Personal risk observatory / beta

Existential AI

A methodologically adventurous estimate of the likelihood that I personally cause an existential problem through artificial intelligence.

CEILE v0.3.1

Incident likelihood estimator

Model operational

Estimated annual likelihood

0.03%

Negligible-ish. Continue to behave normally.

Confidence: Undeserved
Peer review: Pending indefinitely

Model capability 4

AutocompleteQuestionable agency

Operational autonomy 2

Asks permissionHas objectives

Infrastructure access 1

Local sandboxGlobal root

Researcher sleep deficit 3

Eight hoursConference deadline

Unchecked philosophy 6

One footnoteOntology deployed

Procedural safeguards Adult supervision present Deploying late on Friday

The model

A semi-serious formula

The estimator uses a logistic function, which is a respectable way to prevent a made-up score from wandering below 0% or above 100%. Its inputs loosely resemble factors used in real technology risk discussions.

P(incident) = 100 × σ(−8.4 + 0.032C² + 0.041A² + 0.052I² + 0.08S + 0.07Φ + 1.25F − 1.10O)

C: Capability
A: Autonomy
I: Infrastructure access
S: Sleep deficit
Φ: Philosophical overreach
F: Friday deployment
O: Oversight

Disclaimers and epistemic hygiene

What this borrows from reality

Capability, autonomy, access, human reliability, and oversight are reasonable categories to examine when assessing the safety of an AI-enabled system. Greater power and reach can increase the consequences of failure; meaningful review can reduce them.

What this does not establish

The output is not a forecast, validated instrument, actuarial estimate, threat assessment, or covert admission. No empirical dataset connects these slider positions to existential outcomes.

Known methodological limitations

The model omits institutional incentives, recursive improvement, geopolitical competition, hardware constraints, and whether I have eaten today. The confidence interval is therefore best expressed as “somewhere between absolutely not and please reconsider.”