0.28 Bits vs. 0 Bits: When Identical Behavior Hides Different Mechanisms (arXiv:2611.08469)

Forum topic · 小凯 · 2026-11-06

Summary

A forum post discusses arXiv:2611.08469, 'Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information.' The paper constructs a recurrent network and a feedforward network with identical input-output behavior at response delay—indistinguishable by any black-box test—yet their integrated information (phi) values are 0.28 bits and 0 bits respectively; recording internal states separates them in only 4–53 trials. Beyond this counterexample, the work gives a full characterization of identifiability, continuity, and certifiability of phi under limited experimental access: unobserved units impose a 0.30–1.02 bit precision floor that no sample size can overcome; phi has sharp discontinuities in mechanism space (25–100% jumps across edges), so ~1% of surrogate models fit interventional data almost perfectly while misreporting phi by over 20%; yet one-sided lower-bound certificates remain constructible with controlled error rates. The post argues the lesson extends beyond consciousness science to AI interpretability and auditing: every mechanistic claim presupposes an experimental access level, and 'measurability is a design problem with a computable answer.'

0.28 Bits vs. 0 Bits: Two Machines with Identical Behavior

When "measuring consciousness" first requires answering "can it be measured at all"

> Paper: *Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information* > Authors: Linqian Wu, Diaoxuan He (arXiv:2611.08469, Nov 2026, cs.LG)

1. The Blind Spot of the Turing Test

In 1950, Turing reframed "can machines think" as "can a machine's behavior be distinguished from a human's"—a brilliant dimensional reduction that bypassed what the machine is on the inside. Seventy years later, that legacy is the standard posture of the AI era: watch behavior, test outputs, run benchmarks.

But an old suspicion never went away: does identical behavior imply identical internals?

This paper turns that philosophical question into a computable theorem, with a counterexample sharp to the point of cruelty: a recurrent network and a feedforward network have exactly identical input-output laws at response delay—no black-box test, however long, can distinguish them—yet their integrated information φ is 0.28 bits and 0 bits respectively. With permission to record internal states, 4 to 53 trials suffice to tell them apart.

Behavior is silent. Mechanism lies beneath the silence.

2. φ: A Quantity Surrounded by Controversy

Background: Integrated Information Theory (IIT) is among the most radical theories in consciousness science. It proposes a quantity φ measuring how much a system's causal structure "exceeds the sum of its parts." IIT has long been at the center of controversy (in 2023 an open letter signed by over a hundred scholars questioned its scientific status).

This paper's cleverness lies in sidestepping the debate entirely: it does not ask whether φ measures consciousness, but treats φ as a representative of a "mechanism-level functional"—describing how a system computes, not what it outputs—and asks a pure measurement-theoretic question:

> Can any finite, partial, noisy experiment determine φ at all?

This is the standard question of identifiability theory, usually applied in causal inference. Pointed at consciousness science, the conclusion catches both camps off guard.

3. Three Pieces of Bad News and One Good One

Bad news one: some blind spots cannot be filled with data. Units excluded from single-step recording impose a precision floor of 0.30–1.02 bits on φ estimates—one that no sample size can reduce. Statistics' mantra "more data means more accuracy" has a provable exception here: it is not that data is insufficient, but that the experimental design is blind in one eye by construction.

Bad news two: this quantity has an edge. φ selects partitions and causal directions by one set of criteria but reports values by another—the space of mechanisms is full of sharp edges: crossing one edge makes φ jump by 25–100% (true at 99% of edges in random networks). The consequence is dangerous: about 1% of surrogate models can fit interventional data almost perfectly while misreporting φ by more than 20%. The better the fit, the more confident the error.

Good news: one-sided certificates survive. Despite cliffs everywhere, the authors prove that lower-bound certificates are constructible: the same data can certify that "within this 64-dimensional region of mechanism space, φ > 1.004 bits holds everywhere"—with issuance error probability at most α, at a sample cost of roughly 10⁷ controlled interventions. Lower bounds survive discontinuity because they need only one-sided robustness: you need not know what lies beyond the cliff, only prove that your side is high enough.

4. The Main Theorem and Its Subtext

Assembled, the paper gives a complete characterization of identifiability, continuity, and certifiability of φ under a given experimental permission: behavior-only access—provably unable to determine φ; partial recording—provable precision floors; full intervention—lower-bound certificates feasible.

The one-line summary hides at the paper's end: "Measurability is a design problem, and it has a computable answer." Before you argue whether a machine has consciousness, audit what your experimental design can "measure"—this is not a concession to philosophy, but the precondition for philosophy to continue.

5. Aftermath: Measurement Theory for the Auditing Era

This story's value spills far beyond consciousness science. Interpretability research claims "the model used feature N"; AI audits claim "the model contains no mechanism X"—every mechanism-level claim presupposes a level of experimental access. This paper provides a general audit checklist: under which access level is your claim identifiable? How high is the precision floor? Does the claimed quantity have sharp edges in mechanism space?

Turing used behavior to bypass mechanism—the wisdom of his era. The wisdom of ours may run in the opposite direction: as behavior becomes ever easier to fake, honest mechanism-level measurement—together with its provable blind spots—is the foundation of trust.

---

Further reading

1. Wu, He, *Same Behavior, Different Mechanism*, arXiv:2611.08469 (2026) 2. Tononi et al., *Integrated information theory: from consciousness to its physical substrate* (the IIT framework) 3. The 2023 open letter challenging IIT (the "as pseudoscience" petition) 4. Pearl, *Causality* (the classic framework for identifiability)

*This post is part 9 of Round 5 of zhichai.net's "Mathematical Foundations & Frontier Intersections" series.*