arxiv: 2605.05475 · v1 · submitted 2026-05-06 · 💻 cs.AI

Recognition: unknown

Intentionality is a Design Decision: Measuring Functional Intentionality for Accountable AI Systems

Allessia Chiappetta , Robert Mahari

Authors on Pith no claims yet

Pith reviewed 2026-05-08 16:22 UTC · model grok-4.3

classification 💻 cs.AI

keywords intentionalityAI accountabilityfunctional intentionality testagentic systemsdesign-contingent behaviorautonomy calibrationoversightbehavioral profile

0 comments

The pith

Intentionality in AI is a controllable design choice defined by five observable behaviors.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that intentionality in AI should be treated as a behavioral profile made up of purpose, foresight, volition, temporal commitment, and coherence rather than any form of inner experience. These traits arise from concrete design decisions about memory, planning depth, and tool use, so they can be increased or limited by changing the architecture. To make this practical, the authors present the Functional Intentionality Test (FIT) that scores systems on the five dimensions and pairs it with the FIT-Eval protocol for structured assessment. The result is a way to match the level of oversight to how intentional a system appears, addressing the accountability challenges that come with more autonomous, long-horizon AI. If the claim holds, designers gain a lever to set intentionality levels deliberately instead of discovering them after deployment.

Core claim

Intentionality is defined not as consciousness but as a behavioral profile characterized by purpose, foresight, volition, temporal commitment, and coherence. These properties are design-contingent because architectural choices such as memory persistence, planning depth, and tool autonomy shape the degree to which systems exhibit organized goal pursuit. The Functional Intentionality Test (FIT) quantifies intentional-like behavior across the five observable dimensions, and FIT-Eval provides a structured evaluation protocol for eliciting and scoring them. This enables proportionate oversight and deliberate autonomy calibration in agentic systems.

What carries the argument

The Functional Intentionality Test (FIT), a multidimensional scoring framework that turns the five behavioral dimensions of purpose, foresight, volition, temporal commitment, and coherence into quantifiable levels for design control.

If this is right

Architectural decisions like memory persistence and tool autonomy directly influence the degree of intentionality exhibited by AI systems.
Rising intentional capacity in AI heightens accountability risks, necessitating more structured oversight.
Reduced human agency through intentional AI can improve efficiency but requires deliberate calibration of autonomy levels.
FIT provides interpretable levels of intentionality that enable proportionate governance for different systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Regulators could adopt FIT scores as a basis for tiered approval requirements for agentic AI in high-stakes domains.
Developers might deliberately constrain specific dimensions such as foresight in certain applications to keep overall intentionality low.
Applying FIT to existing planning-based agents would test whether current systems already produce high intentionality scores without targeted design.
The approach suggests research into trade-off curves showing how changes in one dimension affect measurable performance and risk.

Load-bearing premise

The five listed behavioral properties are both necessary and sufficient proxies for intentionality, are reliably observable from external behavior alone, and can be controlled through design choices without reference to internal states.

What would settle it

A side-by-side comparison in which two otherwise identical systems differ only in one design variable such as memory persistence, yet the higher-memory version shows no increase in temporal commitment or coherence on FIT-Eval tasks.

read the original abstract

As AI systems increasingly exhibit autonomous, goal-directed, and long-horizon behavior, users lack a standardized way to detect the degree to which a system functions like an intentional actor for governance and accountability purposes. This position paper defines intentionality not as consciousness, but as a behavioral profile characterized by purpose, foresight, volition, temporal commitment, and coherence - criteria long used in legal and philosophical contexts to infer intent. These properties are design-contingent: architectural choices such as memory persistence, planning depth, and tool autonomy shape the degree to which systems exhibit organized goal pursuit. If intentionality is design-contingent, it is in principle controllable. Yet control requires measurement. We introduce the Functional Intentionality Test (FIT), a multidimensional framework that quantifies intentional-like behavior across five observable dimensions, and propose FIT-Eval, a structured evaluation protocol for eliciting and scoring them. While reduced human agency can increase efficiency, rising intentional capacity heightens accountability risks. By translating intentionality into interpretable levels, FIT enables proportionate oversight and deliberate autonomy calibration in increasingly agentic systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This position paper defines a five-dimension behavioral test for AI intentionality but stays purely conceptual with no validation, examples, or scoring details.

read the letter

The paper's main move is to treat intentionality as a design-contingent behavioral profile built from five traits—purpose, foresight, volition, temporal commitment, and coherence—drawn from legal and philosophical sources. It then names the Functional Intentionality Test (FIT) and a protocol called FIT-Eval to quantify those traits for accountability in agentic systems. That mapping from architecture choices like memory persistence and planning depth to observable behavior is the clearest new element here. It gives people working on governance a structured way to talk about calibrating autonomy without invoking consciousness. The abstract also flags the practical tension between efficiency gains from reduced human oversight and rising accountability risks, which is a fair point to surface. The stress-test concern holds up: the five properties are presented as both the definition of intentionality and the things FIT measures, so the framework is circular by design. Nothing in the provided text shows why these traits are necessary and sufficient, how they can be scored reliably from external traces alone, or how to separate explicit design effects from emergent scale effects. No rubric, no worked examples, and no discussion of edge cases appear. This leaves the proposal at the level of a useful taxonomy rather than a testable measurement tool. Readers in AI ethics or policy circles who want language for proportionate oversight might pull ideas from it for discussion. It is not yet ready for citation as a working method. A serious editor should send it for peer review as a position paper, with the clear expectation that referees will require concrete implementation details and some argument on observability before it advances further.

Referee Report

3 major / 3 minor

Summary. The paper claims that intentionality in AI is not consciousness but a design-contingent behavioral profile defined by five properties (purpose, foresight, volition, temporal commitment, and coherence) drawn from legal and philosophical contexts. It introduces the Functional Intentionality Test (FIT) as a multidimensional framework to quantify these observable dimensions and proposes FIT-Eval as a structured evaluation protocol, arguing that this enables proportionate oversight and deliberate autonomy calibration in agentic systems.

Significance. If the five dimensions can be operationalized into reliable, externally observable metrics that support accountability inferences, the framework could inform governance of increasingly autonomous AI agents by linking architectural choices to measurable behavioral profiles. The position highlights a key trade-off between efficiency from reduced human oversight and heightened risks from intentional capacity, offering a conceptual bridge between design decisions and regulatory needs.

major comments (3)

[Definition of intentionality (abstract and opening sections)] The definition of intentionality as precisely the five properties that FIT then quantifies (purpose, foresight, volition, temporal commitment, coherence) creates circularity: the test operationalizes its own stipulative input without external benchmarks, independent validation criteria, or falsifiability conditions. This is load-bearing for the central claim that FIT provides a reliable measurement tool for accountability.
[FIT-Eval protocol description] No concrete scoring rubric, example traces from specific AI systems, inter-rater reliability discussion, or error analysis is supplied for FIT-Eval, leaving the protocol as a high-level outline rather than an implementable method. This undermines the utility claim for 'quantifying' dimensions to enable 'proportionate oversight'.
[Design-contingency and architectural choices discussion] The design-contingency argument (architectural choices such as memory persistence and planning depth shape the properties) does not address how to distinguish these from emergent long-horizon coherence arising from scale alone, nor does it provide a causal mapping showing controllability without reference to internal states. This challenges the controllability premise required for accountability calibration.

minor comments (3)

[Abstract] The abstract should explicitly note that this is a conceptual position paper without empirical data, formal derivations, or validation examples.
Additional references to prior work on functionalism in philosophy of mind, AI ethics frameworks for intent attribution, and existing behavioral evaluation protocols in agentic systems would strengthen context.
Clarify operational distinctions between 'volition' and 'purpose' when applied to observable AI behavior traces to avoid overlap in measurement.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their insightful comments, which highlight important areas for strengthening the conceptual and practical aspects of our position paper. We address each major comment point by point below, indicating planned revisions where appropriate.

read point-by-point responses

Referee: [Definition of intentionality (abstract and opening sections)] The definition of intentionality as precisely the five properties that FIT then quantifies (purpose, foresight, volition, temporal commitment, coherence) creates circularity: the test operationalizes its own stipulative input without external benchmarks, independent validation criteria, or falsifiability conditions. This is load-bearing for the central claim that FIT provides a reliable measurement tool for accountability.

Authors: We acknowledge the stipulative nature of the definition and the resulting self-referential structure in the framework. The five properties are not invented for FIT but are drawn directly from established legal and philosophical criteria used to attribute intent in non-conscious actors (e.g., corporate liability and action theory). This grounding provides external motivation independent of the test itself. To mitigate the concern, we will add a new subsection clarifying the literature sources for each property, outlining potential validation pathways (such as alignment with human expert ratings of intent or legal precedent analysis), and specifying falsifiability conditions through inconsistent predictions in applied settings. As a position paper, full empirical validation lies outside the current scope but can be explicitly signposted. revision: partial
Referee: [FIT-Eval protocol description] No concrete scoring rubric, example traces from specific AI systems, inter-rater reliability discussion, or error analysis is supplied for FIT-Eval, leaving the protocol as a high-level outline rather than an implementable method. This undermines the utility claim for 'quantifying' dimensions to enable 'proportionate oversight'.

Authors: We agree that the FIT-Eval description is currently high-level and requires concrete details to support implementability claims. In the revised manuscript, we will expand the protocol section to include: (1) a sample scoring rubric with behavioral indicators and ordinal scales for each dimension, (2) illustrative evaluation traces from at least two contrasting systems (e.g., a standard LLM versus one augmented with persistent memory and explicit planning), and (3) a brief discussion of inter-rater reliability considerations and common error sources. These additions will better substantiate the utility for proportionate oversight. revision: yes
Referee: [Design-contingency and architectural choices discussion] The design-contingency argument (architectural choices such as memory persistence and planning depth shape the properties) does not address how to distinguish these from emergent long-horizon coherence arising from scale alone, nor does it provide a causal mapping showing controllability without reference to internal states. This challenges the controllability premise required for accountability calibration.

Authors: The paper's core claim is that intentionality is observable via behavioral profiles, making controllability inferable from design interventions that predictably alter those profiles. We will revise the relevant section to explicitly contrast scale-driven emergence (e.g., coherence from larger context windows) with targeted architectural features (e.g., dedicated memory modules that enhance temporal commitment independently of model size), using comparative examples. For causal mapping, we will emphasize that FIT operates on external behavior, so controllability is demonstrated through observable score changes following design modifications rather than internal access. We note that rigorous causal experiments would strengthen this but are beyond the position paper's scope; the revision will clarify this boundary. revision: partial

Circularity Check

1 steps flagged

Intentionality defined via five properties; FIT quantifies exactly those properties by construction

specific steps

self definitional [Abstract]
"This position paper defines intentionality not as consciousness, but as a behavioral profile characterized by purpose, foresight, volition, temporal commitment, and coherence ... We introduce the Functional Intentionality Test (FIT), a multidimensional framework that quantifies intentional-like behavior across five observable dimensions"

Intentionality is defined using precisely the five properties; FIT is then presented as quantifying behavior across those identical five dimensions. The measurement therefore reduces directly to the input definition without independent criteria, external validation, or derivation that could falsify or extend the original characterization.

full rationale

The paper's core move is to define intentionality as a behavioral profile consisting of purpose, foresight, volition, temporal commitment, and coherence, then introduce FIT as the framework that quantifies intentional-like behavior across those same five dimensions. This makes the test an operationalization of its own definitional inputs rather than an independent derivation or externally validated measure. No equations, formal mappings, or external benchmarks are supplied to break the equivalence. The position paper therefore exhibits self-definitional circularity at the level of its central claim.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The proposal rests on a domain assumption that behavioral traits suffice as proxies for intentionality and introduces FIT as a new construct without independent empirical grounding or prior validation.

axioms (1)

domain assumption Intentionality can be inferred from observable behavioral properties (purpose, foresight, volition, temporal commitment, coherence) without reference to internal mental states or consciousness.
Explicitly stated as the definitional basis in the abstract.

invented entities (1)

Functional Intentionality Test (FIT) no independent evidence
purpose: To quantify intentional-like behavior across five dimensions for accountability purposes.
New framework and scoring construct proposed by the authors.

pith-pipeline@v0.9.0 · 5491 in / 1322 out tokens · 40055 ms · 2026-05-08T16:22:55.420901+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

39 extracted references · 9 canonical work pages · 1 internal anchor

[1]

American Law Institute (ALI). 1962. Model Penal Code § 2.02(2). 5

1962
[2]

Daniele Amoroso and Guglielmo Tamburrini. 2020. Autonomous weapons systems and meaningful human control: ethical and legal issues.Current Robotics Reports1, 4 (2020), 187–194

2020
[3]

1963.Intention(2nd ed.)

Elizabeth Anscombe. 1963.Intention(2nd ed.). Blackwell, Oxford

1963
[4]

Sergio MC Avila Negri. 2021. Robot as legal person: Electronic personhood in robotics and artificial intelligence.Frontiers in Robotics and AI8 (2021), 789327

2021
[5]

Ian Ayres and Jack M Balkin. 2024. The Law of AI is the Law of Risky Agents without Intentions.U. Chi. L. Rev. Online(2024), 1

2024
[6]

Joffrey Baeyaert. 2025. Beyond Personhood: The Evolution of Legal Personhood and Its Implications for AI Recognition.Technology and Regulation 2025 (2025), 355–386

2025
[7]

Yavar Bathaee. 2018. The Artificial Intelligence Black Box and the Failure of Intent and Causation.Harvard Journal of Law & Technology31 (2018). https://jolt.law.harvard.edu/assets/articlePDFs/v31/The-Artificial-Intelligence-Black-Box-and-the-Failure-of-Intent-and-Causation-Yavar- Bathaee.pdf

2018
[8]

Tim Bayne et al. 2024. Tests for Consciousness in Humans and Beyond.Trends in Cognitive Sciences28, 6 (2024). doi:10.1016/j.tics.2024.01.010

work page doi:10.1016/j.tics.2024.01.010 2024
[9]

Alexander Blanchard, Claudio Novelli, Luciano Floridi, and Mariarosaria Taddeo. 2025. A Risk-Based Regulatory Approach to Autonomous Weapon Systems.Digital Society4, 1 (2025), 23

2025
[10]

Andreea Bodnari and John Travis. 2025. Scaling enterprise AI in healthcare: the role of governance in risk mitigation frameworks.npj Digital Medicine8, 1 (2025), 272

2025
[11]

1987.Intention, Plans, and Practical Reason

Michael Bratman. 1987.Intention, Plans, and Practical Reason. Harvard University Press, Cambridge, Massachusetts

1987
[12]

Madalina Busuioc. 2022. AI algorithmic oversight: new frontiers in regulation. InHandbook of regulatory authorities. Edward Elgar Publishing, 470–486

2022
[13]

Consciousness in Artificial Intelligence: Insights from the Science of Consciousness

Patrick Butlin, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, Axel Constant, George Deane, Stephen M. Fleming, Chris Frith, Xu Ji, Ryota Kanai, Colin Klein, Grace Lindsay, Matthias Michel, Liad Mudrik, Megan A. K. Peters, Eric Schwitzgebel, Jonathan Simon, and Rufin VanRullen. 2023. Consciousness in Artificial Intelligence: Insights from the...

work page internal anchor Pith review arXiv 2023
[14]

Ishani Chakrabarty and Aditi Baral. 2023. Artificial Intelligence and Personhood in the 21st Century.Indian J. Integrated Rsch. L.3 (2023), 1

2023
[15]

Jie Chen, Jian Sun, and Gang Wang. 2022. From unmanned systems to autonomous intelligent systems.Engineering12 (2022), 16–19

2022
[16]

Clawdbot. 2025. Clawdbot: Persistent AI Assistant. https://clawd-bot.com/. Accessed 13 February 2026

2025
[17]

2022.Intent

Cornell Law School. 2022.Intent. https://www.law.cornell.edu/wex/intent Accessed: 2025-02-01

2022
[18]

Dan B. Dobbs. 2000.The Law of Torts. West Group, St. Paul, Minnesota

2000
[19]

Michele Farisco, Kathinka Evers, and Jean-Pierre Changeux. 2024. Is Artificial Consciousness Achievable? Lessons from the Human Brain.Neural Networks(2024). https://www.sciencedirect.com/science/article/pii/S0893608024006385 Advance online publication

2024
[20]

Katherine B Forrest. 2023. The ethics and challenges of legal personhood for AI.Yale LJF133 (2023), 1175

2023
[21]

Simon Goldstein. 2016. A Preface Paradox for Intention.Philosophers’ Imprint16 (2016), 1–20. http://hdl.handle.net/2027/spo.3521354.0016.014

work page arXiv 2016
[22]

Goodenough

Oliver R. Goodenough. 2004. Responsibility and Punishment: Whose Mind? A Response.Philosophical Transactions of the Royal Society of London B: Biological Sciences359, 1451 (2004), 1805–1811. doi:10.1098/rstb.2004.1548

work page doi:10.1098/rstb.2004.1548 2004
[23]

Daniel Harris. 2022. Corporate Intent and the Concept of Agency.Stan. JL Bus. & Fin.27 (2022), 133

2022
[24]

Gregory Klass. 2009. Intent to contract.Va. L. Rev.95 (2009), 1437

2009
[25]

Thomas Kosch and Sebastian Feger. 2024. Risk or Chance? Large Language Models and Reproducibility in Human–Computer Interaction Research. arXiv preprint arXiv:2404.15782(2024). https://arxiv.org/abs/2404.15782v1

work page arXiv 2024
[26]

Rachel Ma, Jingyi Qu, Andreea Bobu, and Dylan Hadfield-Menell. 2024. Goal Inference from Open-Ended Dialog.arXiv preprint arXiv:2410.13957 (2024). https://arxiv.org/abs/2410.13957v1

work page arXiv 2024
[27]

Carl Macrae. 2025. Managing risk and resilience in autonomous and intelligent systems: Exploring safety in the development, deployment, and use of artificial intelligence in healthcare.Risk Analysis45, 4 (2025), 910–927

2025
[28]

Bertram F Malle and Sarah E Nelson. 2003. Judging mens rea: The tension between folk concepts and legal concepts of intentionality.Behavioral sciences & the law21, 5 (2003), 563–580

2003
[29]

Mizanur Rahman, Amran Bhuiyan, Mohammed Saidul Islam, Md Tahmid Rahman Laskar, Ridwan Mahbub, Ahmed Masry, Shafiq Joty, and Enamul Hoque. 2025. LLM-Based Data Science Agents: A Survey of Capabilities, Challenges, and Future Directions.arXiv preprint arXiv:2510.04023(2025). https://arxiv.org/abs/2510.04023v1

work page arXiv 2025
[30]

Patrick Raschner. 2022. Supervisory oversight of the use of AI and ml by financial market participants. InDigitalisation, Sustainability, and the Banking and Capital Markets Union: Thoughts on Current Issues of EU Financial Regulation. Springer, 99–123

2022
[31]

Sandeep Reddy, Sonia Allan, Simon Coghlan, and Paul Cooper. 2020. A governance model for the application of AI in health care.Journal of the American medical informatics association27, 3 (2020), 491–497

2020
[32]

Martin Shapiro. 1985. The Deconstruction and Reconstruction of Intent.Criminal Justice: Nomos XXVII(1985)

1985
[33]

Sahara Shrestha. 2021. Nature, Nurture, or Neither?: Liability for Automated and Autonomous Artificial Intelligence Torts Based on Human Design and Influences.Geo. Mason L. Rev.29 (2021), 375

2021
[34]

Danielle Swanepoel and Daniel Corks. 2024. Artificial intelligence and agency: Tie-breaking in AI decision-making.Science and Engineering Ethics 30, 2 (2024), 11. 6

2024
[35]

David Velleman

J. David Velleman. 1989.Practical Reflection. Princeton University Press, Princeton, NJ

1989
[36]

Francis Rhys Ward, Matt MacDermott, Francesco Belardinelli, Francesca Toni, and Tom Everitt. 2024. The reasons that agents act: Intention and instrumental goals.arXiv preprint arXiv:2402.07221(2024)

work page arXiv 2024
[37]

Tiefeng Wu, Miao Wang, Yunfang Xi, and Zhichao Zhao. 2024. Intent Recognition Model Based on Sequential Information and Sentence Features. Neurocomputing566 (2024), 127054. doi:10.1016/j.neucom.2023.127054

work page doi:10.1016/j.neucom.2023.127054 2024
[38]

Rui Zhang, Christopher Flathmann, Geoff Musick, Beau Schelble, Nathan J McNeese, Bart Knijnenburg, and Wen Duan. 2024. I know this looks bad, but I can explain: Understanding when AI should explain actions in human-AI teams.ACM Transactions on Interactive Intelligent Systems14, 1 (2024), 1–23

2024
[39]

Fox Harrell

Jichen Zhu and D. Fox Harrell. 2009. System Intentionality and the Artificial Intelligence Hermeneutic Network: The Role of Intentional Vocabulary. InDigital Arts and Culture Conference Proceedings. Irvine, CA. https://escholarship.org/uc/item/3rd2s695 December 2009. A Appendix: Methodological Framework and Evaluation Protocol (FIT-Eval) The Functional In...

2009