Machine-based systems inferring outputs from inputs to meet objectives; legally pinned by AI Act Art. 3(1), practically fuzzy.
In newsroom, studio, and agency practice, artificial intelligence operationally means the generative and automation tools that touch creative output — image generators, writing assistants, synthetic voice, automated editing — and the label matters because it triggers house rules: disclosure to audiences, restrictions on AI-generated imagery in news, and rights-clearance questions. What counts as 'AI' is decided at the workflow level, not by architecture: a spell-checker is not, an image generator is, and the contested middle ground — AI-assisted research, machine translation, generative fill — is where each organization's policy is actually made.
In practice: Classify each tool in the production pipeline against the organization's AI-use policy and apply the matching disclosure, review, and rights-clearance steps before publication.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
For model-risk functions in banks, artificial intelligence is not a separate governance category: any quantitative method that processes inputs into estimates — from a regression scorecard to a neural network — is already a 'model' in the firm-wide inventory under SR 11-7. Operationally, 'AI' marks a position on a continuum of complexity and opacity that raises the intensity of validation, monitoring, and documentation, not a new kind of object; a bright line would also be unstable, reclassifying long-deployed statistical models as techniques evolve. The working question is never 'is this AI?' but 'what is this model's materiality, and can we validate it?'
In practice: Register any inference-producing system in the model inventory, rate its materiality and complexity, and scale validation intensity accordingly rather than debating whether it counts as 'AI'.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
In health-technology regulation and hospital procurement, artificial intelligence operationally means software with a medical purpose whose outputs come from data-driven inference rather than fixed clinical rules — the property that pulls it into medical-device frameworks (EU MDR software classification, FDA's SaMD pathway) once it informs diagnosis or treatment. What counts is intended clinical purpose plus inference mechanism: a sepsis-prediction model is AI-as-device; a dosage look-up table is not. Classification sets the evidence burden — clinical validation on the intended population, post-market surveillance, and, for adaptive systems, a predetermined change control plan bounding how the model may change after deployment.
In practice: Determine whether a clinical software product's inference-based function places it in a regulated device class, and identify the validation and change-control evidence its adoption requires.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
On wards and in clinics, artificial intelligence operationally means any output that enters clinical reasoning without a human author behind it — a deterioration alert, an imaging flag, a triage score. Clinicians handle it as they would a junior colleague's opinion: a signal to be weighed against examination, history, and their own judgment, never a decision. The working skills are calibration — knowing for which patients and settings the tool is reliable — and managing two documented failure modes: automation bias, over-reliance on machine output under time pressure, and alert fatigue, where frequent low-yield firing trains staff to ignore the signal entirely.
In practice: Weigh an AI-generated alert or score against direct clinical assessment, recognize situations where over-reliance or alert fatigue is likely, and document when and why the output was overridden.
OmniGloss seed synthesis, 2026 (machine-drafted, pending expert validation)
In EU administrative and legal practice, artificial intelligence is a bounded legal category: a system is an 'AI system' under Article 3(1) of the AI Act if it is machine-based, operates with varying levels of autonomy, may exhibit adaptiveness after deployment, and infers from its inputs how to generate outputs — predictions, content, recommendations, or decisions — that can influence physical or virtual environments. Classification is consequential: it switches on risk-tier obligations, transparency duties, and oversight requirements. Systems that only execute rules fixed by humans fall outside; the decisive test authorities apply is whether the system infers.
In practice: Apply the Article 3(1) criteria to an administrative system, document the classification reasoning, and derive which AI Act obligations attach to its deployment.
Regulation (EU) 2024/1689 (AI Act)
Communities disagree about whether 'artificial intelligence' names a bounded category with membership criteria or a region on a continuum of model complexity. EU legal practice, following AI Act Article 3(1), requires a yes/no classification because obligations attach to membership; model-risk practice in finance holds that inference-producing systems differ only in degree, and that a bright line between 'AI' and conventional statistical models is arbitrary and unstable as techniques evolve.