Machine Intelligence Is Getting Cheap. Human Judgment Is Not.

The price of machine intelligence is collapsing. Stanford's 2025 AI Index found that querying an AI model at a fixed level of capability cost 280 times less in late 2024 than it did just 18 months earlier, and businesses now openly measure "intelligence per dollar" the way they once measured cost per shipment. That shift does not make human intelligence worthless. It changes which human abilities carry a premium: when answers become nearly free, the value moves to the people who can ask the right questions, judge the answers, and catch the ones that are wrong.

What happened: intelligence became a line item

For most of history, buying more thinking meant hiring more people. That link broke quietly over the past few years. According to the Stanford HAI 2025 AI Index Report, the cost of running an AI model that performs at the level of GPT-3.5 on a standard knowledge benchmark fell from 20 dollars per million tokens in November 2022 to 7 cents per million tokens by October 2024, a reduction of more than 280-fold in about 18 months. Hardware costs declined about 30 percent per year over the same period, and energy efficiency improved roughly 40 percent annually.

Business writers now describe this as the new economics of "intelligence per dollar": companies compare AI providers on how much competent output a dollar buys, and they redesign workflows around the assumption that a first draft, a summary, or an analysis costs almost nothing. When something becomes 280 times cheaper in a year and a half, organizations stop rationing it. Routine cognitive work, the kind that once justified an entry-level analyst's salary, is being repriced toward zero.

Why it matters: cheap answers make judgment expensive

Economics has a reliable rule for moments like this: when the price of a good collapses, the value flows to its complements. When calculators became free, arithmetic stopped being a career skill, but the ability to know which calculation mattered became more valuable, not less. The same logic is now playing out with reasoning itself.

Machine-generated answers are abundant, fast, and confident. What they are not is self-verifying. Every output still needs a human somewhere in the chain to decide three things: whether the question was the right one to ask, whether the answer is actually correct, and whether it should be acted on. Those three judgments are exactly the part of intelligence that has not gotten cheaper. If anything, the flood of fluent, plausible, occasionally wrong machine output has made them scarcer and more consequential.

This is the practical meaning of real-world intelligence in the current decade. It is not the ability to produce information, which machines now do for pennies. It is the ability to evaluate information under uncertainty: to notice when a statistic is implausible, when a chain of reasoning skips a step, when an answer is precisely worded but subtly off. Employers are already reorganizing around this distinction, keeping fewer people whose job is producing routine analysis and paying more for people whose job is deciding which analysis to trust.

The catch: outsourcing thinking can erode it

There is a complication, and it is well documented. In a 2025 study published in the journal Societies, researcher Michael Gerlich surveyed and tested 666 participants and found that heavier use of AI tools was associated with lower measured critical thinking, with cognitive offloading, the habit of handing mental work to a tool rather than doing it yourself, acting as the key mediating factor. Younger participants in the study relied on AI tools more and scored lower on critical thinking assessments. Notably, higher educational attainment buffered the effect: people with strong learned habits of reflection kept more of their edge even while using the tools.

The finding should be read carefully. It is correlational, and it does not show that AI tools make people incapable. What it shows is a plausible trap: the same cheap intelligence that raises the value of human judgment can, if used passively, atrophy the very faculties that judgment depends on. A muscle you never load gets weaker. The people who benefit most from the intelligence-per-dollar era will be the ones who use machine output as raw material for their own reasoning rather than as a replacement for it.

The abilities that are gaining value

Asking better questions

A cheap answer to the wrong question is worth nothing, and organizations generate wrong questions constantly. Framing a problem well, knowing what to ask, in what order, with what constraints, remains a human bottleneck. People who do this well multiply the value of every dollar spent on machine intelligence downstream of them.

Verification and calibrated skepticism

The scarce skill is not spotting obvious errors; it is maintaining calibrated doubt about fluent, well-formatted output. That means estimating orders of magnitude before accepting a figure, checking a cited source before repeating it, and treating confidence as a property of evidence rather than of tone. In hiring terms, this is the difference between someone who forwards an AI-drafted report and someone who catches the error in paragraph four.

Judgment under incomplete information

Models are strongest where the answer is already latent in their training data. Real decisions, whether to take the job, trust the contractor, enter the market, are made with missing information, conflicting incentives, and consequences that arrive slowly. Practical judgment of this kind is what intelligence researchers try to capture when they distinguish real-world problem solving from abstract puzzle performance, and it is precisely what a per-token price cannot buy.

Transfer: applying knowledge out of context

Cheap intelligence is trained on the past. Humans still hold the advantage in transferring a principle from one domain to a genuinely novel one, the move that lets a nurse's triage instinct improve a customer-support queue, or a chess player's sense of tempo inform a negotiation. Transfer is hard to benchmark, which is exactly why it stays valuable.

What happens next

Expect the gap to widen in both directions. Routine cognitive output will keep getting cheaper, probably by another order of magnitude, because the trend lines in the Stanford data show no sign of flattening. At the same time, the premium on evaluative skill will grow, because every additional unit of cheap machine output creates additional demand for someone to judge it. The likely equilibrium is uncomfortable but clear: fewer jobs producing answers, more jobs, and better-paid ones, deciding which answers to trust and what to do about them.

For individuals, the strategy follows directly. Do not compete with the machine on volume; that market is gone. Compete on the three things that got more valuable: question quality, verification, and judgment. And treat those abilities the way an athlete treats conditioning, as something measured and trained rather than assumed.

How to keep your edge measurable

Start with an honest baseline. General intelligence measured by abstract puzzles is only loosely connected to the evaluative skills described above, which is why assessments built around applied reasoning exist. You can take the Real World IQ test to see how you perform on problems drawn from practical situations rather than textbook patterns.

Then train deliberately. Before accepting any machine-generated answer that matters, write down your own estimate first, even a rough one, so the tool confirms or corrects your reasoning instead of replacing it. Verify one claim per day back to a primary source. Practice explaining a decision you made this week to someone outside your field; if the explanation collapses without jargon, the reasoning may have been borrowed rather than owned. These habits are small, but the Gerlich findings suggest that the difference between users whose thinking sharpens and users whose thinking dulls comes down to exactly this kind of active engagement.

Puzzle constructors have watched this dynamic for decades. Timothy E. Parker, the Guinness World Records Puzzle Master, has put puzzles in front of 180 million total solvers across 30 years and more than 80 countries, and the consistent pattern is that solvers stay sharp by doing the work, not by reading the answer key. Cheap machine intelligence is, in effect, a universal answer key. Whether it makes you sharper or duller depends on when you choose to look at it. If you want a benchmark for where your applied reasoning stands today, the Real World IQ test is a practical place to start.

FAQ

What does "intelligence per dollar" mean?

It is a way of measuring how much competent AI output a fixed budget buys. As model prices fall, the same dollar purchases more analysis, drafting, and problem solving, which is why businesses now compare AI systems on cost per unit of capability rather than raw capability alone.

How much cheaper has AI actually become?

Stanford's 2025 AI Index documented a drop from 20 dollars to 7 cents per million tokens, more than 280-fold in about 18 months, for AI performance at the level of GPT-3.5 on a standard benchmark. Hardware and energy efficiency improved sharply over the same period.

Does using AI tools lower your intelligence?

The evidence shows an association, not a sentence. A 2025 study of 666 participants found heavier AI use correlated with lower critical thinking scores, mediated by cognitive offloading. But the effect was buffered by education and reflective habits, which suggests active, questioning use of AI protects thinking while passive dependence erodes it.

Which human skills become more valuable as AI gets cheaper?

Question framing, verification, judgment under incomplete information, and the ability to transfer knowledge to novel situations. These are complements to machine intelligence: the more cheap output exists, the more valuable the people who can evaluate and direct it become.