The most important thing to understand about artificial intelligence in its current form is not what it can do—which is considerable—but what it cannot do, which is nearly everything its most enthusiastic promoters claim.

This is not a criticism born of technophobia or a failure to appreciate genuine progress. The transformer architecture that underlies modern language models represents a legitimate breakthrough in machine learning. These systems can generate coherent prose, translate between languages, summarize documents, and produce working code with a fluency that would have seemed impossible a decade ago. They are useful. They are also, in important ways, profoundly limited.

The reasoning illusion

The central misconception about large language models is that they think. They do not. They predict the next token in a sequence based on statistical patterns learned from their training data. This is not a minor technical distinction; it is the difference between a calculator and a mathematician.

When a language model produces a correct answer to a complex question, it is not because it has understood the question and reasoned toward a solution. It is because its training data contained enough similar questions and answers that the statistical patterns point toward the correct response. This works remarkably well for many tasks. It fails catastrophically for others, often without warning.

The phenomenon of "hallucination"—where models confidently produce false information—is not a bug to be fixed but a feature of how these systems operate. A model that generates plausible text will inevitably generate plausible-sounding falsehoods, because plausibility and truth are different things.

What the benchmarks hide

The AI industry has developed an elaborate apparatus of benchmarks and tests designed to demonstrate progress. Models are evaluated on their ability to pass bar exams, solve mathematics problems, and answer trivia questions. These scores are then presented as evidence of approaching human-level intelligence.

This is misleading in several ways. First, benchmarks measure what benchmarks measure, which is often quite narrow. A model that scores well on a medical licensing exam is not a doctor and cannot practice medicine. It has learned patterns in medical text, including exam-preparation materials. Second, as benchmarks become targets, they become less useful as measures of genuine capability. Models can be optimized for specific tests in ways that do not generalize.

More fundamentally, human intelligence is not a single dimension to be measured by a number. The things that make humans useful in complex situations—judgment, ethical reasoning, the ability to know what we do not know, genuine understanding of context and consequence—are precisely the things these systems lack.

The economic question

None of this means AI is not economically significant. Automation of routine cognitive tasks is already transforming industries, and this process will accelerate. The question is not whether AI will change work but whether the changes will match the extraordinary valuations currently assigned to AI companies.

The gap between demonstrated capability and promised capability is where investment risk lives. Companies valued on the assumption of imminent artificial general intelligence face a different future than companies valued on the ability to automate customer service queries. The former is speculative; the latter is real but perhaps less exciting.

Our take

The useful stance toward AI is neither credulity nor dismissal but calibration. These are powerful tools with genuine applications and real limitations. The people who will benefit most from the current moment are those who understand both—who can identify tasks where statistical pattern-matching excels and recognize situations where it fails. The hype serves those raising capital. The clarity serves everyone else.