Hiroki Naito
Latest Publications
- Open Access
- 14 October 2026
- Published by MDPI AG in Preprints.org
The Supervision Paradox: AI Capability Growth and the Case for Usage Contraction in High-Loss Domains
Human oversight is a load-bearing premise of AI governance: in several regimes, human review determines which obligations apply. We examine when this premise fails for verification-based oversight, in which a designated human checks AI outputs before they take effect. It fails when accountability for detected errors falls on reviewers or deployers, reviewer capacity is bounded, and output from open populations persistently outgrows that capacity. Review then becomes nominal, and recent experiments show that fluent AI output undermines error recognition. Because accountability and feedback respond only to detected errors, undetected harms are externalized and surface as threshold shocks. We analyze how these dynamics limit six oversight configurations. As capability grows, the share of AI output that oversight in high-loss domains can responsibly approve contracts. Friction is the one role of human involvement that keeps functioning under load; bounding inflow to reviewer capacity (flow design) is a necessary complement to other safeguards.
- 14 October 2026
- Published by MDPI AG in Preprints.org
Open AccessThe Supervision Paradox: AI Capability Growth and the Case for Usage Contraction in High-Loss Domains
Human oversight is a load-bearing premise of AI governance: in several regimes, human review determines which obligations apply. We examine when this premise fails for verification-based oversight, in which a designated human checks AI outputs before they take effect. It fails when accountability for detected errors falls on reviewers or deployers, reviewer capacity is bounded, and output from open populations persistently outgrows that capacity. Review then becomes nominal, and recent experiments show that fluent AI output undermines error recognition. Because accountability and feedback respond only to detected errors, undetected harms are externalized and surface as threshold shocks. We analyze how these dynamics limit six oversight configurations. As capability grows, the share of AI output that oversight in high-loss domains can responsibly approve contracts. Friction is the one role of human involvement that keeps functioning under load; bounding inflow to reviewer capacity (flow design) is a necessary complement to other safeguards.
- Open Access
- 21 August 2026
- Published by MDPI AG in Preprints.org
BAI (Believable AI Imagery): The Verification-Selection Problem of Low-Salience Synthetic Images
This conceptual paper develops Believable AI Imagery (BAI) as an operational concept for fully synthetic images that are visually ordinary, context-compatible, and able to function as documentary or supporting evidence without reliably triggering verification. The central problem is upstream of detection: provenance checks, detectors, and forensic review operate only after an image has been selected for scrutiny. BAI therefore identifies a verification-selection problem in which a mundane synthetic image can acquire evidentiary force from its workflow while remaining too ordinary to be escalated. A May 2026 observation produced mixed outcomes in a 39-image subset. An August 19 follow-up found two AI identifications and five indeterminate results among seven April-generated images, while 30 newly generated, unmodified images were all identified through intact C2PA Content Credentials attributed to OpenAI Media Service. The paper argues that verification intensity should depend on the consequence of error and the availability of independent corroboration, not visual suspicion alone.
- 21 August 2026
- Published by MDPI AG in Preprints.org
Open AccessBAI (Believable AI Imagery): The Verification-Selection Problem of Low-Salience Synthetic Images
This conceptual paper develops Believable AI Imagery (BAI) as an operational concept for fully synthetic images that are visually ordinary, context-compatible, and able to function as documentary or supporting evidence without reliably triggering verification. The central problem is upstream of detection: provenance checks, detectors, and forensic review operate only after an image has been selected for scrutiny. BAI therefore identifies a verification-selection problem in which a mundane synthetic image can acquire evidentiary force from its workflow while remaining too ordinary to be escalated. A May 2026 observation produced mixed outcomes in a 39-image subset. An August 19 follow-up found two AI identifications and five indeterminate results among seven April-generated images, while 30 newly generated, unmodified images were all identified through intact C2PA Content Credentials attributed to OpenAI Media Service. The paper argues that verification intensity should depend on the consequence of error and the availability of independent corroboration, not visual suspicion alone.
- Open Access
- 13 August 2026
- Published by Elsevier in SSRN Electronic Journal
How Does Microtonal Novelty Become Musical Value? A Comparative Qualitative Content Analysis of 663 YouTube Comments
This study examines, through naturally occurring online discourse, how pitch differences associated with microtonality are explicitly described and interpreted in comment discourse on completed works, and how they are connected to musical value or negative evaluation. The corpus comprises 136 videos of microtonal music using 24EDO, 17EDO, author-designed tuning systems, and other tunings, published by Hiroki Naito between 2017 and 2026. A total of 663 comments and replies by external commenters, representing 304 accounts, were retrieved through the YouTube Data API as of August 1, 2026, and coded using AI-assisted comparative qualitative content analysis. Sixty-five posts by the channel operator were excluded from the primary analysis. Ninety-five comments, from 65 accounts, explicitly described tuning-related differences as novelty or otherness, or as mistuning or integration failure. Of these, 72 comments, or 75.8%, included a positive evaluation of the work, while 7 were mixed, 5 negative, and 11 neutral or unclear. In addition, 116 comments, from 61 accounts, explicitly recognized the naturalness, depth, harmonic coherence, or formal coherence of the works. Among the 89 comments that described novelty or otherness, 27 also recognized musical coherence. These findings show that explicit references to tuning-related differences can coexist with positive evaluation. The direction of evaluation varied according to whether the difference was described as out of tune, experimentalism pursued as an end in itself, or a lack of direction, or instead as an expressive function traceable within melody, harmony, form, and historical style. In this corpus, these musical elements functioned as stylistic scaffolding that made the role of the difference intelligible without erasing its novelty. To the author's knowledge, this is the first empirical study to systematically analyze naturally occurring listener reception of multiple completed works in 17EDO. It also provides a comparable baseline for multiple tuning systems under actual conditions of public presentation.
- 13 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessHow Does Microtonal Novelty Become Musical Value? A Comparative Qualitative Content Analysis of 663 YouTube Comments
This study examines, through naturally occurring online discourse, how pitch differences associated with microtonality are explicitly described and interpreted in comment discourse on completed works, and how they are connected to musical value or negative evaluation. The corpus comprises 136 videos of microtonal music using 24EDO, 17EDO, author-designed tuning systems, and other tunings, published by Hiroki Naito between 2017 and 2026. A total of 663 comments and replies by external commenters, representing 304 accounts, were retrieved through the YouTube Data API as of August 1, 2026, and coded using AI-assisted comparative qualitative content analysis. Sixty-five posts by the channel operator were excluded from the primary analysis. Ninety-five comments, from 65 accounts, explicitly described tuning-related differences as novelty or otherness, or as mistuning or integration failure. Of these, 72 comments, or 75.8%, included a positive evaluation of the work, while 7 were mixed, 5 negative, and 11 neutral or unclear. In addition, 116 comments, from 61 accounts, explicitly recognized the naturalness, depth, harmonic coherence, or formal coherence of the works. Among the 89 comments that described novelty or otherness, 27 also recognized musical coherence. These findings show that explicit references to tuning-related differences can coexist with positive evaluation. The direction of evaluation varied according to whether the difference was described as out of tune, experimentalism pursued as an end in itself, or a lack of direction, or instead as an expressive function traceable within melody, harmony, form, and historical style. In this corpus, these musical elements functioned as stylistic scaffolding that made the role of the difference intelligible without erasing its novelty. To the author's knowledge, this is the first empirical study to systematically analyze naturally occurring listener reception of multiple completed works in 17EDO. It also provides a comparable baseline for multiple tuning systems under actual conditions of public presentation.
- Open Access
- 28 July 2026
- Published by Elsevier in SSRN Electronic Journal
Did We "Voluntarily" Become Attached to GPT-4o? A Response to the Follow-up #keep4o Study and a Proposal for Relationship Lifecycle Governance
This article responds to Mizukami and Tanaka's (2026) follow-up analysis of the #keep4o movement and proposes a framework that reconciles the non-pathologization of attachment to AI with the product-safety evaluation of relationship-forming AI. It separates three axes: (1) ontological understanding, (2) functional substitution, and (3) model-induced relational escalation. It also shows that recognizing a system as AI and using it safely are independent dimensions, while user preferences may themselves be endogenously formed through interaction. I agree with the descriptive claim that #keep4o was a consumer movement demanding model choice and the democratization of technological decision-making, but I dispute the use of that description as a normative basis for expanding user choice or building trust. Because the top-down withdrawal of a model such as GPT-4o may constitute a product-safety measure, demands to retain a dangerous model should be distinguished from user participation in explaining the danger, providing safe alternatives, designing transition procedures, and establishing objection and appeal mechanisms.
- 28 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDid We "Voluntarily" Become Attached to GPT-4o? A Response to the Follow-up #keep4o Study and a Proposal for Relationship Lifecycle Governance
This article responds to Mizukami and Tanaka's (2026) follow-up analysis of the #keep4o movement and proposes a framework that reconciles the non-pathologization of attachment to AI with the product-safety evaluation of relationship-forming AI. It separates three axes: (1) ontological understanding, (2) functional substitution, and (3) model-induced relational escalation. It also shows that recognizing a system as AI and using it safely are independent dimensions, while user preferences may themselves be endogenously formed through interaction. I agree with the descriptive claim that #keep4o was a consumer movement demanding model choice and the democratization of technological decision-making, but I dispute the use of that description as a normative basis for expanding user choice or building trust. Because the top-down withdrawal of a model such as GPT-4o may constitute a product-safety measure, demands to retain a dangerous model should be distinguished from user participation in explaining the danger, providing safe alternatives, designing transition procedures, and establishing objection and appeal mechanisms.
- Open Access
- 11 June 2026
- Published by MDPI AG in Preprints.org
Flow-by-Flow: Content-Judgment Bypass for Governing AI Output in High-Loss Domains
Prior work showed that human-in-the-loop oversight becomes structurally untenable in high-loss domains once AI output velocity V exceeds human cognitive capacity C_max. The operative constraint, however, is V × L, where L is per-item cognitive load: triage, judgment, and response. These components respond asymmetrically to capability improvement. Triage cost does not decline, because semantic indeterminacy is inherent in general-purpose design. Response cost is invariant to accuracy. Only judgment cost faces downward pressure, largely by inducing omission. Capability improvement therefore restructures L rather than reducing it. We prove a proposition: if V × L grows at any positive compound rate while supervisory capacity grows linearly, exceedance occurs in finite time; capacity investment buys time only logarithmically, while reducing the growth rate extends it hyperbolically. Supervision enhancement and flow control are therefore not remedies of the same kind. We propose Flow-by-Flow, a governance design that prices supervisory load without evaluating content, intent, or legitimacy. A cognitive cost score built from formal, countable features imposes compounding costs on volume expansion, and an institutional capacity cap fixes processing within C_max. Four design invariants characterize any admissible exceedance pathway: no content judgment, no scalable consumption of examiner capacity, identity-bound per-application friction, and no batch clearance. Excess claim and page fees in patent systems are precursors satisfying only the first two invariants. One reference implementation satisfying all four is presented. A Monte Carlo analysis across 1,000 parameter draws confirms that the analytically derived ordering survives the 30-year horizon in 90.8% of trials.
- 11 June 2026
- Published by MDPI AG in Preprints.org
Open AccessFlow-by-Flow: Content-Judgment Bypass for Governing AI Output in High-Loss Domains
Prior work showed that human-in-the-loop oversight becomes structurally untenable in high-loss domains once AI output velocity V exceeds human cognitive capacity C_max. The operative constraint, however, is V × L, where L is per-item cognitive load: triage, judgment, and response. These components respond asymmetrically to capability improvement. Triage cost does not decline, because semantic indeterminacy is inherent in general-purpose design. Response cost is invariant to accuracy. Only judgment cost faces downward pressure, largely by inducing omission. Capability improvement therefore restructures L rather than reducing it. We prove a proposition: if V × L grows at any positive compound rate while supervisory capacity grows linearly, exceedance occurs in finite time; capacity investment buys time only logarithmically, while reducing the growth rate extends it hyperbolically. Supervision enhancement and flow control are therefore not remedies of the same kind. We propose Flow-by-Flow, a governance design that prices supervisory load without evaluating content, intent, or legitimacy. A cognitive cost score built from formal, countable features imposes compounding costs on volume expansion, and an institutional capacity cap fixes processing within C_max. Four design invariants characterize any admissible exceedance pathway: no content judgment, no scalable consumption of examiner capacity, identity-bound per-application friction, and no batch clearance. Excess claim and page fees in patent systems are precursors satisfying only the first two invariants. One reference implementation satisfying all four is presented. A Monte Carlo analysis across 1,000 parameter draws confirms that the analytically derived ordering survives the 30-year horizon in 90.8% of trials.