Refugee status determination: how cooperation with machine learning tools can lead to more justice
Bibliographic record
Abstract
Previous research on refugee status adjudications has shown that prediction of the outcome of an application can be derived from very few features with satisfactory accuracy. Recent research work has achieved between 70 and 90% accuracy using text analytics on various legal fields among which refugee status determination. Some studies report predictions derived from the judge identity only. Additionally most features used for prediction are non-substantive and external features ranging from news reports, date and time of the hearing or weather. On the other hand, literature shows that noise is ubiquitous in human judgments and significantly affects the outcome of decisions. It has been demonstrated that noise is a significant factor impacting legal decisions. We use the term "noise" in the sense described by D. Kahneman, as a measure of how human beings are unavoidably influenced by external factors when making a decision. In the context of refugee status determination, it means for instance that two judges would take different decisions when presented with the same application. This article explores ways that machine learning can help reduce noise in refugee law decision making. We are not suggesting that this proposed methodology should be exclusive from other approaches to improve decisions such as training of decision makers, skills acquisition or judgment aggregation, but rather that it is a path worth exploring. We investigate how artificial intelligence and specifically data-driven applications can be used to benefit all parties involved in refugee status adjudications. We specifically look at decisions taken in Canada and in the United States. Our research aims at reducing arbitrariness and unfairness that derive from noisy decisions, based on the assumption that if two cases or applications are alike they should be treated in the same way and induce the same outcome.
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 | 0.025 | 0.081 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.003 | 0.002 |
| Science and technology studies | 0.003 | 0.005 |
| Scholarly communication | 0.008 | 0.010 |
| Open science | 0.002 | 0.008 |
| Research integrity | 0.003 | 0.004 |
| Insufficient payload (model declined to judge) | 0.007 | 0.003 |
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".