Human Capital, Not Model Benchmarks, Predicts Hybrid Intelligence in Forecasting

Abstract: Whether pairing people with AI helps or hurts is usually reported as a single average effect. Using a real-money prediction market (Polymarket) as an objective, externally resolved benchmark, this pilot shows that the value of human-AI collaboration depends on a specific, measurable form of human capital. Analyzed at the level of the individual forecaster, hybrid performance is trimodal: most people either deferred to the model (matching it) or used it to rubber-stamp a prior guess (performing worse than the model alone), while a minority engaged in genuine complementary reasoning and reached accuracy matching or even exceeding (i.e., lower error than) the market itself. Collaborative traits (perspective-taking, intellectual humility, and curiosity) rather than raw cognitive ability or model benchmarks, distinguished who reached that mode. The results are preliminary but statistically robust, and motivate a pre-registered replication now in preparation.

• The paper shows that hybrid forecasting success hinges on collaborative human capital rather than AI benchmark scores.
• It finds that Cyborgs, who engage in interactive reasoning with LLMs, achieve forecasting errors nearly on par with top market benchmarks.
• The study suggests that optimizing human-AI workflows requires selecting for traits like perspective-taking, curiosity, and intellectual humility.

Human Capital as the Critical Factor in Hybrid Human-AI Forecasting

Overview

This paper, "Human Capital, Not Model Benchmarks, Predicts Hybrid Intelligence in Forecasting" (2607.02467), interrogates the prevailing assumption that higher-performing AI models unilaterally translate to greater hybrid human-AI team efficacy. Using a rigorous experimental design centered on externally-verified real-money prediction markets, it demonstrates that hybrid forecasting performance is neither monolithic nor strongly predicted by raw cognitive ability (“g”) or model benchmark scores. Instead, the results identify distinct modes of human-AI interaction, with a critical minority of “Cyborg” forecasters—characterized by elevated collaborative human capital—achieving performance on par with, or exceeding, both leading LLMs and the prediction market reference.

Experimental Design and Methods

The study sampled 108 adults, with 78 participating in the main experiment across Human-only and Hybrid (with LLM access) conditions. Forecasting spanned 30 economic, geopolitical, and business outcomes, with each participant making individual probabilistic judgments resolved by Polymarket contracts. Four state-of-the-art LLMs served as machine baselines: Llama 3.1 (8B), Qwen3 (8B), GPT-4o, and Gemini 3 Pro. Teams of three collaborated, yet all individual responses were aggregated for analysis.

Participants completed diagnostic assessments of both cognitive ability (Raven’s, ICAR “g”) and collaborative human-capital traits: the IRI Perspective-Taking subscale, Comprehensive Intellectual Humility Scale, a curiosity inventory, and others. Hybrid interaction styles were classified post hoc into three categories: Automators (defer to model output), Validators (use AI to check existing intuition), and Cyborgs (engage in complementary, iterative reasoning).

Empirical Results

Performance Heterogeneity and Interaction Modes

Forecasting error, measured via scaled Brier scores, revealed a pronounced trimodal distribution:

• Human-only participants: mean 14.7
• Hybrid Automators: 10.4 (better than unaided humans, worse than LLMs)
• Hybrid Validators: 31.7 (substantially underperforming both AI and humans)
• Hybrid Cyborgs: 3.8 (lowest error, matching/exceeding best LLMs and the market benchmark)

Only the Cyborg group achieved performance on par with the Polymarket baseline (3.5), outperforming each individual LLM (LLMs ranged 4.6–7.1). Notably, Validator performance strongly regressed when utilizing the AI merely as post-hoc justification, undermining synergy and reducing accuracy below that of human-only interaction.

Predictors of Hybrid Success

Key findings:

• Cognitive ability (‘g’) and fluid reasoning: predicted solo human accuracy (correlations ~−0.45 to −0.66), but not hybrid performance (correlations near zero).
• Collaborative human capital (perspective-taking, curiosity, intellectual humility): significantly predicted superior hybrid outcomes (correlations to lower error: perspective-taking r = −0.32; curiosity r = −0.26; intellectual humility r = −0.25).
• Logistic models indicated a sharp, non-identifiable divide: only high-collaborative-capital individuals attained Cyborg status, suggesting nonlinearity in trait interaction.

Qualitative process analysis suggested that effective human-AI “division of labor” arose only in Cyborgs, who leveraged LLMs for fact-finding and probabilistic anchoring, while humans interrogated strategic uncertainty and model blind spots.

Supplementary Analyses

A further experiment with a “Socratic” LLM (withholds answers, prompts only) did not enhance accuracy per se, but increased Cyborg-type interaction frequency and precluded Automator-style passive reliance. Preliminary EEG data indicated reduced cognitive engagement among Automators relative to both Cyborgs and human-only forecasters.

Implications

Theoretical

These results interrogate and nuance the assumption that simply embedding AI into human workflows enhances decision-making. Specifically, the interplay between human collaborative capacity and AI benchmarking is nontrivial: “g” alone is insufficient for hybrid gains. Instead, the emergence of complementary reasoning—facilitated by perspective-taking, openness, and curiosity—aligns with theoretical models positing that human-AI collective intelligence is governed by interaction strategy rather than constituent ability scores.

Practical

For organizations deploying AI-augmented decision support, these findings imply that user screening, training, and interface design should target collaborative traits over traditional measures of intelligence or technical expertise. AI “readiness” effectively becomes a function of advanced theory-of-mind and epistemic humility, not raw IQ. Furthermore, hybrid deployment should not assume uniform uplift: in specific regimes, especially when AI is used as a validator rather than a partner, performance may degrade compared to either human or AI alone.

Future Prospects

The forthcoming pre-registered replication (outlined in the paper) intends to manipulate interaction style and formally composite human-capital metrics, potentially enabling causal identification. If confirmed, these results could inform adaptive AI interfaces that diagnose user traits and dynamically scaffold hybridization toward the Cyborg regime. They also signal a transition in AI “augmentation” research: from benchmarking state-of-the-art models to optimizing for human-AI collaborative fit.

Conclusion

This study rigorously shows that collaborative human capital—not LLM scale or human intelligence metrics—drives effective hybrid forecasting. Superior outcomes are neither guaranteed by AI integration nor predicted by traditional benchmarks; rather, they depend on specific human capacities for perspective-taking, curiosity, and humility, facilitating genuinely complementary human-AI reasoning. This has broad implications for the evaluation, deployment, and future design of hybrid intelligence systems, underscoring the need to reconceptualize both collaborator selection and interface paradigms in AI-assisted decision domains.