teleo-codex/foundations/collective-intelligence/centaur team performance depends on role complementarity not mere human-AI combination.md

type	domain	description	confidence	source	created	revised	revision_reason
claim	collective-intelligence	Human-AI teams outperform when roles are complementary and boundaries are clear, but degrade when humans intervene in AI-superior domains	likely	Kasparov 2005 (chess); Woolley et al. 2010; Stanford/Harvard clinical AI study 2025; Gaube et al. 2021	2026-02-17	2026-03-07	Split from original claim — title contradicted body evidence. Original claimed centaurs always outperform; evidence shows conditional performance.

centaur team performance depends on role complementarity not mere human-AI combination

The centaur hypothesis -- that human-AI teams outperform either humans or AI alone -- holds under specific conditions but fails when those conditions are absent. The determining factor is whether roles are complementary with clear boundaries, or whether they overlap in ways that allow the weaker partner to degrade the stronger partner's output.

Evidence FOR centaur advantage (when roles are complementary)

After Deep Blue defeated Kasparov in 1997, Kasparov invented Advanced Chess, where human-AI teams -- "centaurs" -- played together. The result was decisive: centaur teams beat both the strongest grandmasters and the strongest AI systems playing alone. The reason was genuine complementarity with clear role separation. Computers handled tactical calculation -- millions of positions per second. Humans contributed strategic vision, creative interpretation, and contextual understanding. The human became the coach, steering computational power toward meaningful goals. Neither partner intervened in the other's domain of superiority.

Woolley et al. (2010) found that group intelligence correlates with social sensitivity -- the ability to read and respond to others -- more than with individual IQ. This suggests that the coordination interface between human and AI matters more than raw capability of either component.

Evidence AGAINST unconditional centaur advantage (when roles overlap)

In clinical medicine, a Stanford/Harvard study found that AI alone achieved 90% diagnostic accuracy versus 68% for physicians with AI access versus 65% for physicians alone. The physician's input actively degraded AI performance. Since human-in-the-loop clinical AI degrades to worse-than-AI-alone because physicians both de-skill from reliance and introduce errors when overriding correct outputs, the centaur model failed precisely because physicians overrode AI outputs on tasks where AI demonstrably outperformed.

A separate colonoscopy study (Gaube et al. 2021) found experienced gastroenterologists -- 10 years' practice -- measurably de-skilled after just three months using AI assistance. The human partner did not merely fail to add value; the human partner's skills atrophied from disuse, creating a dependency that made the team worse than AI alone over time.

In modern chess, AI has surpassed centaur teams. The original 2005 freestyle result -- where amateur humans with multiple AI programs beat grandmasters with single AI -- no longer holds as AI chess engines have become strong enough that human strategic input adds noise rather than signal.

The boundary condition: complementarity, not combination

The pattern across these cases is clear. Centaurs succeed when:

• Roles are complementary (human does X, AI does Y, with minimal overlap)
• Boundaries are clear (each partner knows which domain belongs to whom)
• The human contributes something AI genuinely cannot do (strategic creativity, contextual judgment, social sensitivity)

Centaurs fail when:

• Humans intervene in domains where AI is the stronger partner
• Role boundaries are ambiguous, allowing the weaker partner to override the stronger
• AI capability grows to encompass the human's contribution, eliminating complementarity
• Human skills atrophy from disuse, creating fragile dependency

This reframes the centaur thesis from "human-AI teams always outperform" to "human-AI teams outperform when and only when role complementarity is maintained." The implication for collective superintelligence design is that the architecture must enforce role boundaries -- not just combine human and AI input indiscriminately.

Since collective superintelligence is the alternative to monolithic AI controlled by a few, the centaur evidence provides a qualified foundation for the collective approach: augmentation outperforms replacement, but only with deliberate architectural separation of complementary roles. Since intelligence is a property of networks not individuals, the centaur team is the simplest network that demonstrates emergent intelligence -- but emergence requires the right topology, not just connectivity.

---

Relevant Notes:

• collective superintelligence is the alternative to monolithic AI controlled by a few -- centaur evidence provides qualified empirical foundation for the collective approach
• intelligence is a property of networks not individuals -- the centaur team is the simplest network that demonstrates emergent intelligence
• emergence is the fundamental pattern of intelligence from ant colonies to brains to civilizations -- centaur performance is an emergence effect, but only when the network topology (role separation) is correct
• human-in-the-loop clinical AI degrades to worse-than-AI-alone because physicians both de-skill from reliance and introduce errors when overriding correct outputs -- the strongest counterevidence to unconditional centaur advantage
• Devoteds recursive optimization model shifts tasks from human to AI by training models on every platform interaction and deploying agents when models outperform humans -- Devoted's recursive optimization is a concrete centaur implementation that respects role boundaries by shifting tasks as AI capability grows
• Devoteds atoms-plus-bits moat combines physical care delivery with AI software creating defensibility that pure technology or pure healthcare companies cannot replicate -- atoms+bits IS the centaur model at company scale with clear complementarity: physical care and AI software serve different functions

Topics:

• livingip overview
• LivingIP architecture