# Researchers Show Facility Power Meters Can Detect Hidden AI Training Runs Without Trusting the Operator

On simulated data center power, a Viterbi tracker caught at least 96% of undisguised training runs at a 1% false-alarm rate, and at least 87% of runs over a day even when the operator tried to hide them.

- Published: 2026-10-23T06:19:29.884Z
- Canonical: https://polylog.news/ai/2026-10-23/researchers-show-facility-power-meters-can-detect-hidden-ai
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv cs.CR](https://arxiv.org/abs/2610.30182), [Polylog editors](https://polylog.news), [Tom's Hardware](https://www.tomshardware.com/tech-industry/data-centers/us-senate-kills-bill-that-could-potentially-shield-americans-from-skyrocketing-power-bills-due-to-ai-data-centers-opponents-say-bill-is-toothless-and-doesnt-do-enough-to-protect-citizens)

A new paper on arXiv, "Detecting AI Training in Datacentre Power" by Tom Kimpson, Mauricio Baker and Emlyn Graham, addresses a verification problem in proposals for AI governance. A monitor has to tell whether a cluster is running a large t…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-10-23/researchers-show-facility-power-meters-can-detect-hidden-ai (subscription information: https://polylog.news/pricing).