Remote Physical Detection of AI Agent Energy

Can an independent remote artificial intelligence agent reveal its electrical energy consumption through signals observable at a receiving network endpoint? This paper separates energy prediction, detection of a physical side channel, and quantitative measurement of a specified electrical measurand. A multidisciplinary review examines computing energy, network signal transformation, clock skew, thermal effects, dynamic voltage and frequency scaling, radiometric fingerprints, metrology, and relevant patent claims. Existing experiments establish that some computation dependent information can survive as network timing or encoded clock behavior. They do not establish an ordinary website as a calibrated meter for the total energy of an arbitrary distributed agent. A conditional non-identifiability result is developed: when admissible systems have the same distribution of all permitted observations but different energy consumption, no estimator can recover both values, regardless of receiver precision. A corresponding squared-error lower bound makes this limitation quantitative. More restricted measurement pathways remain possible if a causal channel, source boundary, calibration, and influence quantities are specified. Eleven proposed experiments, including eight principal contrasts and three discriminating controls, test preservation, causal specificity, quantitative recovery, and transfer across hardware and paths. The defensible verdict is non-identifiability under the unrestricted ordinary-website observation model, with conditional physical leakage and cooperative or additional-channel measurement remaining distinct research directions.