Foundations

Why AlphaZero Does Not Count as Intelligent for Bernard Jennings

AlphaZero learned through self-play to defeat leading chess, shogi, and Go programs. Bernard Jennings nevertheless excludes it from his definition of intelligence, which demands more than successful problem-solving. This article explains his classification and examines the evidence still needed for his proposal and its claimed upper limit on intelligence.

Pale mechanical blocks rise through heavy brushwork until they meet a translucent upper plane. An orange line marks the contact, with a lone indigo block visible farther along the ramp above.
Pale mechanical blocks rise through heavy brushwork until they meet a translucent upper plane. An orange line marks the contact, with a lone indigo block visible farther along the ramp above.

AlphaZero learned chess, shogi, and Go through self-play and defeated leading programs.1 Bernard Jennings nevertheless classifies it as non-intelligent. In his account, even this performance is insufficient. A system would have to identify problems of its own and resolve them in ways not already determined by its existing capabilities and input.

Jennings published his proposed definition in The Architecture of Intelligence: Definition, Spectrum, and Ceiling, an independent working paper released on Zenodo on May 22, 2026. It has not been peer-reviewed and is not an established measure of intelligence.2 AlphaZero makes his use of the criteria concrete. Whether those criteria hold up empirically, or justify a fixed ceiling on intelligence, remains to be tested.

The boundary Jennings draws

Demand-ownership means that a system registers a gap, conflict, or inefficiency within its own patterns as something requiring resolution. Receiving an external task is insufficient. Jennings assigns this function to what he calls “Category 2 emotional architecture,” by which he means state changes that operate as internal control signals. The term makes no claim about felt emotion in this paper. The existence and relevance of the proposed architecture remain part of the theory at issue.

Resolution-ownership concerns what happens next. A resolution retrieved from storage or produced by executing a fully specified path does not qualify. The system must generate the path in response to the demand it has registered. Jennings treats this as a property of a particular resolution and allows mixed cases within one system.

The proposed observable marker is residue beyond encoding. “Encoding,” or layer 1, includes the system’s stable competence: architecture, trained parameters, consolidated dispositions, and the configured operating procedure. Residue appears when layer 1 together with the input does not exhaust the behavior involved in detecting and resolving a coherence demand.

Several familiar signs of capability fail this test. Stochastic output remains part of layer 1 when the source of randomness is configured there. A result can be new in history and still follow from stable competence. Expensive prediction shows computational difficulty, not residue.

Applying the criterion is correspondingly difficult. Jennings acknowledges that a complete practical specification of layer 1 may be unavailable and that judgment is required. The paper classifies cases in theoretical terms; it does not itself supply an empirical demonstration of residue.2

Why AlphaZero receives a zero

AlphaZero gives the framework a demanding example. David Silver and colleagues reported in 2018 that systems using the same general reinforcement-learning approach mastered chess, shogi, and Go after separate self-play training for each game. According to the paper, each system received no domain knowledge beyond the rules. A policy-and-value network worked with a general Monte Carlo tree search, and the trained systems defeated leading programs in all three games.1

Those facts establish exceptional performance. They do not directly measure any of Jennings’s three conditions. The game, rules, and learning signal were supplied externally, while the AlphaZero research did not investigate whether the system registered internal pattern conflicts as demands of its own.

Self-play also complicates an intuitive reading of resolution-ownership. AlphaZero does not retrieve moves from a database of human games. Its search and learned network produce them. Jennings nevertheless asks whether a resolution comes from an internally registered demand and exceeds the execution of stable competence. The Science paper has no measure for that distinction.

Within his own framework, Jennings gives a definite answer at inference time. The trained policy and value network and the search procedure belong to layer 1; the board position is the input. He says they determine the output, assigns AlphaZero zero residue, and places it off the intelligence spectrum. Training is treated as the process that produces layer 1.

This classification is a consequence of the new definition. Silver and his co-authors did not report it, and Jennings does not test a complete layer-1 prediction against observed AlphaZero behavior. The example shows where his boundary runs without validating that boundary independently.

A ceiling borrowed from companion work

The paper also proposes an upper limit. Jennings places it at fully integrated and optimally calibrated operation of eight Coherence Resolution Modes. His intelligence paper inherits the claim that these modes are complete from a separate working paper in the same research program.3

The dependency is explicit. A fourth type of constraint, a fifth resolution mechanism, or another admissible combination would overturn the completeness claim and take the ceiling with it. These are meaningful failure conditions. They do not amount to independent corroboration, since the main paper and its support share the same author and theoretical foundations.

For now, the ceiling is a conditional hypothesis. The neatness of an eight-part typology cannot establish that the typology covers every relevant phenomenon.

What organizations can take from the distinction

An organization does not need to accept Jennings’s spectrum to benefit from one distinction raised by the debate. AlphaZero shows how operational capability can grow under a tightly specified goal. Compute, training, and search affect what the system can achieve within the game. They leave the purpose of deployment and responsibility for its consequences with the people and institutions using it.

Neither source addresses the governance of a deployment. In practice, the organization still has to name the capability being increased, the person or body that sets its objective and limits, and the role that may turn an output into action. Observed consequences then need to reach someone authorized to change or stop the arrangement.

Jennings could classify a system as non-intelligent while that system remains highly useful or dangerous. Passing his three-part test would likewise confer no authority to deploy. The paper is therefore most valuable as a provisional attempt to say exactly what performance leaves unanswered. Its own empirical and theoretical work is not yet complete enough to settle the category it proposes.

Footnotes

  1. David Silver et al., “A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go Through Self-Play”, Science 362 (2018), 1140–1144. The earlier open preprint describes the chess/shogi version of the approach. ↩ ↩2

  2. Bernard Jennings, The Architecture of Intelligence: Definition, Spectrum, and Ceiling, version 01, independent working paper, 22 May 2026. The publication status comes from the Zenodo record and the document; neither identifies a peer-reviewed publication. ↩ ↩2

  3. Bernard Jennings, Coherence Resolution Modes in Autonomous Systems, working paper, 2026. The main paper expressly makes its ceiling conditional on the completeness argument attributed to this companion work. ↩

Oliver Wrede writes and teaches on interface design, knowledge systems, and the architecture of intelligence in organizations. He is interested in how humans, institutions, and machines reason together — and how design shapes the quality of that reasoning.

More from Oliver Wrede

Intelligence Architecture

The Business Value of Critical Thinking When AI Does the Work

AI agents can produce analyses and recommendations in seconds. Their business value also depends on whether employees ask useful questions, assess evidence, and learn from the results. Intelligence Architecture therefore needs to develop people's capabilities. How can everyday work support that development, and how can an organization assess its value?

Oliver Wrede·
Intelligence Architecture

AI and the Unwritten Rules of an Organization

Process documents describe only part of how an organization works. AI that answers from documents and stored data can miss informal responsibilities and essential coordination. This article asks how that work can be recorded for AI without eliminating useful discretion or subjecting staff to new forms of monitoring.

Oliver Wrede·