In 1997, IBM’s Deep Blue defeated Garry Kasparov, the world chess champion, in a six-game match. It was treated as a milestone in artificial intelligence. Researchers debated what the victory meant about machine cognition. The cover of Newsweek declared it “The Brain’s Last Stand.”
Today, every smartphone contains a chess engine that plays at a level that would have seemed impossible to the researchers of that era. Nobody calls it artificial intelligence. It is called a chess app.
This is the naming problem.
The Pattern
The chess example is not unusual. It repeats across the history of the field with enough regularity to deserve a name. Larry Tesler, the computer scientist who gave us cut, copy, and paste, put it directly: “AI is whatever hasn’t been done yet.”1 Once a system reliably performs a task, the task stops being AI and becomes software. The label migrates to the next frontier.
The examples accumulate quickly.
Optical character recognition was a central problem in pattern recognition, computer vision, and artificial intelligence for decades. The ability to read printed text and convert it to editable characters required techniques that once looked recognizably intelligent. It is now a background feature in scanner apps, document workflows, and many PDF tools. Most users do not think of that as an AI system, though in a technical sense it is doing work that earlier researchers would have recognized as part of the AI project.
Spam filtering was exciting machine learning research in the early 2000s. Paul Graham’s 2002 essay “A Plan for Spam” introduced a generation of programmers to Bayesian classification, framing it as a technique for building systems that could learn to distinguish messages.2 The technique worked. It was deployed at massive scale. Today, a spam filter processes most email before it arrives without anyone remarking that artificial intelligence is involved.
Face detection in digital cameras was a research problem through much of the 1990s. By the mid-2000s it was a consumer product feature. By the early 2010s it was a chip. The camera systems on modern phones perform real-time object recognition at a level that far exceeds what the leading AI research systems of 2000 could do. This capability is marketed as “smart camera” or “computational photography,” not as artificial intelligence.
Recommendation engines at streaming services and online retailers were, a decade ago, described regularly as AI systems. The algorithms that learned your viewing patterns and suggested the next show were sufficiently novel to warrant that framing. They are now called “the algorithm,” a word that has somehow become more ominous than the phrase it replaced, and that nobody would describe as a form of intelligence.
Route-finding and navigation were flagship AI problems from the earliest days of the field. Heuristic search, path planning, real-time traffic modeling, automatic rerouting: these draw from bodies of work that overlap heavily with artificial intelligence, graph theory, and operations research. To the user, none of that matters. They are called Maps.
Why the Label Moves
The pattern is not arbitrary. It reflects something real about how we understand intelligence.
We tend to define intelligence as the thing we do not yet understand mechanically. When a process is opaque, we call it thinking. When it becomes transparent, we call it computation. The moment we can explain exactly how a system achieves a result, the result stops seeming intelligent. Chess is a perfect example: once researchers understood that sufficiently deep search with good evaluation functions could defeat human grandmasters, the mystery evaporated. Deep Blue was no longer mysterious. It was a very fast search.
This pattern is often called the AI effect: the tendency to redefine intelligence upward whenever a machine achieves a target capability. Drew McDermott’s 1976 essay “Artificial Intelligence Meets Natural Stupidity” is one early expression of the same concern, particularly its critique of the field’s habit of naming programs as if they were already doing what we hoped they might do.3 Each achievement is reclassified. The goalposts move.
This has an important implication for reading the history of AI. Every winter in the field was partly a failure of delivery and partly a failure of definition. Expert systems did not fail to be AI systems. They succeeded at being the AI systems their architects had designed, and then those systems became routine enough that they were reclassified as software. The first AI winter followed initial claims that were too broad. The second followed a period of genuine commercial deployment that was eventually absorbed into ordinary practice. In both cases, the label retreated faster than the capability did.
Tracing the Thread
The history traced here runs from the 1956 Dartmouth Conference, through symbolic AI and expert systems, through the statistical turn, through deep learning, to the transformer and large language models. Each chapter involved a cluster of capabilities that earned the AI label in their moment and then partially lost it as the capabilities were absorbed into the infrastructure of computing.
The term itself was a choice. John McCarthy and his colleagues at Dartmouth selected “artificial intelligence” deliberately. Other phrases were available: “machine intelligence,” “complex information processing,” “automata studies.” The choice of “artificial intelligence” was partly strategic, a term provocative and broad enough to attract interest and funding, and partly genuine conviction that the problems being attacked were the right problems for understanding mind.4
That original provocation has never fully settled. The question of whether the systems that now carry the label are doing something that deserves to be called intelligence is contested in ways that go beyond marketing. Large language models trained on massive text corpora do things that no one fully predicted and that no one fully understands. Whether that constitutes intelligence depends on a definition that the field has not converged on in seventy years of trying.
What We Will Stop Calling AI
The pattern suggests a question worth asking: which of the things we currently call AI will, in ten years, simply be called software?
Code generation is a plausible candidate. Systems that translate natural language descriptions into working code are currently treated as AI tools. If they become sufficiently reliable, they will be absorbed into development environments and called something like “assisted editing.” The AI framing will fade as the capability becomes routine.
Writing assistance is another candidate. The systems currently marketed as AI writing tools are, in many contexts, becoming defaults in word processors and communication platforms. As the assistance becomes seamless, the label may follow the spam filter into the background: present but unremarked.
Medical image analysis, legal document review, financial fraud detection: each is currently in the phase where the AI label is prominent. Each is in the process of becoming infrastructure. The trajectory is familiar.
What remains AI, in the sense of attracting the label and the debate, will be whatever proves genuinely resistant to mechanization. The history suggests this is harder to predict than it seems. Chess was supposed to require intelligence. It turned out to require search. Natural language understanding was supposed to require intelligence. Current systems perform at levels that were once considered definitively human, and the debate about whether they understand anything has not been resolved.
The Governance Problem
If the label “AI” is a moving target, then governance that addresses AI as a category faces a structural difficulty. The specific system that needs regulation today may be reclassified as ordinary software by the time regulatory frameworks are established. The capabilities that pose the actual risks may be distributed across systems that don’t individually trigger AI-specific review.
This is not an argument against AI governance. It is an argument for governance that targets specific capabilities, specific risks, and specific deployment contexts rather than a category label that the field itself has never been able to stabilize. The EU’s AI Act takes an explicitly risk-based approach, categorizing systems by intended use and potential harm rather than by whether they qualify as “artificial intelligence.” That instinct is correct, even if the implementation is contested.
The naming problem is not going to be solved. The label will keep moving because the underlying phenomenon, the boundary between what machines can and cannot reliably do, is itself moving. What can be done is to notice the pattern, resist the temptation to treat the label as a stable category, and govern the capabilities rather than the name.
The chess app on your phone is genuinely impressive. It is also not what anyone means when they talk about artificial intelligence today. In twenty years, some of what we are discussing today will probably be in the same position. The question is which parts, and what will have replaced them at the frontier.
We will not know until the label moves again.
Footnotes
-
Tesler’s formulation is widely cited but difficult to trace to a single original source. The version most commonly attributed to him is “AI is whatever hasn’t been done yet.” A related formulation, “As soon as it works, no one calls it AI anymore,” is attributed to various sources including John McCarthy. The phenomenon itself is well documented regardless of the precise provenance of either phrase. ↩
-
Graham, P. (2002). A plan for spam. The essay introduced Bayesian spam filtering to a broad programmer audience and demonstrated that a statistical classifier trained on examples could outperform hand-crafted rule systems for this problem. It is an early example of the pattern in which ML research becomes invisible infrastructure. ↩
-
McDermott, D. (1976). Artificial intelligence meets natural stupidity. SIGART Newsletter, 57, 4-9. McDermott’s essay did not coin the phrase “AI effect”; it criticized the rhetorical habits and overclaiming that later discussions of the AI effect point back to. The term was popularized later; the underlying observation is older. ↩
-
McCorduck, P. (2004). Machines Who Think, 2nd ed. A.K. Peters. McCorduck’s history of AI includes detailed accounts of the Dartmouth Conference and the deliberate choice of the term “artificial intelligence.” The original 1955 proposal by McCarthy, Minsky, Rochester, and Shannon is available online. ↩