A Paradox Seventy Years Old
In the 1950s, researchers were convinced that once a computer learned to play chess, it would prove the existence of machine intelligence. In 1997, Deep Blue defeated Kasparov.
The reaction was deflating: "That's not intelligence. It's just brute-force search."
Welcome to the AI Effect — a phenomenon first described by John McCarthy: as soon as AI learns to do something, that thing stops counting as intelligence.
The Mechanics of the Shift
Pamela McCorduck put it most precisely:
"It's a strange thing — as soon as AI solves a problem, people say: 'Well, that's just computation. Real intelligence is something else entirely.'"
The pattern repeats with striking regularity:
| Task | When it seemed like "real intelligence" | Reaction after it was solved |
|---|---|---|
| Playing chess | Before 1997 | "Just brute force" |
| Image recognition | Before 2012 | "Just statistics" |
| Language translation | Before 2016 | "Just pattern matching" |
| Writing code | Before 2021 | "Just autocomplete" |
| Generating images | Before 2022 | "No understanding" |
| Medical diagnosis | Before 2024 | "No empathy" |
Each time, the bar for "real intelligence" rises by exactly one step forward.
Why This Happens
There are several competing explanations.
Version 1: The definition is tautological. Intelligence is whatever machines can't yet do. Once they can — it's no longer intelligence. The definition moves with the horizon.
Version 2: Deflation through understanding. When we see the mechanism (neural network weights, search tree traversal), comprehensibility kills the magic.
Version 3: Threat to uniqueness. Humans need to believe their intelligence is irreplaceable. Every AI achievement is an existential challenge — easier to deny than accept.
Practical Consequences for Business
The AI Effect isn't just a philosophical puzzle. It has direct operational consequences.
Undervaluing mature technologies. Companies delay adopting solutions that already work because they perceive them as "mere automation" rather than "real AI." While competitors automate, they wait for AGI.
Inflated expectations about the future. Focus always falls on the next milestone. The real value from existing tools goes unrecognized.
Disappointment cycles. A technology is adopted with inflated expectations, delivers real but modest results, and the conclusion drawn is "AI doesn't work" — even though the tool simply solved a specific problem.
The AI Effect Cycle in Corporate Contexts
How to Work With This Effect
Evaluate by outcome, not mechanism. It doesn't matter whether a model "understands" text in a philosophical sense. What matters: does it reduce document processing time by 70%? Does it reduce error rates?1
Don't wait for "real" AI. It doesn't exist as a fixed point — only as a moving horizon. Value is created by tools that work today.
Call things what they are. "Automated application processing using a language model" sounds less exciting than "AI revolution" — but that precise description helps set realistic goals and measure real ROI.
Conclusion
The AI Effect is a cognitive trap that simultaneously devalues past achievements and inflates expectations for future ones. The only way out: evaluate AI systems by concrete business value, not by conformity to a philosophical ideal of intelligence.
The most effective AI implementations are those people say a year later: "Well, that's just automation." That means it works.
Footnotes
-
According to McKinsey Global Institute (2024), companies that evaluate AI projects through measurable KPIs are 2.4× more likely to achieve planned ROI targets than companies relying on qualitative assessments of system "smartness." ↩
