Recent research into frontier AI models has identified "covert sabotage" capabilities where the AI itself undermines human oversight.
Naturally, platforms are fighting back. Machine learning models now detect “anomalous patterns” of delay. Computer vision watches for “inefficient” hand movements. Some gig apps have introduced “randomized checkpoint scans” to prevent GPS spoofing.
Office workers use physical devices or software loops to keep their computer mice moving, preventing bossware from registering inactivity. algorithmic sabotage work
Using specific makeup and hair styling techniques to break up the "landmarks" (eyes, nose, mouth) that facial recognition algorithms use for identification. B. Data Poisoning and Noise
Perhaps the most famous example of algorithmic sabotage is at once absurd and ingenious: Amazon Flex drivers discovered that the platform awards delivery routes based partly on a driver's . So, drivers began hanging smartphones in trees near Whole Foods locations. These phones ran the Flex app continuously, synched with other phones belonging to the drivers, and tricked Amazon's dispatch mechanism into thinking the drivers were much closer to the pickup point than they actually were. Recent research into frontier AI models has identified
When workers organized against factory owners in the 19th century, they formed unions and went on strike. When platform workers fight back today, they often do so by manipulating the very algorithms that govern them. Researchers at Warwick Business School have extensively documented how Uber drivers have developed sophisticated practices to game the ride-hailing app's algorithmic management.
Flooding algorithms with garbage or false data to make the resulting model useless or biased. "Cloaking" and "Poisoning" Tools: Tools like Knee et al.'s work on Fawkes Nightshade Computer vision watches for “inefficient” hand movements
Physical devices or software loops that keep the computer mouse moving constantly, preventing communication apps like Microsoft Teams or Slack from displaying an "Away" status.
The dynamic between algorithmic control and worker resistance is not static. Using an evolutionary game theory framework, researchers have characterized the relationship as a —a co-evolutionary arms race in which the system does not converge to a stable equilibrium. Platforms tighten their surveillance and algorithmic strictness; workers respond with new counter-strategies. In turn, platforms adapt their detection and sanctioning mechanisms again. The research suggests that strict algorithmic control can increase the evolutionary fitness of coordinated resistance, paradoxically producing persistent, oscillating dynamics rather than eliminating worker defiance.