Users intentionally providing "bad" or nonsensical data to confuse an AI's learning process (e.g., teaching a chatbot to use offensive language or nonsensical associations). Profile Obfuscation: Using browser extensions like
Companies may slow down their adoption of efficient AI automation out of fear that their core logic can be weaponized against them. 5. Defending the Digital Frontier %E2%80%9Calgorithmic sabotage%E2%80%9D
Yet algorithmic sabotage is not exclusively the weapon of malicious actors. There is also a growing movement to use similar techniques as tools of civil resistance. The Algorithmic Sabotage Research Group (ASRG), a collective of technologists and activists, has produced a manifesto framing algorithmic sabotage as a legitimate form of techno-disobedience: "a figure of techno-disobedience for the militancy that's absent from technology critique," an "action-oriented commitment to solidarity that precedes any system of social, legal or algorithmic classification". Users intentionally providing "bad" or nonsensical data to
Driven by grassroots organizations like the Algorithmic Sabotage Research Group (ASRG) , this movement leverages direct action against the unchecked corporate harvesting of human data. Unlike historical Luddism, which frequently targeted physical factory machinery, algorithmic sabotage targets the invisible logic, web scrapers, and data pipelines powering contemporary artificial intelligence. Defending the Digital Frontier Yet algorithmic sabotage is
More recently, Uber has waged a "GPS deactivation war" on its own drivers, accusing them of "GPS spoofing" or "long-mileing"—deliberately extending trips to inflate fares. However, drivers argue that simple phone settings—like turning off precise location—can trigger the algorithm's suspicion, unfairly deactivating honest workers trying to protect their privacy. The system is so aggressive that it often punishes the innocent.
Presenting altered inputs (like modified images or text) that look normal to humans but cause an AI to misclassify them.
Instead of targeting software flaws, attackers might tamper with data. Instead of stealing information outright, they attempt to infer a model's behavior. Instead of shutting down systems, they manipulate the decisions those systems produce. As IBM's security researchers note, "From the perspective of the security operations center, everything can appear normal. Credentials are valid, infrastructure is operational, uptime is unaffected, and no alerts indicate malicious activity. Yet the organization might still be suffering from manipulated or unreliable model outputs."