What Counts as Artificial Intelligence?

A calculator, a payroll workflow, a sales dashboard, a fraud detector. All four process data. Calling all four AI would erase the differences that make any of them worth examining.

An AI system is a machine-based system that uses inputs to infer how to produce outputs such as predictions, content, recommendations, decisions, or actions. The system operates toward explicit or implicit objectives and influences a physical or digital environment.

“Infer” is the important word. The system derives an output from inputs through a model, search procedure, learned representation, or other computational strategy. This definition does not require consciousness, a robot body, or human-like conversation.

That wording is not a teaching device. The OECD Council revised its own definition of an “AI System” on 8 November 2023. The revision travelled almost unchanged into binding law. Article 3(1) of the EU AI Act, in force since 1 August 2024, carries it. An AI system there is machine-based and “designed to operate with varying levels of autonomy”. It “may exhibit adaptiveness after deployment”. For explicit or implicit objectives, it “infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.” Recital 12 of the same Regulation states plainly why the verb carries the weight. “A key characteristic of AI systems is their capability to infer.”

These categories can coexist inside one product. The point is to identify which mechanism handles which part of the job.

Getting that identification wrong can be an enforceable misstatement. On 18 March 2024 the U.S. Securities and Exchange Commission announced settled charges against two investment advisers. Delphia (USA) Inc. and Global Predictions Inc. had told clients and prospective clients they were using AI in ways they were not. “Investment advisers should not mislead the public by saying they are using an AI model when they are not,” said SEC Chair Gary Gensler. “Such AI washing hurts investors.”

Capability labels are useful only when they clarify scope. They become misleading when they are treated as settled scientific thresholds.

The disagreement is countable. On 4 November 2023 Morris and colleagues published “Levels of AGI for Operationalizing Progress on the Path to AGI”. The paper works through “nine prominent examples” of existing AGI definitions. Among them are the Turing Test, consciousness, and analogies to the human brain. Others are the ability to learn new tasks and the capacity for economically valuable work. The authors find each one ambiguous enough to propose a leveled scale in its place. One of them, Shane Legg, is credited in the paper itself with popularizing the term AGI among computer scientists. Nine competing definitions inside a single survey is what a contested concept looks like from the inside.

Borderline cases rarely need a philosophical verdict. They need a transparent classification that explains the mechanism.

Use “AI” for machine-based systems that infer outputs from inputs toward objectives, especially when pattern recognition, search, adaptation, or generation handles complexity that fixed rules cannot fully capture. Then name the exact task and system boundary.

This definition is intentionally less dramatic than science fiction. Its strength is that it makes systems easier to inspect, compare, test, and govern.

Joseph Weizenbaum explained why the border keeps moving in the opening paragraph of the paper that introduced ELIZA (Communications of the ACM, January 1966). Once a program is unmasked and its inner workings are explained plainly, he wrote, “its magic crumbles away; it stands revealed as a mere collection of procedures, each quite comprehensible.” The observer thinks “I could have written that” and moves the program off “the shelf marked ‘intelligent’”. A definition anchored in mechanism survives that reflex. A definition anchored in wonder does not.