CHI 2026 · Human–AI interaction
Rewrite the rules behind an AI decision.
CoExplain lets people express their knowledge as decision rules, then work with AI to refine the thresholds and structure. The explanation becomes a way to shape the model—and understand how it changes.
01 / Guided interface tour
Follow the house-price example to see what people can change, what AI assistance adds, and how the two are compared.
01 / Write
The user’s tree starts with grade > 9, then checks living area. Following the upper path, a house with living area > 2000 receives a high-price prediction.
These are editable rules on familiar attributes. CoExplain parses the written tree into a neural network, so the rules guide the model’s predictions.
02 / Enhance
Enhance Values refines numeric thresholds while keeping the tree’s structure. Enhance Flowchart can also reorganize the rules, adding or rearranging decision nodes.
Someone confident in the logic can ask for narrower adjustments; someone reconsidering the logic can allow structural changes.
03 / Constrain
Prediction Similarity encourages the enhanced model to preserve the current tree’s predictions. Structure Similarity encourages it to retain the current arrangement of rules.
The edit history records updates, insertions, and removals. People can inspect what changed before accepting a proposal.
04 / Compare
The proposed tree begins with living area, while the original begins with grade. Side-by-side rules and accuracy indicators make that change visible.
Users can accept the proposal or selectively incorporate parts they agree with. A separate simulation tests the rules on a test dataset.
A guided tour of the paper’s interface (§5), with the original figure enlarged at each step. No model is trained in this page.
02 / How it works
CoExplain connects symbolic decision rules and a trainable neural network. Each interaction has a corresponding model operation.
A decision tree learns to reproduce the neural network’s predictions, giving people an explanation in terms of attributes and thresholds.
The authored tree is mapped into a neural network: decision tests, paths, and outcomes become connected computations.
Training improves fit to data while regularization encourages similar predictions and tree structure. The updated model is distilled into a tree for inspection.
A distilled tree approximates the network. The paper separately evaluates explanation–model faithfulness (§6.3.7); readability alone does not establish that the two behave identically.
03 / Study findings
43 participants were assigned to Read-only explanations, manual editing, or CoExplain. Each worked on income, house-price, and heart-disease prediction tasks, using supplied guideline rules as background knowledge.
| Task | Read-only | Editable | CoExplain |
|---|---|---|---|
| Income | 77.1% | 87.5% | 91.9% |
| House price | 61.9% | 89.6% | 87.6% |
| Heart disease | 74.4% | 92.8% | 93.9% |
Mean user–AI faithfulness with explanations visible; higher means better anticipation. Both editable conditions significantly outperformed Read-only overall; they did not differ significantly from each other. Source: §6.3.2, Figure 7.
Less effort with assistance
Compared with manual editing, CoExplain reduced the number of operations, revision rounds, and time spent editing.
Group means · §6.3.6, Figure 9| Measure | Manual editing | CoExplain |
|---|---|---|
| Editing timeminutes | 7.18 | 3.36 |
| Edit operations | 14.05 | 9.00 |
The tradeoff
Manual editing produced rules closest to the supplied guidelines, but lowered accuracy on the model’s original distribution. CoExplain recovered part of that accuracy while retaining closer rule alignment than Read-only.
| Condition | Accuracy ↑ | Rule distance ↓ |
|---|---|---|
| Read-only | 78.9% | 13.74 |
| Editable | 59.8% | 3.13 |
| CoExplain | 69.9% | 6.64 |
Rule distance is Tree Edit Distance to the supplied guideline tree; lower means closer structure. These tasks deliberately used different guideline and pre-trained data distributions. Source: §6.3.3–6.3.5, Figure 8.
Scope of the evidence. A controlled study with university participants across three prediction tasks. The results concern understanding, alignment to supplied guidelines, and editing effort; they do not establish clinical or other real-world deployment outcomes.
Paper & resources
CHI ’26 · ACM Conference on Human Factors in Computing Systems · Barcelona, Spain
@inproceedings{chen2026editablexai,
author = {Chen, Haoyang and Bai, Jingwen and Fang, Tian and Lim, Brian Y.},
title = {Editable XAI: Toward Bidirectional Human-AI Alignment with
Co-Editable Explanations of Interpretable Attributes},
booktitle = {Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems},
year = {2026},
publisher = {Association for Computing Machinery},
doi = {10.1145/3772318.3791375},
url = {https://doi.org/10.1145/3772318.3791375}
}