Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information
arXiv cs.AI · 6 Nov 2026, 04:55 AEST
The study explores how finite experiments can identify mechanism-level quantities in AI systems, particularly focusing on integrated information theory. It demonstrates that behavior can be misleading, as different network structures can yield identical outputs while having distinct internal dynamics.
Key facts
- Interpretability, AI auditing and assessments of machine consciousness increasingly rest on mechanism-level quantities: functionals of a system's causal transition law that describe how it computes, not what it outputs.
- Behavior can be silent: a recurrent and a feed-forward network with identical input-output laws at their response delay carry 0.28 and 0 bits, yet internal recordings separate them within 4-53 trials.
- The measure has sharp edges: because $6$ selects a partition, and the cause and effect it specifies, by one criterion and reports another, infinitesimal changes of mechanism that cross an edge change it by 25-100% at 99% of the edges we found in random networks.
- One-sided claims survive: the same data certify a lower bound on both sides of an edge, and we prove $6>1.004$ bits throughout a 64-dimensional region of mechanisms and certify it from about $10^7$ controlled transitions with false-certificate probability at most $α$.
Read aloud
The study explores how finite experiments can identify mechanism-level quantities in A I systems, particularly focusing on integrated information theory. It demonstrates that behavior can be misleading, as different network structures can yield identical outputs while having distinct internal dynamics. Interpretability, A I auditing and assessments of machine consciousness increasingly rest on mechanism-level quantities: functionals of a system's causal transition law that describe how it computes, not what it outputs. Behavior can be silent: a recurrent and a feed-forward network with identical input-output laws at their response delay carry zero point two eight and zero bits, yet internal recordings separate them within four-fifty three trials. The measure has sharp edges: because $six$ selects a partition, and the cause and effect it specifies, by one criterion and reports another, infinitesimal changes of mechanism that cross an edge change it by twenty five-one hundred percent at ninety nine percent of the edges we found in random networks. One-sided claims survive: the same data certify a lower bound on both sides of an edge, and we prove $six one point zero zero four$ bits throughout a sixty four-dimensional region of mechanisms and certify it from about ten dollars^seven$ controlled transitions with false-certificate probability at most $α$.
Sources in this cluster
- Same Behavior, Different Mechanism: What Finite Experiments Can Certify About Integrated Information
arXiv cs.AI · 6 Nov 2026, 04:55 AEST