Hybrid Confirmation Tree (HCT)

• The paper introduces HCT, a method combining independent human and AI binary decisions with a tiebreaker to resolve disagreements.
• HCT improves efficiency by reducing human inputs by 28%-44% and outperforming traditional majority human aggregation in various real-world datasets.
• The method’s success depends on maintaining independent, complementary judgments, thereby preserving human agency and auditability in high-stakes applications.

The hybrid confirmation tree (HCT) is a hybrid intelligence strategy for binary classification in which a human and an AI make independent initial decisions, agreement is accepted immediately, and disagreement triggers a second human tiebreaker. Introduced as a simple, transparent, and sequential aggregation method, it is designed to combine human and artificial intelligence while maintaining human agency, because every final decision is always approved by a human (Berger et al., 2 Feb 2026). Subsequent comparative work positions HCT as an alternative to the standard AI-as-advisor paradigm by preserving the independence of human and AI judgments rather than presenting AI output as advice to be accepted or rejected by a human (Berger et al., 31 Mar 2026).

1. Decision protocol and organizational logic

The HCT has a three-step operational structure. First, a human decision-maker $H1$ and an AI system $M$ make independent decisions on the same binary classification task. Second, if $H1$ and $M$ agree, that consensus is accepted as the final decision. Third, if they disagree, a second human $H2$ is consulted, and $H2$'s decision determines the final outcome (Berger et al., 2 Feb 2026).

The defining feature of the protocol is independence. In the comparative formulation against AI-as-advisor, both the human and the AI are required to make their initial judgments in parallel and without knowledge of the other’s choice. Disagreement is not handled by negotiation or revision of the first human’s judgment after exposure to AI output; instead, it is resolved by a fresh, independent human judgment. The “tree” terminology refers to this branching structure: the root split is agreement versus disagreement, and only the disagreement branch extends to a tiebreaker (Berger et al., 31 Mar 2026).

A central design property is preservation of human agency. In the original formulation, every final decision is always approved by a human, whether through the initial agreement path or through the tiebreaking path (Berger et al., 2 Feb 2026). This makes HCT structurally distinct from purely automated adjudication and from advisory interfaces in which AI recommendations can shape the human’s first-order judgment before an independent human opinion has been recorded.

2. Analytical formulation and cost structure

The analytical treatment in the original HCT paper defines $p_H$ as human accuracy and $p_M$ as machine accuracy. Under independence, HCT is correct when the human and machine are both correct, or when one of them is correct and the tiebreaking human is also correct. Its accuracy is therefore given by

$$\pi_{HCT} = p_H p_M + p_H (1 - p_M) p_H + (1 - p_H) p_M p_H$$

The expected human cost follows from the fact that $H1$ is always consulted, while $M$0 is consulted only on disagreement. With disagreement probability

$M$1

the expected cost is

$M$2

For comparison, a three-person human majority vote with independent, identical human accuracy $M$3 has accuracy

$M$4

and expected cost

$M$5

These derivations support the paper’s central claim that HCT can match and exceed the accuracy of a three-person human majority vote while requiring fewer human inputs, particularly when AI accuracy is comparable to or exceeds human accuracy (Berger et al., 2 Feb 2026).

The model is also generalized to correlated decisions using Cohen’s $M$6. In that setting, the benefit of aggregation falls as correlation rises: if decisions are highly correlated, aggregating them is less beneficial. This provides the formal basis for treating independence and partial error decorrelation as central resources in hybrid collective intelligence (Berger et al., 2 Feb 2026).

3. Complementarity, correlation, and operating thresholds

The HCT paper defines complementarity as the regime in which the hybrid system outperforms individual humans, AI, and the majority vote. Its analytical result is that HCT’s ability to achieve complementarity is maximized when human and AI accuracies are similar and their decisions are not overly correlated (Berger et al., 2 Feb 2026).

The reported boundary conditions are specific. HCT achieves its greatest advantage when machine accuracy is similar to or just above average human accuracy, when human-machine correlation is lower than human-human correlation, and when neither agent is below chance accuracy. The paper further reports that if the machine is far superior, the machine alone is best; if it is far inferior, majority human aggregation works better. Figure 3B is described as showing that the complementarity window shrinks as correlation increases, and that the complementarity region grows when human-AI decisions are less correlated than human-human decisions (Berger et al., 2 Feb 2026).

HCT also has an operating-threshold interpretation when the AI output is probabilistic. By tuning the threshold at which the AI counts a case as positive, the system can move between two limiting human-only heuristics. At threshold $M$7, the AI always votes positive and HCT behaves like a polyarchy. At threshold $M$8, the AI only votes positive if certain and HCT behaves like a hierarchy. The paper states that this gives HCT greater flexibility in navigating true and false positive trade-offs than fixed human-only heuristics such as hierarchies and polyarchies (Berger et al., 2 Feb 2026).

A common misunderstanding is that HCT is simply a hybridized majority vote. The analytical framing suggests a more precise characterization: HCT is a conditional aggregation rule whose performance depends not only on marginal accuracies but also on disagreement structure, because the second human is invoked only when the first human and the AI diverge.

4. Empirical performance relative to human-only aggregation

The original empirical validation reanalyzes six real-world datasets: skin cancer diagnosis using dermoscopy images, skin cancer diagnosis using non-dermoscopy images, deepfake video detection, criminal rearrest prediction, the Hybrid Forecasting Competition, and ForecastBench for LLM-based geopolitical forecasts (Berger et al., 2 Feb 2026).

Across these datasets, HCT reportedly outperformed a three-human majority vote. The reported gains ranged from $M$9 to $H1$0 percentage points, including $H1$1 points in skin cancer diagnosis with dermoscopy images, $H1$2 in deepfake detection, and $H1$3 in criminal rearrest prediction. These differences are described as credible Bayesian differences with high statistical confidence. The same study reports that HCT achieved these gains with $H1$4 to $H1$5 fewer human inputs than majority vote, including a $H1$6 reduction in criminal rearrest and a $H1$7 reduction in hybrid forecasting (Berger et al., 2 Feb 2026).

The empirical claims are not limited to comparison with three-human groups. The study also states that HCT outperformed even large human-only groups, up to 15-person majorities in the reported experiments (Berger et al., 2 Feb 2026). This is significant because the comparison is not merely between a hybrid scheme and a minimal human baseline; it is between HCT and larger human collectives assembled under conventional majority aggregation.

These results support the paper’s broader claim that HCT is a practical, efficient, and robust strategy for hybrid collective intelligence. A plausible implication is that the benefit is not reducible to substituting a strong AI for a weak human; rather, the reported gains depend on selective escalation on disagreement and on the complementarity of error patterns.

5. Comparison with the AI-as-advisor paradigm

In the AI-as-advisor approach, AI issues a recommendation and a human decides whether to accept or reject it. The comparison paper argues that this standard paradigm has several limitations: people frequently ignore accurate advice, rely too much on inaccurate advice, and their decision-making skills may deteriorate over time. HCT is presented as an alternative because it preserves the independence of human and AI judgments instead of turning AI output into advice for a human to revise around (Berger et al., 31 Mar 2026).

The comparison uses 10 datasets from various domains, including medical diagnostics and misinformation discernment, and reports more than 41,000 human decisions. On this benchmark, HCT outperformed AI-as-advisor on all datasets. The reported improvement ranged from $H1$8 to $H1$9 percentage points per dataset, with an average improvement of $M$0 percentage points over AI-as-advisor and a $M$1 HDI of $M$2, together with a $M$3 chance of practical significance (Berger et al., 31 Mar 2026).

The paper also evaluates settings in which AI provided explanations of its judgment. Across 16 explainable AI conditions and more than 50,000 choices, HCT was more accurate in 11 of 16 cases and at least as good in two more; only in rare “coin-flip” accuracy settings did AI-as-advisor outperform HCT. Reported gains held across domains and across both high- and low-expertise humans, with low performers improving the most, though even top-tier experts benefited. Mechanistically, HCT captured correct AI decisions more frequently: when AI was correct and disagreed with the human, the tiebreaking human selected the correct answer $M$4 of the time, compared with $M$5 for a human updating an initial disagreement in AI-as-advisor (Berger et al., 31 Mar 2026).

These findings directly challenge the view that explanations or exposure to AI output are sufficient to recover hybrid accuracy. The reported evidence indicates that independent aggregation can outperform advisory integration even when AI explanations are available.

6. Signal detection account, limitations, and deployment implications

The comparison paper interprets the HCT advantage using signal detection theory. With $M$6 denoting human accuracy and $M$7 denoting AI accuracy, AI-as-advisor accuracy is written as

$M$8

where $M$9 is the rate of accepting correct AI advice and $H2$0 is the rate of accepting incorrect AI advice. The associated discrimination and response-bias measures are

$H2$1

The paper reports that participants’ discrimination ability $H2$2 was not sufficient to reliably separate correct from incorrect AI advice, and that response bias $H2$3 was generally conservative: participants hesitated to accept advice even when AI was more accurate. On this account, HCT outperforms AI-as-advisor because people cannot discriminate well enough between correct and incorrect AI advice, and the signal detection model reportedly predicted which paradigm was superior in 37 of 41 empirical comparisons (Berger et al., 31 Mar 2026).

The published limitations are equally specific. HCT increases per-decision labor for a fraction of cases: depending on dataset, $H2$4 to $H2$5 of cases require a tiebreaker. The approach also requires organizational enforcement of independence, because its mathematical and practical justification depends on humans making judgments without AI influence at the initial stage. The comparison paper notes that HCT can be less effective when human accuracy is at or below chance, since tiebreakers are then no better than random, and that workflow integration and personnel management are nontrivial despite the protocol’s conceptual simplicity (Berger et al., 31 Mar 2026).

At the same time, both papers emphasize properties that are relevant in high-stakes deployment. HCT maintains clear human agency and auditability, is described as favorable under legislation such as the EU AI Act, and may better preserve human skills over time than AI-as-advisor, which risks deskilling through overreliance (Berger et al., 31 Mar 2026). The original HCT paper adds that the method is immediately implementable in workflows that require or benefit from a second opinion or consensus, such as medical double-reading, and that it is suitable for domains including medicine, law, and forecasting (Berger et al., 2 Feb 2026).

Taken together, the reported results characterize HCT as a conditional aggregation mechanism rather than an advisory interface. Its documented advantages arise when independence is preserved, human and AI accuracies are similar or the AI is somewhat stronger, and their errors are not overly correlated. Its documented limitations arise when these conditions fail, when human judgments approach chance, or when the cost of occasional tiebreaking is unacceptable.