AI Catastrophic Risk Calculator & Threshold Explorer

Anthropic Safety Report Horizon

Existential Risk Horizon Projection (0–25 Years)

Cumulative probability trajectory modeling cumulative threshold exceedance vs. alignment mitigation.

Threshold Watch
Elevated Risk Threshold Exceeded 10-Year Risk: 14.2%
10-Year Projected Risk
14.2%
Critical Horizon Exceedance
8.4 yrs
Mitigation Confidence
62.1%
IGN Report 10% Marker
+4.2%
Projected Risk Curve (P_existential)
Confidence Interval Band (±1σ)
Anthropic >10% Decade Threshold
Mitigation Buffer Intersection

Model Parameters

Adjust dynamic levers
4.2
Annual effective FLOP expansion across frontier autonomous clusters.
1.1
Rate of mechanistic interpretability and steering verification progress.
10.5%
Prior anchor from cited Anthropic researcher threshold (>10%).
78%
Maximum unmonitored decision-making scope allowed before interlocks engage.
10 yrs
Policy lag and global regulatory response coordination window.

Source Citation & Evidence Context

“An Anthropic safety researcher has said there is a greater than 10% chance that AI ‘could kill all humans’ within the next decade.”
Source Account: @IGN (Verified X Monetization Post)
Canonical Status ID: 2097656369390166496
Original Article Link: IGN News Article Coverage
Analytical Purpose: Explores the sensitivity of the stated >10% probability estimate against differential rates of compute scaling vs. alignment safety research.

Model Assumptions & Methodological Limitations

1. Hazard Function: Risk is modeled as a cumulative hazard function where compute acceleration compounds faster than linear alignment velocity unless capped by strict international governance.

2. Subjective Probabilities: Frontier researcher statements reflect subjective Bayesian beliefs under incomplete visibility into autonomous agency and latent goal alignment.

3. Countermeasures: Increased mitigation confidence (green metric) dynamically compresses the hazard accumulation rate over the 10 to 20-year horizon.

A local risk fixture differs from its assigned hazard curve

Read the explanation

The saved default assigns compute four point two, alignment one point one, baseline risk ten point five, autonomy seventy eight and governance ten years. Its representative branch directly returns fourteen point two percent risk, eight point four year horizon and sixty two point one mitigation. The generic mitigation formula instead gives alignment contribution thirty three, governance eighteen and autonomy five point five, totaling fifty six point five. At three pixels per point forced mitigation sixty two point one measures one hundred eighty six point three, generic fifty six point five one hundred sixty nine point five and difference five point six sixteen point eight. Moving just outside the representative tolerances changes branches. Neither branch runs a safety model or measures real extinction probability; these are source-authored assumptions with a special default override. The risk curve computes a hazard constant from one minus assigned decade risk, using negative natural logarithm divided by ten. But time enters as time raised to one point zero eight. At ten years, the exponent is about twelve point zero two, not ten. With assigned risk fourteen point two percent the drawn probability is about sixteen point eight one seven percent at ten years, a difference about two point six one seven. This plot uses twenty horizontal pixels per year and ten vertical pixels per percentage point. The ten-year curve point is two hundred pixels right of the origin and one hundred sixty eight point one seven upward; the assigned fourteen point two level is one hundred forty two pixels upward. Their visible separation is twenty six point one seven pixels. These are calculations of the intended source curve; the original local D3 curve cannot draw when its library is unavailable. The source banner calls risk greater than ten threshold exceeded. Values from five up to ten are elevated precaution, while below five gets constrained safety margin. Those branches inspect the assigned number alone. At twenty pixels per percentage point five measures one hundred, ten two hundred and difference five one hundred. The critical horizon uses another algebraic rule rather than finding a crossing on the hazard curve, and the interval shading combines a sine term with elapsed-time growth rather than a statistical confidence estimate. Native listeners and exports are registered before the initial draw, so local JSON and Markdown can preserve entered settings despite missing D3. No actual researcher statement, confidence interval, governance effectiveness, emergency horizon or future catastrophe is independently verified.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.