Does the whole exceed its parts? Effect of AI Explanations on Complementary Team Performance

Between-subjects crowdsourcing experiments on three datasets, with around 100 participants per condition, preceded by a pilot on one dataset.

The defining design move. The authors first measured unassisted human accuracy on each task in separate pilot runs, then selected the study items so that the AI's accuracy matched it. For the two sentiment datasets they observed human accuracy of 87% (Beer) and 85% (Amzbook) and then chose 50 unambiguous test-set examples on which the AI scored 84%, with equal false positive and false negative rates. For LSAT the model scored 65% on the selected 20 items against 67% human accuracy. Matching the two parties is what creates a wide complementary zone; it is the methodological contribution that separates this study from the eight it critiques.

Tasks. Two sentiment classification datasets, beer reviews and Amazon book reviews, both binarised to positive or negative, and a set of Law School Admission Test logical reasoning questions with four options and one correct answer. Sentiment items took around 17 seconds each, LSAT items around a minute.

AI models. RoBERTa text classifiers fine-tuned per sentiment dataset, with confidence scores recalibrated by isotonic regression because the raw neural scores were poorly calibrated. For LSAT, a RoBERTa model fine-tuned on ReClor, a logic-reasoning dataset drawn from LSAT and GMAT questions.

Explanation strategies. Explain-Top-1 highlights the most influential words for the predicted class only. Explain-Top-2 also colour-codes and highlights words supporting the alternative class. Adaptive, the authors' own proposal, switches between them on the AI's confidence: it explains only the top prediction when confident and adds the counter-argument when below a threshold set at the classifier's median confidence, 89.2% for Beer and 88.9% for Amzbook. This yielded 25 items shown each way, with 18 correct and 7 incorrect predictions falling in the Explain-Top-2 bucket.

Explanation sources. AI-generated explanations came from LIME, selected because the RoBERTa-plus-LIME combination ranked highest in a separate explainer comparison with human judges. Expert explanations, written by one author, served as an upper bound on quality: a short convincing phrase span per class for sentiment, and for LSAT condensed prep-book rationales for the correct answer plus manually crafted supporting arguments for the other three options. Experts wrote only explanations; they did not alter predictions or confidences.

Conditions. Two baselines in every domain: unassisted Human, and Team (Conf), which shows the AI's recommendation and calibrated confidence but no explanation. The authors chose the confidence baseline deliberately as a stronger and more honest comparison than recommendation-without-confidence. Sentiment then added Team (Explain-Top-1, AI), Team (Explain-Top-2, AI), Team (Adaptive, AI) and Team (Adaptive, Expert), six conditions in all. LSAT used the expert source throughout, five conditions.

Interface. Explanations appear as inline highlights in the review text, coloured to match the positive and negative buttons, so that the user does not move between input and explanation. The AI, personified as "Marvin", hovers above its recommended button with the confidence score beneath. For LSAT all AI information sits in a separate right-hand pane, explicitly to discourage blind following. Crucially, in all conditions the AI's recommendation and explanation are visible before the participant answers.

Procedure. Payment scheme and task explained, then a screening phase requiring four of six correct labels (two of two for LSAT, a stricter gate against cheating). Those who passed were randomly assigned to a condition, given condition-specific instructions on the meaning of prediction, confidence and explanation, then labelled all 50 items (20 for LSAT) one by one, in a fixed order shared across conditions, with immediate correctness feedback after each item. A post-task survey asked whether the AI assistance and the explanations in particular had helped, and how the participant had used the assistance.

Participants. Amazon Mechanical Turk, United States only, 97% approval and 1,000 approved tasks minimum. Data were dropped from participants whose median labelling time fell below 2 seconds (3 for LSAT) or who gave the same label throughout. 566 recruited for Beer and 552 for Amzbook, of whom 84% passed screening and filtering; 508 for LSAT, of whom 35% did. Final cells ranged from 93 to 101 participants.

Incentives. Base pay $0.50, $0.25 for the survey, and a bonus scheme combining a linear and a step component: $0.05 per correct decision plus $0.50 for exceeding 90% accuracy or $1.00 for exceeding 95%. Because the AI was fixed at 84%, blind following could not earn the step bonus. LSAT paid $0.30 per correct answer with step bonuses at 30%, 50% and 85%. Average payment $3.35 (sentiment) and $6.30 (LSAT), roughly $16 and $23 per hour.

Analysis. Student's t-tests with Bonferroni correction on team accuracy, an analysis split by whether the AI's recommendation was correct, human-AI agreement rates split at the adaptive confidence threshold, and a qualitative coding of 409 free-text responses on collaboration strategy by two annotators (mean Cohen's kappa 0.71, mean raw agreement 93%).

Every human-AI team beat both the human alone and the AI alone. No explanation condition beat simply showing the AI's confidence.

Complementarity was achieved by the plain baseline. For Beer, Team (Conf) reached 0.89 (SD 0.05) against 0.84 for the AI and 0.82 (SD 0.09) for unassisted humans. Amzbook and LSAT likewise landed in the complementary zone, with relative gains over unassisted workers of 2.2% and 20.1%. This is the first result: matched accuracy plus a confidence display is enough to make the team worth forming.

Explanations added nothing on top. Explain-Top-1 against Team (Conf) on Beer: 0.88 (SD 0.06) against 0.89 (SD 0.05), z = -1.18, p = .24; the same null on Amzbook (z = 1.23, p = .22) and LSAT (z = 0.427, p = .64). H2 could not be rejected. Explain-Top-1 did not beat Explain-Top-2 either, so H1 was unsupported (Beer z = 0.85, p = .40; Amzbook z = 0.81, p = .42; LSAT z = 0.42, p = .68).

The authors' own Adaptive strategy also failed. No significant difference against Team (Conf) for Beer (z = -1.02, p = .31), Amzbook (z = 1.08, p = .28) or LSAT (z = 0.16, p = .87). Nor did switching to expert-written explanations help: Adaptive Expert against Adaptive AI gave z = 1.31, p = .19 on Beer and z = -0.78, p = .43 on Amzbook. H3 and H4 both unsupported. The negative result therefore covers not only off-the-shelf saliency but a purpose-built confidence-adaptive strategy and a human-written quality ceiling.

The reason the net effect is zero is the mechanism that matters for the framework. Splitting by whether the AI was right, explanations raised accuracy when the recommendation was correct and lowered it when the recommendation was wrong. Explanations increase acceptance of the AI's answer irrespective of its correctness, so the gains and losses cancel. The authors note this is the first large-scale demonstration of the effect on end-to-end decision accuracy rather than on subjective trust ratings.

Adaptive explanations did move behaviour, just not enough to move accuracy. Agreement rates split at the confidence threshold show Adaptive reduced agreement with the AI on low-confidence items and preserved it on high-confidence ones, which is the desired direction. The authors give two reasons it did not convert into accuracy: on high-confidence items participants would have had to detect confident mistakes, which are unknown unknowns, and on low-confidence items participants had to both disagree and then get the answer right themselves.

Comparable accuracy does not guarantee complementary partners. Mapping items by average unassisted human accuracy showed Beer and LSAT had many items the AI got right and humans found hard, while Amzbook did not: there, AI errors correlated with human errors. Amzbook accordingly gained least from assistance, and its participants reported ignoring the AI most (30%, against 23% for Beer and 11% for LSAT). What matters is error independence, not equal accuracy.

Behaviourally, most participants neither followed nor ignored the AI. In the coded free-text responses, 47% used the AI as a prior guide, 25% as a post-check, 22% mostly ignored it and only 6% mostly followed it. The prior-guide pattern was more common in sentiment analysis and the post-check pattern in LSAT, matching the interface, where LSAT put the AI in a separate pane. Confidence was the salient cue: of the 90 participants who mentioned using it, 27% reported an explicit personal threshold below which they distrusted the AI, mostly between 80 and 100 (83 ± 8 Beer, 89 ± 7 Amzbook, 90 ± 0 LSAT). The same proportion of participants reported using confidence whether or not they also saw explanations.