Doom Index vs p(doom)

The shortest explanation

A Doom Index value of 44 does not mean a 44% probability of catastrophe. It means the source-backed evidence assessed by DoomBench currently places its qualitative trajectory at 44 on a scale from no meaningful AI capability to the terminal AI takeover scenario.

What the Doom Index measures

The Doom Index records whether dated evidence moves toward or away from loss of human control. Every item has a direction, assessed magnitude, confidence, category, company attribution, and relevant model attribution. The chronology can be revised when older evidence is backfilled or a judgment changes.

This makes the index useful for comparing pressure over time and seeing which developments moved it. It is an editorial measurement framework, not a calibrated forecast of how often catastrophe will occur.

What p(doom) measures

p(doom) is shorthand for the probability a person or model assigns to a catastrophic AI outcome. Estimates are only comparable when they define the same outcome, time horizon, assumptions, and level of catastrophe. One person may mean human extinction, while another includes permanent disempowerment or a less terminal global disaster.

Because those definitions and prior beliefs differ, two careful readers can inspect the same evidence and reach different p(doom) estimates.

How to use DoomBench to update an estimate

1. Write down your current p(doom) before reviewing the latest movement.
2. Define your outcome and time horizon so the estimate has a stable meaning.
3. Open the evidence items that moved the index and inspect their original sources.
4. Ask how reliable, surprising, and independent each item is.
5. Use categories, companies, and model versions to identify concentration or repeated signals.
6. Update in the direction and by the amount the evidence justifies, then record why.

The index movement is a prompt to reassess evidence, not a number to add directly to your probability estimate.

An example update

Suppose the Doom Index rises from 40 to 44 after several capability and autonomy developments. A reader should not automatically add four percentage points to their p(doom). They should inspect whether those developments were genuinely new, whether they came from independent evidence, and whether safeguards or counterevidence changed at the same time.

One reader might make a modest upward probability update. Another might leave their estimate unchanged because the news was already anticipated. Both responses can be coherent if their priors and reasoning are explicit.

Important limits

• There is no direct conversion from a Doom Index value to p(doom).
• A backfill can revise historical index values without representing a new event today.
• Several headlines may trace back to one underlying development and should not be counted as independent evidence.
• Model Doom Scores compare model risk profiles. They are also not probabilities.
• DoomBench provides evidence and transparent judgments, not a required personal estimate.