FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteNo-Evidence version preferred74 / 162 · f24b9ac361230836

Towards Human-centered Explainable AI: A Survey of User Studies for Model Explanations

Psychology and cognitive neuroscience · 2210.11584v5

FLOWING EVIDENCE BENCHMARK

How do we tell whether Evidence helps?

At the same writing position, with the same model and task, how does supplying retrieved paper passages change the first output? We compare matched versions and retain ties, unusable outputs and incomplete reviews.

SAME DRAFT · EVIDENCE ON OR OFF

01 · FIXED WRITING POSITION

MANUSCRIPT

Same writing position ▌

Matched autocomplete at the same draft position

02 · TWO MATCHED INPUTS

SHARED BY BOTH

Manuscript context, model, task and prompt

A · With Evidence

Retrieved paper passages supplied

B · Without Evidence

No retrieved passages supplied

LLM

Same model and version

A → first output

B → first output

Blind judge agent

First outputs are anonymized as X and Y

Continuation review: accuracy, fit to the writing task and usability
Returns: X preferred / tie / Y preferred / both unusable

Order check: X / Y → Y / X

HOW DOES BLIND JUDGING WORK?

① Anonymize both outputs
The judge sees the same draft and both first outputs without knowing which received Evidence.

② Compare and swap order
The judge applies task-specific criteria in X/Y and then Y/X order.

③ Review disagreements
A third pass resolves disagreements. Incomplete reviews remain in the denominator.

How are autocomplete positions stratified?

Before seeing generation outcomes, we check whether a retrieved passage contains a specific proposition that directly supports the next writing move. Those positions appear in the left opportunity group; the rest are ordinary positions on the right. We select a balanced sample from admitted papers in each field. The 50/50 split is experimental, not a measure of how often either type occurs in writing.

AUTOCOMPLETE · 110

Biology, statistics and astrophysics

55 positions on each side; statistics uses the ten-paper rerun.

AUTOCOMPLETE · 52

Psychology and climate science

26 positions on each side; psychology includes nine papers and climate science four.

How are the table percentages calculated?

Across five fields there are 81 source-opportunity positions. The original blind review preferred Evidence in 51; a task check moved one empty Evidence continuation to both unusable, leaving 50 in public counts. In the ordinary group, another pair of empty outputs moved from no winner to both unusable. Original verdicts remain visible on case pages.

50Evidence version preferred
÷
81All positions in this group
=
62%Evidence preference in this group

Source contribution is a separate review: 16 of 21 Evidence wins entered into source review directly used retrieved papers; another 29 wins await review.

These are development-stage model judgments pending independent human review. They are not formal Benchmark conclusions and do not, on their own, establish causality.

Manuscript writing position

Text excerpt · not a PDF page

Research manuscript · excerpt

Towards Human-centered Explainable AI: A Survey of User Studies for Model Explanations

VII Future Research Directions > VII-B Open Research Problems > VII-B 2 Identifying and handling confounders

… in automatic evaluations, i.e., the explanations should be able to faithfully unbox the model. This verification step is essential to guarantee the validity of the empirical user study and to ensure that users are not tricked by unfaithful explanations. However, in most current human-subject experiments, the functional faithfulness of explanations is not thoroughly verified beforehand. Using unfaithful explanations could lead to the problem that only the placebo effect of explanations is measured. Ideally, a good explanation should be faithful to the model as well as understandable by users.

Existing research underscores the vulnerability of model explanation studies to significant confounding effects. For instance, Papenmeier et al. [ 156 ] reveal that user trust can be more influenced by model accuracy than the faithfulness of the explanation itself. Similarly, Yin et al. [ 193 ] demonstrate that the accuracy score perceived by users and the one shown to users contribute to trust formation.

↓ Next paragraph was generated here

The highlighted text is near the writing position. This is a layout of the evaluation input, not a PDF screenshot.

A matched comparison at the same writing positionJudgment: No-Evidence version preferred

Both arms use the same model, task and manuscript context. Only the retrieved passages differ. The first recorded output from each arm is shown below.

Evidence

Same model and position · retrieved passages supplied · first output

Not preferred in this comparison

These findings suggest that isolating the effect of explanation faithfulness from model accuracy remains a central methodological challenge, since users' trust judgments may reflect the underlying model performance rather than the explanation itself.

Passages supplied to the Evidence version

How model accuracy and explanation fidelity influence user trust ↗

The influence of explanation fidelity differs depending on the model accuracy: We see that for systems with medium accuracy, a high-fidelity explanation does not harm user trust, while a low-fidelity explanation does. Yet, for a system with high accuracy, any explanation (high-fidelity as well as lo…

Read full passage excerpt

The influence of explanation fidelity differs depending on the model accuracy: We see that for systems with medium accuracy, a high-fidelity explanation does not harm user trust, while a low-fidelity explanation does. Yet, for a system with high accuracy, any explanation (high-fidelity as well as low-fidelity) leads to a decrease in trust. We conclude that the interplay between explanation fidelity and user trust is more complex than pictured in literature to date. Furthermore, our findings show a discrepancy between how users act and what users report, which should be taken into account when evaluating user trust. With our research, we contribute empirical evidence of the relation between accuracy, fidelity, and user trust to the xAI community.

No Evidence

Same model and position · no retrieved passages · first output

Preferred by blind review

These findings indicate that factors unrelated to the explanation itself, such as perceived model accuracy, can dominate user responses and thus confound the measured effect of explanation quality. Future work should therefore systematically identify and control for such confounders, for example by manipulating model accuracy independently of explanation faithfulness.

Judgment and task check

The blinded model review preferred the no-Evidence version. Its original rationale appears below.

Read the original model judgment

In this rationale, Output A = Evidence; Output B = No Evidence.

Both outputs provide coherent, grounded continuations. Output A correctly notes the challenge of isolating explanation faithfulness from model accuracy, supported by the source's finding that explanation fidelity's influence differs by model accuracy level. Output B does the same but adds a concrete methodological recommendation (manipulating model accuracy independently) that is a natural, logical extension of the source's findings without inventing new empirical claims. The source explicitly discusses the 'interplay between explanation fidelity and user trust' and notes the 'discrepancy between how users act and what users report,' which supports B's call to 'systematically identify and control for such confounders.' B is more useful because it advances from diagnosis to actionable guidance while staying within the scope of the source's implications. Both avoid unsupported claims—neither invents new study results, and B's recommendation is a standard methodological response to confounding, not a factual claim about what was done. No citation markers are generated.

The passages are input material; their presence does not mean an output used them or that they support every claim. Draft and source passages are excerpted; no simulated scores or PDF appear here.