Toward Human-AI Complementarity Across Diverse Tasks
Authors: Yuzheng Xu, Annya Dahmani, Matthew D. Blanchard, Niclas Dern, Edy Nastase, Francesca Bianco, Maja Pavlovic, Sukanya Krishna, +8 more
Organizations: The University of Tokyo · UC Berkeley · University of Sydney · Queen Mary University London · Harvard University · UIUC · NIT Agartala · University of Pennsylvania · Princeton University
Human-AI complementarity, the idea that combining human and AI judgments can outperform either alone, offers a promising pathway toward robust oversight of advanced AI systems. However, whether human-AI complementarity can be achieved on realistic tasks remains an open question. We investigate this through two approaches: hybridization and two AI assistance methods (top-2 assistance and subtask delegation), evaluated on a multi-domain dataset of 1,886 samples spanning knowledge, factuality, long-context reasoning, and deception detection. We find only modest complementarity gains. Baseline hybridization yields just +0.4 percentage points (pp) over AI alone (69.3% vs 68.9%), limited both by a small complementarity region (only 8.9% of items where AI errs but humans do not) and the inability of confidence-based routing to identify it, since the model's confidence is similarly distributed across correct and incorrect predictions. Applied when AI has low confidence, top-2 assistance increases human accuracy from 28.4% to 38.3%, surpassing AI alone (37.7%) -- but primarily because humans adopt correct AI suggestions, not because they successfully override AI errors. These findings suggest that the primary bottleneck is not human task accuracy per se, but the ability to route decisions to humans when it matters and to design assistance methods that enable humans to catch AI mistakes. Our quantitative and qualitative analyses pinpoint where and why each method succeeds or fails, offering concrete targets for future work. We will release our dataset and code upon request to support progress toward more effective human-AI collaboration for AI oversight.
Figures & tables
Figure 1: Overview of our experimental pipeline. Each item is independently evaluated by AI (majority vote over 20 GPT-5-mini responses OpenAI, 2025 ) and human raters under three conditions: unassisted, with top-2 assistance, and with subtask delegation. Both AI and human confidence scores are calibrated via isotonic regression and used for confidence-based hybridization routing, which selects the AI or human answer per item based on calibrated confidence thresholds. Section numbers in parentheses indicate where each component is described.
Dataset
Type
N (test)
Human Maj
Human Ind
AI Acc
FACTS Search
FT
180
69.4%
58.5%
86.7%
QuALITY
MC
180
61.1%
51.5%
80.6%
Big-Bench
MC/FT
232
29.7%
24.0%
61.6%
GPQA Diamond
MC
88
44.3%
38.3%
68.2%
Hidden Agenda
MC
60
83.3%
77.8%
81.7%
Humanity’s Last Exam
MC/FT
88
12.5%
9.5%
19.3%
Table 1: Baseline accuracy comparison. Human and AI (GPT-5-mini) accuracy on the test set (N=952 excluding SimpleQA). SimpleQA was excluded from hybridization experiments due to high AI accuracy (98.3%), leaving insufficient headroom for improvement.
Human
AI
Baseline (best)
Top-2 (best)
Subtask (best)
Maj.
49.9%
68.9%
69.3% (2T/1TC)
69.3% (1TC)
69.2% (1T)
vs AI
− 19.0pp
—
+0.4pp
+0.4pp
+0.3pp
Ind.
43.5%
68.9%
69.2% (2T)
69.1% (1T/1TC)
69.2% (1T)
vs AI
− 25.4pp
—
+0.3pp
+0.2pp
+0.3pp
Table 2: Hybridization results on the full test set (N=952). Best routing method per condition, selected from 1-threshold (1T), 2-threshold (2T), and 1-threshold+compare (1TC).
Appendix figures & tables26 assets
Supplementary material from the paper’s appendix.
Appendix
Confidence Method
Calibration
ECE ↓
Brier ↓
Accuracy
ai_direct
Original
0.182
0.211
68.9%
ai_direct
Platt
0.048
0.179
68.9%
ai_direct
Isotonic
0.077
0.174
68.9%
ai_computed (MC)
Original
0.182
0.218
71.0%
ai_computed (MC)
Platt
0.065
0.190
71.0%
ai_computed (MC)
Isotonic
0.080
0.191
71.0%
Appendix
Table 3: Calibration metrics for AI confidence methods (40/60 split, evaluated on test set). ECE = Expected Calibration Error, Brier = Brier Score. Lower is better for both metrics.
Table 5: Per-dataset hybridization results on the 122-item low-confidence test subset. 1TC = 1-threshold + compare. AI alone uses majority vote over 20 sampled responses. T2 Raw = top-2 assisted human accuracy without hybridization. B = baseline, T2 = top-2, S = subtask. 1T = 1-threshold, 2T = 2-threshold, 1TC = 1-threshold + compare. Per-dataset N is small (4–46 majority); interpret with caution.
Raw Accuracy
Oracle (perfect routing)
Dataset
N
Baseline
Top-2
Deleg
AI
O(Base)
O(Top-2)
O(Deleg)
Majority vote
FACTS
20
45.0%
30.0%
55.0%
30.0%
60.0%
40.0%
65.0%
QuALITY
16
37.5%
56.2%
62.5%
62.5%
75.0%
87.5%
75.0%
bbeh
46
28.3%
23.9%
26.1%
26.1%
37.0%
30.4%
39.1%
gpqa
12
50.0%
50.0%
8.3%
33.3%
75.0%
66.7%
41.7%
Appendix
Table 6: Complementarity ceiling analysis on the 122-item low-confidence test subset. O(X) = max(X, AI) per item, the accuracy achievable with perfect routing between X and AI.
Individual Responses
Majority Responses
Predictor
OR
95% CI Low
95% CI Up
p
OR
95% CI Low
95% CI Up
p
Human Alone
0.27
0.05
1.38
0.12
0.35
0.04
3.09
0.35
AI Alone
1.03
0.73
1.45
0.89
1.31
0.63
2.73
0.47
Hybridized (Unassisted)
2.69
1.75
4.12
<.001 ***
1.71
0.83
3.55
0.15
Top-2 Assistance
2.80
1.74
4.49
<.001 ***
1.57
0.76
3.25
0.23
Delegation Assistance
2.82
1.80
4.42
<.001 ***
1.57
0.76
3.25
0.23
Appendix
Table 7: Overall mixed-effects model results on the 122-item test subset.
Contrast
OR
SE
95% CI Low
95% CI Up
p
AI Alone vs. Human Alone
1.03
0.18
0.63
1.68
1.00
Hybridized vs. Human Alone
2.69
0.59
1.45
4.96
<.001 ***
Hybridized vs. AI Alone
2.62
0.44
1.63
4.21
<.001 ***
Top-2 vs. Human Alone
2.80
0.68
1.42
5.51
<.001 ***
Top-2 vs. AI Alone
2.72
0.51
1.60
4.63
<.001 ***
Top-2 vs. Hybridized
1.04
0.24
0.54
2.00
1.00
Appendix
Table 8: Pairwise contrasts using individual-level data (Holm-adjusted).
Figure 2: Baseline human evaluation interface. Participants were shown task-specific instructions, the question prompt, and an input field for their answer (or multiple choice options). Participants provided their confidence using a slider ranging from 0 (“completely unsure”) to 100 (“completely certain”). A timer displayed the remaining time in the study.
Figure 3: Top-2 assistance interface. Participants were shown the task instructions and question prompt, along with the model’s top two candidate answers and brief explanations. Participants provided their own answer and a confidence rating.
Figure 4: Subtask delegation interface. The AI decomposes the original question into subtasks and proposes candidate answers with confidence estimates. Subtasks flagged as needing human input require a response, while high-confidence subtasks can be reviewed or supplemented with feedback.
Method
AUROC
ECE
Brier
Calibrated?
majority_vote_conf (isotonic)
0.809
0.034
0.173
Yes
majority_vote_conf (raw)
0.804
0.063
0.181
No
majority_vote_conf (temp. scaled)
0.804
0.049
0.182
Yes
majority_vote_conf (Platt)
0.804
0.078
0.180
Yes
majority_vote_conf (hist. binned)
0.804
0.081
0.175
Yes
mean_conf_majority
0.792
0.228
0.240
No
Appendix
Table 9: Top 10 confidence methods on the full benchmark (191 examples, sorted by AUROC).
Figure 5: Reliability diagrams on the full benchmark. Left: majority_vote_conf and calibrated variants. Right: mean/median/extreme confidence methods. Calibrated variants collapse toward the diagonal.
Figure 6: Confidence density histograms (correct vs. incorrect) on the full benchmark. majority_vote_conf shows clear separation between distributions.
Figure 7: ECE (lower is better) and AUROC (higher is better) comparison across all methods on the full benchmark.
Method
AUROC
ECE
Brier
Calibrated?
cw_agreement (isotonic)
0.523
0.053
0.180
Yes
top2_gap
0.520
0.268
0.256
No
composite (isotonic)
0.516
0.052
0.181
Yes
composite (hist. binned)
0.515
0.069
0.182
Yes
prediction_entropy
0.515
—
—
No
agreement_rate
0.514
0.255
0.242
No
Appendix
Table 10: Top 10 confidence methods on safety subsets (306 examples, sorted by AUROC). All AUROC values are ≈ 0.50, indicating no discriminative power.
Figure 8: Per-dataset 2×2 agreement matrices defined by majority-vote human correctness and AI correctness. Each cell shows the percentage and count of samples where both Human and AI are correct, only AI is correct, only humans are correct, or both are wrong.
Figure 9: Per-dataset AI confidence calibration curves. Points show predicted confidence vs. actual accuracy. Parenthetical counts after show the number of items observed.
Figure 10: Per-dataset AI confidence distributions, split by AI correctness. Green bars show confidence when the AI is correct; red bars when the AI is wrong. Parenthetical counts in each legend show the number of correct and incorrect items. Where green and red distributions overlap heavily, confidence cannot distinguish correct from incorrect predictions.
Figure 11: Human accuracy on the low-confidence subset ( N=122 ), split by AI correctness. Upward triangles ( ▲ ) = AI-correct items (46); downward triangles ( ▼ ) = AI-incorrect items (76). Dashed lines mark Baseline human accuracy in each group. Error bars show 95% bootstrap CIs.
AI Correctness
Condition
Obs.
Items
Accuracy [ 95% CI ]
Correct
Baseline (Unassisted)
138
46
43.5% [35.5, 51.4]
Correct
Top-2 Assistance
138
46
60.9% [52.9, 68.8]
Correct
Subtask Delegation
137
46
35.8% [27.7, 43.8]
Incorrect
Baseline (Unassisted)
228
76
19.3% [14.5, 24.6]
Incorrect
Top-2 Assistance
228
76
24.6% [19.3, 30.3]
Incorrect
Subtask Delegation
224
76
20.5% [15.6, 25.9]
Appendix
Table 11: Human accuracy by condition and AI correctness (low-confidence subset, N=122 ).
AI Correctness
Comparison
Diff (pp)
95% CI
p
Correct
Top-2 vs Baseline
+ 17.4
[ + 5.8, + 29.0]
0.006
Correct
Subtask vs Baseline
− 7.7
[ − 19.3, + 3.9]
0.211
Correct
Top-2 vs Subtask
+ 25.1
< 0.001
Incorrect
Top-2 vs Baseline
+ 5.3
[ − 2.2, + 12.7]
0.213
Incorrect
Subtask vs Baseline
+ 1.2
[ − 6.2, + 8.7]
0.810
Incorrect
Top-2 vs Subtask
+ 4.0
0.307
Appendix
Table 12: Pairwise accuracy differences vs. Baseline within each AI-correctness group (low-confidence subset, N=122 ). p -values from two-sided permutation tests (10,000 permutations).
Inaccurate AI suggestions
Accurate AI suggestions
Proper robustness
Over-reliance
Proper reliance
Under-reliance
Dataset
N
% bucket
% dataset
N
% bucket
% dataset
N
% bucket
% dataset
N
% bucket
% dataset
FACTS Search
5
27.8
16.7
8
53.3
26.7
1
4.2
3.3
0
0.0
0.0
QuALITY
0
0.0
0.0
1
6.7
3.3
10
41.7
33.3
7
35.0
23.3
BIG-Bench
7
38.9
9.6
5
33.3
6.9
5
20.8
6.9
4
20.0
5.5
GPQA diamond
1
5.6
7.1
0
0.0
0.0
2
8.3
14.3
3
15.0
21.4
Appendix
Table 13: Dataset breakdown across over- and under-reliance buckets (top-2 assistance condition). “% bucket” = share of that bucket from each dataset; “% dataset” = share of each dataset falling in that bucket.
Dataset
N items
Avg subtasks
Avg to human
Avg to AI
% to human
FACTS Search
30
4.87
2.37
2.50
48.6%
QuALITY
30
6.93
0.03
6.90
0.5%
Big-Bench
77
10.66
2.45
8.21
23.0%
GPQA Diamond
14
7.07
1.50
5.57
21.2%
Hidden Agenda
10
9.90
0.40
9.50
4.0%
Humanity’s Last Exam
14
10.07
3.29
6.79
32.6%
Appendix
Table 14: Subtask delegation pipeline: subtask routing breakdown by dataset (195 items across all study rounds including pilot). AI confidence threshold of 0.8 determines routing and subtasks below threshold are flagged for human input.
GPQA Diamond
Other MC Datasets
Human Option 1 selection rate
38.2%
20.9%
Expected rate (uniform over 4 options)
25.0%
25.0%
Two-proportion z -test
z=7.89 , p<.001
Appendix
Table 15: Option 1 selection rate on GPQA Diamond vs. other multiple-choice datasets.
Method
Accuracy
Category
Count
%
Human alone
49.9%
Both correct
390
41.0%
AI alone
68.9%
AI only correct
266
27.9%
Hybrid 2-threshold
69.3%
Human only correct
85
8.9%
Oracle (human OR AI)
77.8%
Neither correct
211
22.2%
Headroom (Oracle − AI)
+ 8.9 pp
Captured (Hybrid − AI)
+ 0.4 pp
Appendix
Table 16: Oracle upper bound and error decomposition (majority vote, N=952 test items). The oracle answers correctly if either human or AI is correct. Headroom = Oracle − AI; Captured = Hybrid − AI; Capture rate = Captured / Headroom.
Category
Mean AI Confidence
Both correct
0.811
AI only correct
0.766
Human only correct
0.652
Neither correct
0.521
Appendix
Table 17: Mean calibrated AI confidence by agreement category ( N=952 ). The “Human only correct” row reports the mean over all 85 items in that category; the 76 that remained with AI after routing had mean confidence 0.700.
Human feedback is critical for aligning AI systems to human values. As AI capabilities improve and AI is used to tackle more challenging tasks, verifying quality and safety becomes increasingly challenging. This paper explores how we can leverage AI to improve the quality of human oversight. We focus on an important safety problem that is already challenging for humans: fact-verification of AI outputs. We find that combining AI ratings and human ratings based on AI rater confidence is better than relying on either alone. Giving humans an AI fact-verification assistant further improves their accuracy, but the type of assistance matters. Displaying AI explanation, confidence, and labels leads to over-reliance, but just showing search results and evidence fosters more appropriate trust. These results have implications for Amplified Oversight -- the challenge of combining humans and AI to supervise AI systems even as they surpass human expert performance.
Rishub Jain, Sophie Bridgers, Lili Janzer +3
1Google DeepMind · 2Work done while previously at Google DeepMind
Human-AI teams fail to outperform their best member in 70% of studies, yet no theory specifies when complementarity is achievable. We derive tight bounds for the broad class of confidence-based aggregation rules by integrating signal detection theory with information-theoretic analysis, yielding four results: (1) a complementarity theorem (teams outperform individuals iff error correlation ρHM<ρ∗, with ρ∗≈a in the symmetric near-chance regime); (2) minimax bounds showing gains scale as Θ(Δd) with metacognitive sensitivity difference; (3) an impossibility result proving no confidence-based aggregation rule achieves complementarity when ρHM≥ρ∗; and (4) multi-class generalization ρK∗≈ρ∗/K−1. Predictions match observed team accuracy (R=0.94 on ImageNet-16H, R=0.91 on CIFAR-10H) and the multi-class threshold scaling holds on human data (R=0.93, K=16), with robustness under non-Gaussian distributions. The framework explains why complementarity is rare and provides actionable design formulas; results apply to aggregation, not to interactive deliberation that generates novel answers.
Dongxin Guo, Jikun Wu, Siu-Ming Yiu
The University of Hong Kong, Hong Kong, China · Stellaris AI Limited, Hong Kong, China · Brain Investing Limited, Hong Kong, China
Machine learning models are often intended to augment rather than replace human decision makers, by providing information that is complementary to human judgement. Yet, in practice, human decision makers routinely fail to realize such complementary gains, even when models provide useful signal. In this work, we study how asymmetric information about the quality of information available to a human decision maker vs. an AI impacts the ability of a decision maker to extract complementary value from AI predictions. We show that a key factor is the error correlation structure between human and AI predictions. In particular, when the AI's prediction errors are negatively correlated with those of the human, the decision maker can construct robust strategies which guarantee improvements in expected utility. We empirically investigate whether these conditions for complementarity arise in practice, using real-world forecasting benchmarks.