- The paper argues that the undifferentiated use of 'AI' obscures essential operational, ethical, and technical distinctions in military systems.
- It introduces a comprehensive taxonomy that categorizes systems like decision support, autonomous weapons, and process optimization tools with context-specific risks.
- The analysis advocates for precise, system-level terminology in research and policy to enable more accurate risk assessments and regulatory approaches.
Revisiting the Semantics of “AI”: Toward Greater Specificity in Technical, Ethical, and Regulatory Discourse
Introduction
The paper "Stop Saying 'AI'" (2602.17729) rigorously problematizes the continued use of “artificial intelligence” (AI) as a monolithic term across technical, scholarly, regulatory, and operational contexts, with a specific focus on the military domain. The authors present a comprehensive taxonomy of systems typically grouped under the “AI” umbrella—an umbrella they argue is both overbroad and operationally misleading. By tracing concrete deployments (decision support, autonomous weapon systems, logistical optimization, bureaucratic streamlining, and more) and their unique challenges, the work underscores that ethical, legal, and practical risks are highly system- and context-specific. The principal thesis urges researchers, practitioners, and policymakers to abandon the indiscriminate invocation of “AI” and instead employ precise system-level terminology and risk typologies, mirroring the approach taken with other general-purpose enabling technologies.
Limitations of the “AI” Construct and the Need for a Systemic Taxonomy
The paper notes that “AI” is now a ubiquitous—and highly ambiguous—signifier within military, governmental, and industrial discourse. This imprecision, the authors argue, leads to category errors in ethical debate, confounds regulatory efforts, and obfuscates operational risk assessments by eliding crucial heterogeneities in system architectures, training paradigms, and deployment modalities. For example, computer vision-enabled target detection, reinforcement learning-based mission planning, generative LLM-driven document processing, and last-mile fail-safe autonomy in kinetic platforms are all described as “AI,” despite having divergent epistemic profiles and risk surfaces.
The authors’ taxonomy parses the class of so-called “military AI” systems into:
- AI-enabled Decision Support Systems (AI-DSS) for intelligence processing and targeting,
- AI-enabled Autonomous Weapon Systems (AI-AWS) with varying autonomy regimes (last-mile, full mission, swarms),
- Process optimization tools for logistics and R&D leveraging ML and MARL,
- Bureaucratic streamlining LLMs and GenAI platforms for policy and resource management.
Such taxonomization reveals that normative and operational issues—including reliability, opacity, cognitive offloading, bias, adversarial vulnerability, and legal ambiguity—manifest differently across classes and, in many instances, cannot be meaningfully generalized to “AI” as such.
Case Studies: System-Specific Challenges
Decision Support and Targeting
The enumeration of AI-DSS (e.g., Palantir MetaConstellation for sensor fusion, Clearview AI for facial recognition, Primer for SIGINT summarization) illustrates that system rationale, data provenance, and interface design heavily modulate risk. For instance, automation and action bias are acute in target nomination (e.g., Israel's Lavender and Gospel systems), while linguistic/cultural embedding and out-of-distribution generalization are dominant concerns in narrow NLP-enabled intelligence support. The “AI” label, as the analysis underscores, is analytically vacuous without concurrent specification of the architectural and deployment affordances.
Autonomous Engagement and Swarms
Autonomy in lethal engagement further complicates the landscape. The paper distinguishes “last mile” fallback autonomy (e.g., Ukrainian EW-contested drone modules, Hanwha’s Cheongeom ATGM) from full-mission CCA autonomy (e.g., Helsing’s Centaur on Gripen-E, Boeing’s MQ-28 Ghost Bat) and multi-agent swarm systems. The shift from human-in/on-the-loop to fully autonomous operations introduces new human-machine teaming breakdowns, explainability deficits, and command responsibility questions. The severity of these concerns is not coextensive among systems subsumed under “AI,” but tracks closely with both ML/DRL/MARL methodology and operational envelope.
Process Optimization and Bureaucratic Applications
The integration of ML/AI into logistics, mission support, and bureaucracy (e.g., USAF PANDA, GAMECHANGER, Ask Sage, iFinder) foregrounds distinct technical, ethical, and security liabilities. Predictive maintenance algorithms can instantiate survivor bias and lack adaptability under adversarial distribution shift. LLM-based document and advisory systems remain intrinsically at risk for hallucination, legal imprecision, and data poisoning, recapitulating conceptual and practical problems specific to large generative models rather than to “AI” generally.
Implications for Governance, Regulation, and Scientific Discourse
The authors assert that treating “AI” as a singular target for regulation, strategic assessment, or ethical adjudication leads to conceptual incoherence and policy failure. They explicitly advocate that the analytic community must:
- Decompose “AI” into operationally and ethically salient categories (e.g., LLMs, CV classifiers, reinforcement learning agents),
- Specify the precise affordances, vulnerabilities, and interface modalities of each system,
- Contextualize normative debate at the level of deployment scenario and specific interaction,
- Recognize that general-purpose enabling technologies require disaggregated, use-case-driven governance models.
This para-ontological clarification holds for all domains (medical, legal, financial, etc.), but the military use-case is methodologically productive due to its broad system diversity and high-stakes downstream effects.
Prospective Developments in AI Research and Policymaking
The adoption of the paper’s recommendations would necessitate a fundamental shift in academic, policy, and engineering practice. Technical research agendas would prioritize fine-grained empirical study and taxonomization of reference classes of ML/AI system deployment. Regulatory regimes would move toward post-market, context-aware risk assessments rather than abstracted “AI” bans or certifications. Interdisciplinary practitioners (e.g., military ethicists, legal theorists) would now be tasked with dynamic mapping of challenge vectors to system archetypes, rather than relying on ill-defined meta-level critiques.
Conclusion
"Stop Saying 'AI'" (2602.17729) presents a decisive critique of undifferentiated references to “AI” and demonstrates, through system-level taxonomy and analysis, that ethical, legal, and operational debates are intractable without rigorous specification of system architecture, affordances, and context. The principal implication is that technical specificity must become foundational in both research and regulatory discourse, with broad linguistic categories replaced by concrete system references. This argumentative stance will require reshaping the epistemic standards of AI scholarship as well as the legislative and policy apparatuses that attempt to manage the proliferation of general-purpose enabling technologies.