How “AI Doomer” Became a Label in the AI Risk Debate

The word doomer sounds like someone who has already bought canned food.

The reality is less tidy.

In AI conversations, AI doomer can mean someone who thinks advanced AI might pose a serious extinction risk.

It can mean someone who wants stronger safety rules.

It can mean someone who thinks development is moving too quickly.

And sometimes it means:

a person whose level of concern is higher than the speaker thinks is reasonable.

That flexibility is why the term became useful.

It is also why it is a poor substitute for asking what somebody actually believes.

Before AI, there were already doomers

Doomer did not begin with ChatGPT.

It was already internet slang for someone with a bleak view of the future — someone expecting environmental, economic or social decline rather than a smooth continuation of normal life.

AI supplied a new possible catastrophe.

As increasingly capable language models reached the public, older arguments from AI safety and alignment communities became much easier to imagine.

What happens if AI systems become more capable than expected?

What if humans cannot reliably control them?

What if a system pursues a goal in ways its creators did not intend?

These questions had existed for years.

What changed after ChatGPT was the audience.

The hypothetical machine was no longer hiding inside research papers.

Millions of people had one open in a browser tab.

2023 turned the label mainstream

The year 2023 was when AI doomer became a recognizable mainstream media label.

GPT-4 arrived.

A prominent open letter called for a pause in the development of more powerful AI systems.

Researchers and executives were publicly discussing risks ranging from misinformation and cyberattacks to human extinction.

Then Geoffrey Hinton left Google.

Hinton is one of the foundational figures in modern deep learning. In May 2023, WIRED published a profile titled:

“What Really Made Geoffrey Hinton Into an AI Doomer.”

The headline showed how far the language had travelled.

But the article also demonstrated why the label can oversimplify.

Hinton was worried about increasingly capable AI and argued that much more effort should go into preventing serious harms.

At the same time, he did not call for AI development to stop and said the technology could bring major benefits.

Put those positions into a paragraph and you get nuance.

Put them into an online argument and you may get:

doomer.

Then people started asking for a number

AI risk culture developed another piece of vocabulary:

P(doom).

The phrase means, roughly, the probability someone assigns to an AI-related catastrophic outcome.

There is no single universally agreed event inside the parentheses. Different people may mean human extinction, irreversible loss of control, civilizational catastrophe or something slightly different.

That ambiguity did not stop the number from becoming socially useful.

Asking for someone’s P(doom) turns a sprawling philosophical argument into something that looks wonderfully precise:

What’s your P(doom)?

Ten percent.

That’s high.

Is it?

The number gives people something to compare even when the assumptions behind the numbers differ.

By December 2023, the phrase had reached Washington.

Axios reported that then-Senate Majority Leader Chuck Schumer opened an AI Insight Forum by asking participants for their P(doom), described there as the probability that AGI would be disastrous for humanity.

An internet-community shorthand had become part of a Senate AI discussion.

The internet prefers two teams

A complicated debate becomes easier to post about once both sides have names.

On one side:

doomers.

On the other:

accelerationists.

In 2023, e/acc — short for effective accelerationism — became a visible online label associated with faster technological progress and resistance to what supporters saw as excessive pessimism or restriction.

During the November 2023 upheaval at OpenAI, Marc Andreessen and others publicly used e/acc language while criticizing the “doomer” movement.

The internet had acquired a clean opposition:

e/acc vs. doomers

It worked well as a meme.

It worked less well as a map of what people actually believed.

Someone could think advanced AI posed serious risks and still oppose a pause.

Someone could favor rapid technological progress while supporting substantial safety research.

Two researchers with similar P(doom) estimates could disagree completely about regulation.

And some prominent people rejected the categories altogether.

When WIRED asked Meta chief AI scientist Yann LeCun whether he considered himself an accelerationist rather than a doomer, he said he did not like either label.

Labels, inconveniently, do not require the consent of the labeled.

“AI safety person” and “AI doomer” are not synonyms

This is one of the easiest mistakes to make from outside the debate.

AI safety is a broad field.

It can include practical engineering problems such as making systems follow instructions reliably, preventing harmful outputs or protecting agents from security attacks.

It can also include research into much more extreme scenarios involving future systems that might exceed human capabilities.

Calling everyone working on those problems a doomer would erase distinctions that matter.

Likewise, not everyone described as a doomer believes the same catastrophe is likely or favors the same response.

Some focus on technical alignment.

Some focus on governance.

Some want slower development.

Some want better evaluations and security.

Some simply think organizations should be more cautious about deploying systems they do not fully understand.

A label that can cover all of those positions is socially powerful precisely because it is imprecise.

Sometimes “doomer” describes the speaker too

This is one of the more interesting things about the word.

Suppose someone says:

Maya is an AI doomer.

You may have learned something about Maya.

But you have also learned something about the person speaking.

They consider Maya’s level of concern high enough to deserve a label.

That makes doomer partly relational.

The same person might look alarmist in one group and moderate in another.

Among people who think catastrophic AI risk is negligible, a 5% P(doom) may sound extreme.

Among people who think the risk is much higher, the same estimate may sound optimistic.

The word behaves less like a technical category and more like a position on an invisible cultural map.

Where the border sits depends partly on where you are standing.

The label survived because the argument survived

AI doomer could have remained a phrase tied to the first wave of ChatGPT anxiety.

It did not.

The label has continued to appear in later reporting and public debates about frontier models, AI safety and catastrophic risk.

That persistence makes sense because the underlying disagreement never disappeared.

As models became more capable, the details changed but the questions remained:

How powerful will these systems become?

How quickly?

How controllable?

Who decides what level of risk is acceptable?

Those are not slang questions.

But slang is what people use while arguing about them.

A useful word, if you remember how imprecise it is

AI doomer works because it compresses a complicated position into two words.

That is what slang is good at.

It saves time.

It signals a stance.

It carries attitude.

It also removes detail.

If you actually want to understand a supposed AI doomer, the label is only the beginning.

The better questions are:

What outcome are they worried about?

How likely do they think it is?

What evidence would change their mind?

And what do they think anyone should do about it?

Those answers tell you far more than the label.

They are just harder to fit into a post.