Language shapes AI policy more than technology, and misframed terminology undermines effective governance.
When Naming Becomes a Constraint
The very act of labeling a technology can pre-empt the range of solutions we consider; the term “artificial intelligence” conjures images of autonomous agents, while the underlying systems remain narrow, data-driven pipelines. This mismatch is not merely rhetorical—it directs legislative drafts, funding allocations, and public expectations toward a mythic horizon that never materializes. Lianne Potter’s recent analysis of cybersecurity terminology illustrates how a twenty-page treatise can reveal that the words we choose embed power structures, privilege certain actors, and silence dissenting perspectives. As she notes:
“When a field adopts a catch-all label, it creates a veneer of completeness that masks the fragmented reality beneath; policymakers then legislate against a phantom rather than the concrete mechanisms that actually operate.” – Lianne Potter, author of Naming is Framing: How Cybersecurity’s Language Problems are Repeating in AI Governance
In AI governance, the same pattern repeats: “AI” becomes a catch-all, erasing distinctions between narrow-task models, generative systems, and emergent autonomous agents. The consequence is a legislative focus on “AI safety” that often translates into check-boxes for algorithmic audits, while overlooking the sociotechnical scaffolding—data provenance, labor practices, and institutional incentives—that truly determine outcomes. The result is a governance architecture that appears comprehensive on paper but fails to address the lived complexities of deployment.
The Cognitive Offloading Paradox in Policy Discourse
AI Governance Fails to Translate Ethics Photo: pexels
Framing AI as a cognitive offloading tool—an efficiency booster that frees humans from routine mental labor—has become a dominant narrative in policy circles. This narrative, however, carries an implicit assumption: that the relinquished mental work is expendable and that the system’s fallibility is an acceptable trade-off. The paradox emerges when the very act of offloading creates dependency loops, eroding the capacity for critical oversight.
This narrative, however, carries an implicit assumption: that the relinquished mental work is expendable and that the system’s fallibility is an acceptable trade-off.
Zhipu's founder, Tang Jie, advocates for open access to AI technology, emphasizing its importance for safety and innovation in a rapidly evolving landscape.
Consider the systematic literature review on AI governance that occupies pages 3265–3279 of a leading journal; its breadth underscores how pervasive the offloading frame has become across disciplines. The review catalogues dozens of policy proposals that prioritize “automation readiness” over “human-in-the-loop resilience.” By embedding efficiency as the primary metric, regulators risk codifying a bias toward speed and cost reduction, while sidelining accountability mechanisms that require human judgment.
Our analysis suggests that the offloading frame also reshapes educational curricula, prompting institutions to prioritize tool proficiency over ethical reasoning. Graduates enter the workforce equipped to deploy models but lack the vocabulary to interrogate the underlying value judgments embedded in those models. The linguistic shortcut—“AI does the thinking”—therefore becomes a structural blind spot, allowing systemic risks to proliferate unchecked.
Bridging the Linguistic Governance Gap
If language can both illuminate and obscure, the challenge for AI governance is to construct a lexicon that reflects nuance without sacrificing clarity. The “Linguistic Governance Gap” describes the space between the aspirational language of policy drafts and the technical realities engineers confront daily. Closing this gap requires a two-pronged approach: first, a disciplined audit of terminology used in statutes, standards, and corporate policies; second, the co-creation of a living glossary that evolves alongside technological advances.
Such an audit would reveal, for instance, that the term “autonomous system” appears in over thirty legislative proposals yet is rarely defined beyond “capable of independent decision-making.” Without a shared definition, courts interpret the term inconsistently, leading to fragmented jurisprudence. A living glossary, maintained by a consortium of scholars, industry leaders, and civil-society representatives, could embed contextual footnotes that tie each term to concrete system characteristics—model architecture, training data scope, and intended use cases.
Moreover, the glossary should be coupled with a “framing impact assessment” akin to environmental impact statements. Before a policy adopts a new term, its potential to shape public perception, allocate resources, and set regulatory thresholds would be evaluated. This procedural safeguard would transform framing from an afterthought into a design consideration, ensuring that language serves as a conduit for precision rather than a source of distortion.
Our Path Forward: Reframing without Reducing
AI Governance Fails to Translate Ethics Photo: unsplash
We believe that the remedy lies not in abandoning terminology altogether but in cultivating a meta-awareness of its power. Our view is that every policy draft should begin with a brief “lexical justification” that explains why a particular label was chosen, what alternatives were considered, and how the chosen term aligns with the system’s technical profile. Such a practice would force drafters to confront the implicit assumptions that often go unquestioned.
Moreover, the glossary should be coupled with a “framing impact assessment” akin to environmental impact statements.
Additionally, interdisciplinary training programs must embed linguistic analysis alongside technical instruction. By exposing engineers to the work of scholars like Lianne Potter and Tifany Petricini, future developers will acquire the tools to critique the frames that shape their design choices. This cultural shift, though subtle, can ripple outward, prompting regulators to demand more granular disclosures and encouraging the public to engage with AI debates on a more informed footing.
In sum, the limits of language in AI governance are not an insurmountable barrier but a design flaw that can be rectified through deliberate, systematic interventions. By recognizing framing as a policy lever and by institutionalizing mechanisms to audit and refine our lexical choices, we can steer AI governance toward a future where regulation is anchored in technical reality rather than linguistic illusion.
The stakes are clear: without a disciplined approach to naming, AI governance will continue to chase shadows, leaving society vulnerable to the very risks it seeks to mitigate.
AI & Technology
AI & Technology
