Summary

  • Researchers used aggregated traffic from one French mobile operator to model commune-level vote shares in the 2019 and 2024 European Parliament elections. Adding service-demand measures to a socioeconomic baseline raised mean model correlation by 28.23% in the 2024 analysis.
  • The study's public code starts after the operator has classified and transformed the traffic. The raw network feed and the production classifier remain private. That makes the institution holding the first lens as important as the institution reading the final chart.

The polling booth stayed locked. The network kept taking notes.

A ballot box is a wonderfully stubborn machine. It will not tell a phone company how a customer voted. The phone company can still see a great deal about the neighbourhood around that customer. Its network already measures traffic by application, time and antenna. A research team has now shown how those ordinary engineering observations can be assembled into a political map without opening a ballot.

At RIPE 93, Orlando Eduardo Martínez-Durive presented the result: a study of two French European Parliament elections that linked mobile-service demand to vote shares at the commune level. The original paper deposited by IMDEA Networks supplies the machinery behind the striking figure. In 2024, adding mobile-service indicators to the researchers' socioeconomic model raised its mean correlation with party vote shares by 28.23% relative to that baseline. That is a relative improvement in this model across parties, not 28.23 percentage points, and certainly not a score for guessing any voter's choice. The unit on the chart is a commune.

The operator was unnamed. Its two brands together represented 31.9% of the French mobile market in 2019 and 31.5% in 2024. The campaigns gathered minute-by-minute service-demand byte counts from roughly 17,600 base stations in 2019 and 25,900 in 2024. The classifiers tracked 22 services in the first election and 27 in the second. The researchers then translated antenna geography into municipal geography, using overlapping coverage areas and a uniform-distribution assumption inside each constructed cell. Dense urban networks make that interpolation more usable. Sparse rural networks do not.

The final analysis therefore covered 4,102 urban and suburban communes, roughly 66% of eligible voters and more than 70% of measured mobile demand.

This is an awkwardly good instrument. It does not need a social platform's API or a voluntary poll respondent. The operator has the equipment in the ground and the classifier in the network. It can measure across services and across time. The paper's examples associate greater Facebook or TikTok consumption with support for particular parties and other consumption patterns with other parties; the point for governance is the instrument, not a fresh set of political stereotypes. An association at commune level stays at commune level. A strong fit can coexist with a bad story about why people behave as they do.

The authors' ethics appendix says the operator processed the traffic on its own premises into anonymous antenna-level aggregates, mixing tens to hundreds of users and removing personal identifiers. The operator's data-protection officer approved collection and processing for this collaborative research project. Researchers received aggregates. These are concrete controls, and the study deserves to be read with them intact.

The next question begins where the approved project ends: if operators become a standing source for electoral observation, who specifies the purpose, validates the classifier and decides what level of geographic or service detail may leave the network?

The authors' public reproduction package makes part of the answer inspectable. It releases commune-level transformed tables and analysis code. It also marks the first missing step. Raw operator traffic is under a nondisclosure agreement; the production deep-packet-inspection classifiers, their ground truth and their per-service accuracy are not public. A reader can reproduce much of the modelling downstream of the shipped tables. A reader cannot independently rerun the proprietary conversion from network flow to the service labels that enter those tables. If the instrument mislabels a service, the error enters the political model before the public audit begins.

There is a second, more prosaic limit. The study is observational, and the authors say so. They cannot tell whether a platform changes political views, political views change platform use, or some third feature of a town shapes both. The 2019 and 2024 repetitions show a durable association under their definitions, not a causal experiment. Their town-level result says nothing about whether one named voter used one named service. The paper also excludes rural France and has no second operator against which to test subscriber mix or classifier behaviour.

Those limits do not make the instrument dull. They make its owner important. A network built to carry private communications has acquired a municipal telescope. Researchers can use it to study the public sphere; regulators may want a view that no platform can withdraw. The rights to collect, classify, aggregate, retain, compare and publish that view sit at different desks. The DPO's approval answers a defined research project. A standing civic dashboard would need its own dated instructions, independent checks and a line the operator cannot quietly move when a new service becomes politically interesting.

A lens can be independent of platforms and dependent on a phone company

The attraction of network measurement is clear: platform companies can ration an API, change a definition or close the door. An operator's traffic vantage survives those product decisions. Yet swapping a platform gatekeeper for a network gatekeeper only moves the hinge. The classified service list, the interpolation method and the release threshold become governance decisions. A chart cannot explain its own chain of custody. The ballot box remains sealed; the measurement room needs a window.