Exploring Subjective Notions of Explainability through Counterfactual Visualization of Sentiment Analysis
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
The generation and presentation of counterfactual explanations (CFEs) are a commonly used, model-agnostic, approach to helping end-users reason about the validity of AI/ML model outputs. By demonstrating how sensitive the model's outputs are to minor variations, CFEs are thought to improve understanding of the model's behavior, identify potential biases, and increase the transparency of ‘black box models’. Here, we examine how CFEs support a diverse audience, both with and without technical expertise, to understand the results of an LLM-informed sentiment analysis. We conducted a preliminary pilot study with ten individuals with varied expertise from ranging NLP, ML, and ethics, to specific domains. All individuals were actively using or working with AI/ML technology as part of their daily jobs. Through semi-structured interviews grounded in a set of concrete examples, we examined how CFEs influence participants' perceptions of the model's correctness, fairness, and trust- worthiness, and how visualization of CFEs specifically influences those perceptions. We also surface how participants wrestle with their internal definitions of ‘explainability’, relative to what CFEs present, their cultures, and backgrounds, in addition to the, much more widely studied phenomena, of comparing their baseline expectations of the model's performance. Compared to prior research, our findings highlight the sociotechnical frictions that CFEs surface but do not necessarily remedy. We conclude with the design implications of developing transparent AI/ML visualization systems for more general tasks.
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.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
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 it