AI Risk Index Methodology
v1.1 · published since 17 September 2026The AI Risk Index is a daily 0–100 measure of the risk that artificial intelligence causes serious harm or acts beyond human control, computed by World Tension Meter’s own engine from world news. This page is the permanent, versioned record of how it works.
Who publishes it. World Tension Meter, an independent open observatory. First daily score: 17 September 2026. Every number since then is in the public dataset and never revised.
What it reads. Forty-one public news feeds, read twice a day. Every headline gets an AI-risk weight from 0 to 99: escapes, self-replication and AI-enabled attacks weigh most; capability jumps, deepfakes and military uses weigh high; regulation and policy stories weigh moderate. A language model sharpens the weights when available; the keyword tiers carry on alone when it is not, so the index never stops.
The headline score
Three components, fixed weights. The share of AI headlines that read as risk events (40%), the severity of the strongest eight headlines (40%), and public attention on AI-safety topics ranked against three years of history (20%).
Smoothing. Today’s reading is averaged 50/50 with yesterday’s, so one viral story cannot swing the index ten points and back in a day, while a real trend cannot hide.
Zones. 0–39 calm, 40–54 stable, 55–67 moderate, 68–77 high, 78+ extreme — the same five zones as every WTM index.
The seven sub-indexes (v1.1)
Every AI headline is sorted into up to three risk categories by term matching. Each category’s sub-index is 100 × (1 − e−I/1.5), where I is the average per-day weighted headline count over 30 days (a headline weighs its AI-risk score ÷ 100). One and a half weighted headlines a day reads 63; four a day reads about 93.
- Misuse & cyber. deepfake, deep fake, scam, fraud, voice clone, cloned voice…
- Capability jumps. frontier model, agi, superintelligen, new model, benchmark, outperform…
- Regulation & control. ai act, regulat, lawsuit, sues, sued, copyright…
- Military AI. military ai, autonomous weapon, autonomous drone, ai weapon, pentagon, targeting…
- Safety incidents. safety incident, misaligned, resisted shutdown, escaped, self-replicat, copied itself…
- Jobs & economy. layoff, lay off, replaced by ai, jobs, workforce, unemployment…
- Chips & power. chip, gpu, nvidia, data center, data centre, datacenter…
Incidents. A headline weighing 55 or more counts as an AI incident: harm done, attempted or narrowly avoided. The page shows the 30-day count.
Country attribution. Each AI headline is tagged with the countries it names, using the same registry as our country pages, so you can see where AI risk is being made.
Version history
| Date | Version | Change |
|---|---|---|
| 18 September 2026 | v1.1 | Seven risk-category sub-indexes, AI incidents counter, country attribution, 30-day AI headline store, public dataset, badge and citation formats. |
| 17 September 2026 | v1.0 | Daily AI Risk Index published: AI-risk share of coverage 40%, top-headline severity 40%, public attention on AI-safety topics 20%; EMA 0.5 smoothing; same 0-100 zones as every WTM index. |
How to cite
World Tension Meter (2026). AI Risk Index, daily 0–100 score, v1.1. worldtensionmeter.com/ai-risk/. Licence CC BY 4.0. Machine-readable: ai_risk.json.
Questions people ask
Is this the first numeric AI risk index?
It is the first we know of that publishes a daily 0–100 score from live news with an open method, a permanent unrevised record and a public dataset. The publication date above is the record.
Does it predict an AI catastrophe?
No. It measures how much of the world’s reporting is about AI going wrong, and how severe those stories are. Treat it as a barometer, not a forecast.
Can I use the number in my product or paper?
Yes, free of charge, with attribution to World Tension Meter and a link. The badge and the JSON snapshot exist for exactly that.
Why is it not inside the World Tension Index?
Mixing a new signal into a score with a longer record would break comparability. It stands alone until it has a full year of data.