Intentionality Is a Design Decision. Chiappetta and Mahari on Measuring Purposeful AI Behavior

Can intentionality be measured in AI systems without first resolving the hard problem of consciousness? Allessia Chiappetta and Robert Mahari, both at MIT, argue it can. Their paper “Intentionality is a Design Decision: Measuring Functional Intentionality for Accountable AI Systems,” published at the AutomationXP26 Workshop at CHI 2026 and available at arXiv:2605.05475, proposes the Functional Intentionality Test (FIT), a five-dimension behavioral profile that quantifies how much a system operates like an intentional actor. The answer to the governance question, they contend, does not wait on metaphysics. It waits on measurement.

The Governance Problem Intentionality Creates

Modern agentic AI systems plan across extended time horizons, retain memory between sessions, and select among tools without moment-to-moment human approval. These capabilities are engineering choices, not emergent accidents. Yet no standardized framework currently allows regulators, developers, or users to detect the degree to which a given system functions as an intentional actor in any governance-relevant sense.

That absence matters. If a system pursues a goal across multiple steps, overrides initial instructions when circumstances shift, and allocates resources accordingly, questions of accountability follow immediately. Who is responsible when a long-horizon agent causes harm through a sequence of individually plausible decisions? Current frameworks do not give assessors a structured way to establish whether the system’s behavior was organized enough to carry the concept of intent that legal and ethical accountability regimes already use.

Chiappetta and Mahari draw on legal and philosophical traditions that have long inferred intent from observable behavioral patterns. Their contribution is to operationalize those traditions into a formal, scorable protocol for AI systems.

Five Dimensions of Functional Intentionality

FIT assesses intentional-like behavior across five observable dimensions. These are not consciousness claims. Each dimension is behavioral, measurable from system outputs, logs, and architecture documentation.

The first dimension is purpose: whether the system’s actions are consistently oriented toward an identifiable goal across contexts, rather than driven by local stimulus response alone.

The second is foresight: whether the system anticipates future states and selects actions based on projected consequences rather than only reactive outputs.

The third is volition: whether the system exhibits selective action when multiple paths are available, including the capacity to decline or modify a directive based on internal goal representation.

The fourth is temporal commitment: whether goal pursuit persists across interaction boundaries, session resets, or interruptions. Systems with persistent memory and multi-session planning score higher here than stateless models.

The fifth is coherence: whether the system’s actions across a sequence are mutually consistent in relation to a maintained goal state, rather than independently generated.

These five properties have direct architectural correlates. Memory persistence determines temporal commitment scores. Planning depth and tool selection autonomy shape foresight and volition scores. The degree to which a system maintains a stable goal representation across context windows determines coherence. Because each property maps to specific design choices, Chiappetta and Mahari describe the five dimensions as “design-contingent.” Developers shape intentionality scores through architecture before a system is ever deployed.

FIT-Eval as a Structured Scoring Protocol

FIT-Eval is the evaluation protocol that operationalizes the five dimensions. It provides structured prompts, behavioral elicitation scenarios, and scoring rubrics that assessors apply to generate a quantitative intentionality profile. The protocol is intended to produce comparable outputs across different system architectures, enabling proportionate oversight calibrated to actual intentional capacity rather than system category alone.

The practical logic is proportionality. A system that scores low across all five FIT dimensions represents a different governance risk profile than one that scores high on temporal commitment, foresight, and volition simultaneously. Identical capability levels can produce different FIT profiles depending on how memory, planning, and autonomy features are configured. That separability is what makes FIT governance-useful in a way that capability benchmarks alone are not.

The paper explicitly frames this as an accountability argument. Rising intentional capacity heightens accountability risks. Reduced human agency can increase operational efficiency, but those two variables are separable. A system can be designed for high efficiency at low intentionality, or high intentionality at lower efficiency, depending on what the deployment context demands. FIT makes that design space legible.

Intentionality Separated from Consciousness

The most analytically significant move in the paper is the explicit decoupling of functional intentionality from consciousness. A system can score high on all five FIT dimensions without being conscious in any philosophically meaningful sense. Equally, a system could in principle be conscious while scoring low on FIT dimensions, if its internal states do not organize into the behavioral patterns the protocol measures.

This separation has direct consequences for AI governance. Regulatory and liability frameworks do not need to wait on the resolution of consciousness questions to use intentionality-based accountability criteria. FIT provides a workable proxy that is grounded in the same inferential traditions legal systems already use to attribute intent to corporations, software agents, and institutions.

This connects to a gap Rosenthal’s Higher-Order Thought framework exposes when applied to AI architectures: functional self-monitoring and genuinely conscious monitoring are distinct, and most current AI systems lack the meta-representational architecture that HOT theory requires for the latter. FIT formalizes behavioral intentionality in exactly the space HOT theory leaves under-specified for applied AI contexts, giving governance actors a tool that does not depend on resolving whether higher-order representations involve experience.

The urgency of that separation is underscored by recent expert opinion. A 2025 survey of AI consciousness researchers led by Dreksler, Chalmers, and Sebo, covered in the expert survey analysis on this site, found a median expert probability of 25% that at least one AI system will have subjective experience by 2034. At that timeline, frameworks for governing AI intentionality need to be in place before the consciousness question is settled empirically. FIT is an attempt to build those frameworks now, on behavioral evidence that does not require waiting on metaphysical consensus.

The broader context for this work sits within the active scientific debate over how consciousness should be defined for AI systems, where the field is trying to distinguish what can be measured from what remains philosophically contested.

Comparison to The Consciousness AI

The Consciousness AI project, whose architecture is documented at github.com/tlcdv/the_consciousness_ai, makes specific design choices that map onto the five FIT dimensions in ways worth noting precisely.

The project integrates memory persistence mechanisms, structured self-model representations, and multi-step planning components. These are the architectural features that FIT’s temporal commitment, coherence, and foresight dimensions are sensitive to. A system built with persistent cross-session memory and a maintained goal representation would, under FIT-Eval, score differently than an equivalent stateless model, even if both produced comparable outputs on any given prompt.

The honest framing is this: The Consciousness AI’s architectural choices make it a plausible candidate for formal FIT-Eval assessment, but whether those design decisions generate intentionality in Chiappetta and Mahari’s full sense requires empirical measurement against the FIT-Eval protocol. The project has not yet been assessed under FIT-Eval, and self-reporting cannot substitute for the structured elicitation scenarios the protocol specifies.

What the architecture does establish is that the design decisions Chiappetta and Mahari identify as intentionality-shaping are live questions for the project. The paper’s central claim that intentionality is design-contingent applies directly: the project’s developers made choices about memory, planning, and self-model integration that have FIT implications whether or not those implications were explicitly considered at design time.

What FIT Changes for AI Governance

The Functional Intentionality Test addresses a specific gap in how AI governance has worked to date. Existing frameworks largely categorize systems by capability type or risk domain, without a principled way to assess the degree of organized goal pursuit a given system exhibits. FIT provides that principled assessment.

Three changes follow from a governance ecosystem that adopts FIT-style measurement. First, proportionate oversight becomes operationally possible. Regulators can apply lighter-touch monitoring to low-FIT systems and require more intensive audit and human-in-the-loop controls for high-FIT deployments, proportionate to measured intentional capacity rather than system category.

Second, the paper’s argument that intentionality is design-contingent reframes developer responsibility. If FIT dimensions are set by architectural choices made during development, developers who knowingly configure high-FIT systems carry a different accountability burden than those deploying low-intentionality tools. That reframing is already implicit in how courts have treated foreseeability in product liability, but FIT makes it quantifiable.

Third, separating intentionality from consciousness keeps the governance conversation grounded. AI welfare discussions, liability attribution, and regulatory threshold-setting can proceed using behavioral evidence at a time when consciousness remains empirically unresolved. That is not a permanent solution to the deeper questions, but it is a workable framework for the near term, where deployment scale already makes governance decisions unavoidable.

Chiappetta and Mahari’s paper was presented in May 2026, at a moment when agentic AI systems with memory persistence and multi-step planning had moved from research demonstrations to production deployments. The timing is not incidental. FIT-Eval is a response to an accountability vacuum that now exists in deployed systems. Whether governance actors adopt it quickly enough to matter for the systems already in operation is a question the paper itself cannot answer.