Governance for Sensory AI: Regulation, Ethics & the EU AI Act

Governance for Sensory AI: Treat Models Like Instruments

Sensory AI should be governed as carefully as any lab instrument:

This is not bureaucracy; it is how we keep models scientific and trustworthy.

The EU AI Act: Regulatory Implications for Sensory Science

The regulatory environment is now formalizing these principles. The EU AI Act, entering full implementation on August 2, 2026, introduces a four-tier risk classification system with direct consequences for sensory AI. Systems that use biometric data (including physiological responses to olfactory or gustatory stimuli) for identification or categorization fall into the “High-Risk” category, requiring pre-market conformity assessments, mandatory registration in the EU database, and continuous post-market monitoring. The forthcoming “Digital Omnibus” proposals will further clarify GDPR obligations for AI systems, making it clear that reliance on “legitimate interests” for processing biometric data faces heightened scrutiny.

Companies must now implement “AI Literacy” obligations for employees and design systems with human oversight as a core requirement. For sensory science practitioners, this means that biometric-informed research tools (EEG headsets at retail counters, GSR during panel sessions, heart-rate-based preference models) must be governed under a clear regulatory framework. Notably, the AI Act also permits the processing of special-category personal data to uncover and rectify bias, provided strict safeguards are in place, a provision that supports efforts to make sensory models more inclusive and representative of diverse populations.

Persistent AI Agents and Pragmatic Governance

The pragmatic personhood framework extends naturally into governance. As AI agents become persistent participants in formulation pipelines, maintaining state, remembering past evaluations, and planning across product-development horizons, the governance toolkit must evolve beyond treating them as passive instruments. Assigning clear “bundles” of obligations (mandate adherence, transparency, systemic non-harm) to these agents, and establishing human stop-rules that can “arrest” an agent's operations when thresholds are breached, provides a practical mechanism for accountability that does not require resolving intractable debates about machine consciousness.

Responsibilities and Risks: Design for Flourishing, Not Just Acceptance

Optimization has a dark side. If we optimize only for short-run liking, we risk hyper-palatable cul-de-sacs and homogenization of taste, eroding cultural variety. Sensory scientists should:

Generative Ghosts and Synthetic Provenance

The rise of “Generative Ghosts” (AI agents designed to represent deceased individuals using LLMs and multimodal synthesis) sharpens this responsibility in unexpected ways. Research by Morris and Brubaker (2025) identifies a design space in which such agents can generate novel content, remember past interactions, and plan future actions, blurring the boundary between archival legacy and dynamic evolution.

For sensory science, the parallel is instructive: when AI systems generate “synthetic provenance” for products (fabricated origin stories, AI-authored tasting notes presented as human expertise, or digital replicas of heritage processes), they risk what hauntological theory calls a “false ontology”: collapsing a complex relationship between artifact and history into a misleading narrative of authenticity.

The “second death” problem identified in generative-ghost research, where the platform hosting a digital legacy goes bankrupt, leaving survivors to grieve anew, has a commercial analogue: brands that build consumer rituals around AI-generated provenance narratives are vulnerable to a “second loss” of trust if the synthetic origins are exposed. Sensory scientists, as stewards of the link between cues and meaning, have a responsibility to advocate for “thanatosensitive” design principles: clear consent, transparent sourcing of narratives, and the humility to distinguish between what the data show and what the model imagines.

Champion Open Ontologies

Champion open ontologies. Shared descriptors and data hygiene will outcompete proprietary vagueness.

LeCun's argument for open-source AI reinforces this point from a structural angle: allowing any single commercial entity to control the “repository of all human knowledge” encoded in foundation models would be a strategic and democratic mistake. For our field, this translates into championing open sensory ontologies, shared descriptor lexicons, and transparent data practices. Open-source world models, trained on diverse sensory data and accessible to researchers across cultures, are more likely to preserve the rich variety of human taste than proprietary systems optimized for a narrow demographic. Openness accelerates progress, fosters diversity, and serves as a hedge against the homogenization of taste that closed systems risk producing.