AI’s Intelligence Mirage: Pattern Mastery Without True Understanding

“But is it intelligence?”

Bertrand Meyer posed that question in Communications of the ACM this month, capturing a debate raging through tech circles. AI systems dazzle with feats once thought human-only. They translate languages flawlessly. Diagnose diseases better than some doctors. Yet skeptics insist: no real smarts here. Just clever tricks.

Meyer, chief technology officer at Eiffel Software and professor emeritus at ETH Zurich, pins the confusion on clashing definitions. One camp sees intelligence as grasping concepts—true understanding. The other? Adapting to new situations. Learning from data. Predicting outcomes.

Europeans lean conceptual, he notes. Americans? Pragmatic. Back at Stanford’s AI lab in the 1970s, luminaries like John McCarthy defined it as “the ability to adapt to new situations and learn from experience.” Meyer found that view shocking. Intelligo means “I understand,” after all.

But here’s the rub. Modern AI thrives on the practical side. Large language models pore over billions of examples. Spot patterns. Extrapolate. A translation tool nails nuances no expert system ever could. An image analyzer catches tumors with fewer misses than radiologists. Is that intelligence? Or rote prediction?

Short answer: outcomes matter. Yet claims of understanding crumble under scrutiny.

Consider hallucinations. Ask an LLM to compute eigenvalues of a matrix. It often succeeds. Probe deeper for proofs? Errors creep in. Humans err too. But we claim insight anyway. AI? Mere statistics, critics say.

Meyer calls such arguments unfalsifiable. No experiment proves “real” grasp. Turing Tests measure results. Both humans and AI pass. Ramp up complexity. AI pulls ahead.

Pattern Matching, Not Minds

Recent studies hammer this home. A Carnegie Mellon team tested 14 top LLMs—GPTs, Claude, Gemini—on 500 problems pitting keywords against logic, per their March 2026 preprint ‘The Model Says Walk’. Think the “car wash problem”: closest wash, but no car in sight. Models fixate on “distance.” Ignore the missing vehicle. No model topped 75% under strict rules. A single cue outweighed goals by 8.7 to 38 times.

“Heuristic override,” the authors term it—Yubo Li, Lu Zhang, and colleagues. Keyword vibes trump reasoning. Add a hint? Scores jump 15 points. Strip constraints? Performance tanks. Pattern sniping, not inference.

Iowa State researchers echo this in a April 19 analysis (ScienceDaily). News calls AI “smart.” It “knows.” But “AI does not possess beliefs or feelings,” they warn. It crunches data patterns. Phrases like “ChatGPT knows” inflate illusions. Jo Mackiewicz notes: “The language we choose shapes how readers understand AI systems.”

Black boxes deepen the mystery. New York Times (April 15) contrasts IBM’s Deep Blue—transparent chess brute force—with today’s neural nets. AlexNet revolutionized vision in 2012. Trillions of parameters now. Creators can’t fully decode why they work. Rich Sutton’s “bitter lesson”: mimicking human thought fails long-term. Data devours all.

X chatter amplifies doubts. Peter Jukes: “Apple just proved that the AI is not thinking… It is pattern matching.” Mehdi: LLMs lack “persistent memory… causal reasoning… understanding of physics.” Guri Singh flags CMU: Models “weren’t reasoning. They were pattern-matching on vibes.”

And failures abound. Throw curveballs. AI buckles. No inner model of reality. Yann LeCun bolted Meta for world-modeling billions. Calculators ace math. No clue what numbers mean.

Old AI—logic rules—flopped. Today’s inductive wins. But is prediction enough?

Outcomes Over Ontology

Pragmatists say yes. Compilers churn million-line code flawlessly. Humans can’t. Medical tools slash errors. Vibe-coding boosts programmers. If it acts smart, it is.

Yet risks loom. Overtrust hallucinations. Black boxes hide biases. Anthropomorphism warps policy, expectations.

Meyer favors falsifiability. Theories predict. Einstein bent light—verified. Freud? Zero hits. AI’s predictive power shines. Understanding? Elusive.

So where next? Tools help. Chains of thought nudge better paths. Hints fix overrides. But core limits persist: no consciousness, no intent.

Industry insiders know. Hype sells. Reality demands caution. AI masters patterns. Adapts brilliantly. Understands? Not yet. Maybe never. That’s no knock. It’s a feature—for now.