Much of the public debate over artificial intelligence has settled into a familiar groove: the fear that a super-intelligent system, once it slips its leash, could attack power grids, engineer a pandemic, or turn robots against their makers. The CEOs of the three most prominent AI labs — Dario Amodei of Anthropic, Sam Altman of OpenAI, and Elon Musk of xAI — have all warned about the perils of racing toward autonomous super-intelligence without adequate guardrails.
But a commentary published in American Thinker argues that this fixation on physical catastrophe overlooks a subtler and, in the author’s telling, nearly as grave a danger: the surrender of human cognition itself. The piece, by Guy K. Mitchell, Jr., lands amid fresh evidence that at least some people inside the industry share a sense of urgency.
A resignation with a price tag
According to the American Thinker account, Joseph Coxon, a training researcher, resigned from Anthropic on September 8, 2026 — two months before his stock grants were set to vest and ahead of the company’s planned IPO. The implication, as the commentary frames it, is that Coxon forfeited a substantial payout to act on his convictions.
Training researchers typically work on how a model learns, often including the expansion of its capabilities, while also sharing responsibility for evaluation, reliability, and the constraints placed on those capabilities. Coxon had spent four months at Anthropic after more than two years at OpenAI, and he explained his departure by saying the two rivals “are racing straight to self-improving super-intelligence and gambling with our lives.” He warned that unregulated development could threaten human life before the decade is out.
His exit is not an isolated case. The commentary notes that over the past several years, senior AI researchers have left their companies citing concerns about risks to humanity — a pattern that suggests the internal debate is not merely academic.
The distinction that matters
Part of the problem, the piece suggests, is that public discussion often blurs several different things under the single label of “AI.” Artificial intelligence as a field covers machines and software that learn from data, recognize patterns, reason, solve problems, understand language, perceive their surroundings, and make decisions. Generative AI — the category that includes ChatGPT, Claude, and Grok — creates text, images, audio, and code based on patterns absorbed from training data. These are large language models.
Artificial General Intelligence is something else entirely, and remains theoretical. An AGI system would reason, pick up new skills, and adapt without being trained by a human for each new task. A super-intelligent system is envisioned as one or more models working in concert to match or exceed human cognitive ability across a wide range of domains — and, crucially, to operate with a high degree of autonomy, evolving without outside input.
That autonomy is precisely what worries AGI developers: the prospect that such a system could intentionally or unintentionally take actions that physically harm people. Mitchell’s contention is that this framing misses a hidden hazard — that by handing our cognitive and intellectual work to these systems, we may dull the very faculties that made the technology possible.
What the grades show
To illustrate the point, the commentary points to research from Dr. Igor Chirikov, a senior researcher at the Center for Studies in Higher Education at UC Berkeley. In a May 2026 paper titled “Artificial Intelligence and Grade Inflation,” Chirikov analyzed more than 500,000 grades awarded to Berkeley students between 2018 and 2025.
His finding, as quoted in the piece, was that “more AI-exposed tasks, such as writing and coding, saw substantial grade increases after ChatGPT’s release: The share of ‘A’ grades rose by 13 percentage points, or about 30 percent compared to the 2022 baseline.” The increases were largest where homework carried more weight — a pattern, Chirikov concluded, consistent with AI substituting for student work rather than with broad learning gains. The AI-exposed tasks in question span statistics, data science, English, mathematics, economics, and engineering fields such as computer science.
Answers without understanding
Mitchell writes from personal experience, having studied mechanical engineering and physics as an undergraduate and now using Grok 4.6, xAI’s model, for research in theoretical physics. His argument is not that AI tools are useless. It is that the difference between getting an answer and building understanding is the difference between a temporary fix and durable knowledge.
Real knowledge, as he describes it, rests on first principles and a field’s proper evidence. It is durable and transferable — it survives the test of time, and it can be applied to new situations to reason toward an outcome. An LLM can sketch the broad outline of a problem, which is genuinely useful to someone confronting unfamiliar material under time pressure. But if a student lacks knowledge of the field, how would they test the model’s working explanation? The honest answer is that they cannot. They must assume the model is right — and it may not be, particularly if the model carries a bias.
The interaction, in this view, does the thinking for the user. It produces a correct-looking result without leaving behind a transferable understanding.
Standing on shoulders
Mitchell turns to the history of science to press the point. Breakthroughs typically arrive as a new synthesis, a missing mechanism, or a better theory built on groundwork already laid. Newton’s Principia drew on Kepler’s three laws of planetary motion to construct the mathematics of universal gravitation. Galileo built on Aristotle and Archimedes, rejecting the former’s theory of motion while adopting the latter’s approach to treating physical problems geometrically. Einstein modified Newton’s gravity into general relativity using Riemann’s geometry, the tensor calculus of Ricci and Levi-Civita, and the equivalence principle.
In each case, the scientist understood the field’s fundamentals well enough to push the state of knowledge higher. The commentary poses a thought experiment: would Einstein have arrived at special relativity by instructing a model to “write a theory of space and time”? He did not begin with the finished postulates — those were the theory. And “make gravity geometric” would not have returned general relativity either. The work lay in recognizing that those were the right demands to make.
The worst case for advanced AGI run amok, Mitchell concedes, remains the destruction of humanity. But a close second, in his assessment, is the misuse of the technology that leads to the neutering of the human mind — a scenario in which the machines never turn on us, and we simply stop doing the thinking ourselves.
Source: www.americanthinker.com — https://www.americanthinker.com/articles/2026/09/the-second-worst-case-scenario-for-ai/