AI effect

• 1. Introduction

  ›

  The AI effect is a puzzle that has haunted artificial intelligence research since the field began. A machine beats a grandmaster at chess, and the world marvels. Then, within years, the same achievement is waved aside as mere number-crunching. Historian Pamela McCorduck named it an "odd paradox" in her 2004 book Machines Who Think: the moment a problem is solved, it stops counting as intelligence. How does a genuine breakthrough keep getting reclassified as nothing special? And what does that pattern say about what we actually mean when we use the word intelligence?

• 2. Moving the Goalposts

  ›

  Pamela McCorduck put her finger on the core mechanism. Once an AI system succeeds at a task, that task gets absorbed into other domains, and AI researchers move on to the next unsolved problem. The phenomenon is widely described as an instance of moving the goalposts. Researcher Rodney Brooks made the same observation in 2002, noting that once a system is fully understood, people tend to dismiss it as "just computation". The most compact formulation belongs to Tesler's theorem, often expressed as "AI is whatever hasn't been done yet". That phrase does something useful: it reveals that the word "intelligence" is being used as a placeholder for the unknown, not as a stable description of any fixed capacity. When problems resist full formalisation, researchers sometimes reach for models involving human-assisted computation to describe what is happening.

• 3. Chess, Checkers, and Brute Force

  ›

  IBM's Deep Blue defeated world chess champion Garry Kasparov in 1997, a moment frequently cited as a landmark in AI history. Critics did not celebrate it as proof of machine intelligence. Instead, they argued that Deep Blue relied on brute-force search methods rather than genuine understanding. The same reinterpretation had happened earlier with checkers. Early systems capable of playing both games were initially held up as demonstrations that machines could think. As those systems became better understood, their achievements were reclassified: not intelligence, just computation. The goalposts had moved again. The pattern was consistent enough that it became a recognized feature of the field rather than an isolated reaction to any single system.

• 4. Invisible Technologies

  ›

  Optical character recognition and speech recognition were once central research problems in artificial intelligence. Both required machines to interpret ambiguous real-world input, which seemed to demand something like understanding. As these technologies became reliable and commercially widespread, they were reclassified as standard engineering. Michael Swaine reported in 2007 that AI advances are routinely presented to the public as developments in other fields entirely. Marvin Minsky observed that successful AI innovations tend to evolve into separate disciplines, detaching from their origins in intelligence research. Nick Bostrom noted in 2006 that once a technology is widely adopted, it is often no longer labeled as AI at all. The techniques do not disappear; they become invisible, folded into marketing systems, automation pipelines, and software applications, stripped of the label that once made them remarkable.

• 5. The Age of Large Language Models

  ›

  A 2016 survey of artificial intelligence noted that AI systems were increasingly embedded in everyday applications, reinforcing earlier observations that successful AI technologies tend to become normalized. That trend accelerated sharply with the rise of large language models and other generative AI systems. As these systems gained widespread commercial use, researchers and commentators began describing their outputs as statistical or mechanical, rather than as evidence of intelligence. The current moment is unusual in one respect: unlike earlier periods, including the AI winters when researchers sometimes avoided the term "artificial intelligence" due to negative associations, the label AI is now prominent in public discourse and marketing. Visibility has increased even as the underlying reclassification dynamic continues.

• 6. What We Are Really Protecting

  ›

  Michael Kearns suggested that people may seek to preserve a special role for humans, and that this desire shapes how AI achievements are evaluated. Herbert A. Simon noted that artificial intelligence can provoke strong emotional reactions, which hints at the personal stakes involved in how intelligence gets defined. Some philosophers argue that reclassification is not bias at all, but reflects genuine conceptual distinctions: a system that searches billions of positions per second may truly differ from understanding in some meaningful sense. Similar patterns of reclassification have been observed in studies of animal cognition, suggesting the impulse to redraw the boundary of intelligence is not unique to reactions toward machines. McCorduck's original insight from Machines Who Think remains the sharpest frame: the boundaries of intelligence shift precisely when machines get close enough to cross them.