Gradual Disempowerment at Scale: Measuring Cumulative Agency Erosion in Algorithmic Management of Indian Delivery Workers

Great presentation. Would be good to also add some context on addressing the privacy, security and anonymizatoin protocols required for safely extract the actual worker data

Can also expand on the discussion to explain exactly how the dimensions of interpretability, appeals, and lock-in compound mathematically or theoretically would strengthen the justification for the GDS weighting architecture

- interesting attempt to turn gradual disempowerment into measurable questions.

- underexplored lens for AI governance

- The main improvement needed is to define more clearly when ordinary platform coordination becomes disempowerment, and to validate the proxy score against worker-level evidence or platform disclosures.

Bridging AI safety's gradual disempowerment framing with labor platform auditing via a replicable metric is genuinely novel and worth developing further. The biggest weakness is D4 (predictive lock-in), which is entirely synthetic and carries the most interesting safety risk. Validating this dimension with even a small set of opt-in worker app traces would transform the paper's most important claim from an assumption into evidence.

This is a thoughtful and regionally grounded AI safety governance submission. The project applies the idea of gradual disempowerment to a concrete Global South setting: algorithmic management of Indian delivery workers on platforms such as Swiggy and Zomato. The proposed Gradual Disempowerment Score is a useful exploratory diagnostic that tries to connect AI safety concerns about loss of human agency with platform-labor evidence, using dimensions such as decision space, interpretability, appeals, predictive lock-in, and exit viability. The report is especially strong in its careful framing: it does not claim that disempowerment monotonically worsens, clearly labels the predictive lock-in component as synthetic, and treats the score as a diagnostic rather than proof of harm. The concise visual presentation and weight-sensitivity check also make the submission easy to evaluate.

That said, the project should be understood as an early operationalization rather than a mature or strongly validated eval. Algorithmic management, platform labor, and gig-worker precarity are already substantial research areas, and the AI safety connection here is broader and more governance-oriented than in submissions that directly evaluate model behavior or deployed AI systems. The main limitation is construct validity: GDS has not yet been validated against worker-reported agency, ability to contest decisions, health, income stability, or actual allocation outcomes.

The most AI-specific component, predictive lock-in, is synthetic because platform-level worker logs are unfortunately unavailable, and the weighting scheme remains subjective even with the included sensitivity analysis.

Overall, this is a promising exploratory submission with clear Global South relevance. A stronger next version would add worker interviews, platform-specific evidence, uncertainty intervals, broader weight sensitivity, and validation against real worker experiences or opt-in app traces.