Both mean the same basic thing: this one deserves your attention.

9: Stock the Shelves

A 9 would require substantially more than an alarming experiment or unsettling research result.

We would be looking at a development with serious, credible, and potentially widespread consequences. That might involve major real-world harm, loss of control over a highly capable deployed system, dangerous capabilities advancing faster than existing safeguards can reasonably manage, or governments and AI companies taking emergency measures in response to a credible threat.

And no, Stock the Shelves does not literally mean I am ordering you to purchase seventeen years’ worth of canned goods.

It means we have moved beyond interesting AI safety research and into territory that could begin affecting ordinary people’s real-world decisions.

Fortunately, this should be an extremely rare rating.

10: Get in the Bunker

A 10 would require credible evidence of an immediate or rapidly developing threat with potentially catastrophic consequences that existing safeguards were struggling to contain.

We would be talking about something happening in the real world, not a frightening prediction, an alarming interview, an experimental result being stretched far beyond what it actually showed, or someone deciding their chatbot achieved consciousness because it remembered their favorite soup.

The threshold is intentionally very high.

The entire point of having a 10 is that almost nothing should qualify for it. I would be delighted if The Bunker Scale runs for the next thirty years and this category remains nothing more than an unnecessarily dramatic piece of branding.

Ratings Can Change

A Bunker Scale rating is not carved into stone.

A story might begin at a 4 and move to a 7 when independent researchers confirm the original finding. A 7 might drop to a 5 when we discover that an alarming result depended on unusually artificial testing conditions.

New evidence matters. So do corrections.

The goal is not to defend yesterday’s number. It is to make sense of what we know today.

What the Bunker Scale Is Not

The Bunker Scale is not a scientific probability of catastrophe. An 8 does not mean there is an 80 percent chance something terrible is going to happen.

It is not a judgment about whether AI itself is good or bad, and it is not meant to tell you what to believe.

Think of it as a temperature check on a developing story. I look at the evidence, context, limitations, and consequences, then explain why I think a development belongs where it does.

You are perfectly free to disagree with the number.

In fact, I hope readers occasionally do.

The rating is where the conversation starts, not where thinking is supposed to stop.

One Final Question

There is one question sitting underneath every Bunker Scale rating:

What actually changed?

If the answer is mostly the headline, probably not much.

If the answer involves what AI systems can do, what they have actually done, what access we are giving them, or how humans are responding to those capabilities, then it is worth taking a closer look.

And if something ever truly earns a 10?

Well.

I suppose we will finally have to discuss where we are building the bunker. 🌀