Algorithmic versus human workplace monitoring: source and stated purpose alter behavior

Original study: “Algorithmic versus human surveillance leads to lower perceptions of autonomy and increased resistance,” Communications Psychology 2024; four studies (N=107 recalled personal experiences; N=157 and 117 Zoom experiment; N=814 hypothetical call-center vignette). A partial human comparator/design effect outside state coercion and without a genuine tool-using agent.

In the third experiment, participants doing the same idea-generation task saw one Zoom observer labelled “AI Technology Feed” or a human assistant; contact duration and feedback medium were held constant. Perceived autonomy averaged 3.91 versus 4.61 (d≈−.70), unsolicited criticism of oversight 11.7% versus 0%, and generated ideas 6.18 versus 7.74 (d≈−.43). The purported AI feed was an experimenter-controlled label, not a production AI actually processing their work: the result identifies response to a belief about who is watching, not an AI system’s detection accuracy or automated reporting. Study 2’s stronger difference was confounded by human observer not staying in the Zoom room and speaking rather than posting feedback; Study 3 addresses those differences.

In a 2×2 scenario experiment, N=814 imagined a call-center employer tracking either for evaluation or for employee development. Algorithmic evaluation reduced perceived autonomy relative to human evaluation (d≈−.30), whereas developmental framing removed that detectable source gap. Intention to quit exhibited an interaction, but the direct AI-vs-human comparison within the evaluative condition was only marginal (p=.072); saying it proved a higher observed quit rate would be false. The test varies declared purpose and measurement expectation, not a verified data-retention or employer-sanction policy. Neither trial measured protected political participation, disparate enforcement, multi-month effects or an agent pursuing leads.

Use as evidence that subject experience and design intentions matter when comparing humans and algorithms; do not transplant effect sizes into police or immigration systems. A genuinely more protective design would require actual purpose-bound access and constraints on onward disclosure, measured behavior rather than a friendly label.