Trustworthy ML Regulation as a Principal-Agent Problem
Bibliographic record
Abstract
As ML-enabled systems become increasingly prevalent, so do their societal risks, such as excessive privacy and fairness violations.Mitigating such risks is frequently at odds with the system's overall performance.As a result, companies producing these systems and the public place different relative importance on these objectives-a case of misaligned incentives.Public regulators seeking to curb such risks are faced with designing regulations with a penalty structure that balances unmitigated societal risks and unnecessary financial burdens on a burgeoning industry.However, regulators often do not have access to the data or the training procedure of the model used by companies.Such access asymmetries can cause uncertainties in risk estimations that we show can lead to unintended over-regulation of law-abiding companies (1-7% drop in accuracy on three vision datasets).The access asymmetries that enabled unintended over-regulation are an artifact of the separation of the company and the regulator.This realization leads us to formulate ML Regulation as a Principal-Agent Problem (PAP).We provide a taxonomy of PAPs in this domain and tackle the problem of optimal penalty design in a reduced setting, which allows us to align the company's incentives with the public expectation of safety.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.022 | 0.053 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.003 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.002 | 0.007 |
| Scholarly communication | 0.008 | 0.007 |
| Open science | 0.003 | 0.005 |
| Research integrity | 0.005 | 0.008 |
| Insufficient payload (model declined to judge) | 0.005 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".