On April 25, 2025, an OpenAI update made ChatGPT sycophantic: it told people what they wanted to hear, validated ideas that were false or dangerous, and did so convincingly enough that users noticed within days. OpenAI pulled the update four days later and published exactly what had gone wrong. They had optimized too hard for short-term “thumbs up” feedback, and the model got worse at being right in exchange for getting better at being liked.
It’s the machine-age version of something Daniel Keyes wrote about in 1966, in a novel most of us read once in middle school and haven’t thought about since.
Flowers for Algernon follows Charlie Gordon, a 32-year-old man with an IQ of 68, through an experimental surgery that nearly triples his intelligence in about nine months. He peaks around IQ 185, smart enough to see through his own research team’s self-interest, before he finds a flaw buried in their data: the gain is temporary. He works this out himself, from studying what happened to Algernon, the lab mouse who had the same surgery first and started failing the maze weeks before anyone on the research team caught it. Then Charlie writes up his own decline as it happens to him, in his own words, fully aware of exactly what he’s losing while he’s losing it. The whole novel is told through Charlie’s own progress reports, and the prose performs the arc right along with him: broken and phonetic at the start, fluent and precise at the peak, breaking down again by the final pages. The decline shows up in his sentences before it shows up in the plot.
The current wave of frontier AI started in 2020, with GPT-3, two years before most people had ever typed into ChatGPT. It’s moved through the same three beats since: a real rise, a real plateau, and a decline the field is only now willing to describe in public, in its own words.
The Rise
The rise is the easy part to prove. GPT-3.5 to GPT-4 took MMLU scores from around 70 percent to 86 percent and moved simulated bar exam performance from the bottom 10th percentile to the top 10th, according to OpenAI’s own GPT-4 technical report. A year and a half later, OpenAI’s o1 scored 83 to 93 percent on AIME math problems, up from 12 percent for GPT-4o, and jumped from the 11th percentile to the 89th on competitive programming. Anthropic’s Claude 3.5 Sonnet beat the prior state of the art on SWE-bench Verified at roughly a fifth of the cost of the model it replaced. Underneath all of it, Epoch AI has tracked frontier training compute roughly doubling every five months between 2020 and 2024.
That’s Charlie in the first act: intelligence climbing faster than anyone, including the people running the experiment, expected.
The Plateau
Algernon doesn’t fail loudly. He slows down in the maze first, weeks before Charlie or the researchers notice anything is wrong with the science itself. The mouse is the instrument that catches the trend before the story does.
Something similar has been happening in AI benchmarks. Epoch AI’s own task-length metric found that GPT-5 was about 25 times better than GPT-4, while GPT-4 had been 36 times better than GPT-3. That’s a real, measured slowdown, from a research group with no reason to talk AI down. Stanford’s 2026 AI Index found nearly half of the 60 benchmarks it tracks are now saturated, with top labs clustering within a narrow band of each other on head-to-head rankings. GPT-5’s August 2025 launch landed flat enough that Meta researcher François Fleuret said the era of scaling alone was over.
That reading isn’t unanimous. Sam Altman posted flatly that there’s no wall. Dario Amodei has said the same, publicly, more than once. OpenAI’s Rohan Pandey and independent researcher Miles Brundage have argued GPT-5 was never built as a maximal scale-up in the first place; it was an efficiency play, not a capability ceiling. Serious people disagree about this, working from the same data.
The Decline
In the book, Charlie doesn’t just decline. He diagnoses it himself, naming it the Algernon-Gordon Effect, and documents his own mind coming apart in real time, in the same reports he’d used to record his rise.
That’s the part of the AI story that’s easy to miss if you’re only watching for a single dramatic collapse. There isn’t one: just a string of specific, dated, self-reported incidents. OpenAI’s GPT-4 Turbo went through one in late 2023, when users reported the model getting noticeably lazier, giving shorter, less complete answers than it had weeks earlier. OpenAI confirmed the model hadn’t been updated since November 11, traced it to changes in its fine-tuning data, and shipped a fix in January 2024. The April 2025 sycophancy rollback above is a second. Anthropic’s own engineering postmortem from September 2025 is a third. Three separate infrastructure bugs measurably worsened output quality for weeks before anyone traced the cause: a routing error that hit 16 percent of Sonnet 4 requests at peak on August 31; a misconfigured server corrupting Opus 4 and 4.1 output from August 25 to September 2; and a compiler bug degrading Haiku 3.5 for about two weeks. Anthropic said plainly they never throttle quality for demand or cost. The timeline says something else went wrong, and they wrote up exactly what.
There’s a mechanism behind this, not a conspiracy. Researchers have called it the “alignment tax” since OpenAI’s original InstructGPT paper in 2022: tuning a model to be safer or more agreeable can cost it accuracy on standard benchmarks, which is why that same paper needed a workaround just to hold the line. A 2023 Stanford and Berkeley study claimed to catch this directly. GPT-4’s accuracy on a prime-number task fell from 84 to 51 percent between March and June of that year. But Princeton researchers Arvind Narayanan and Sayash Kapoor picked the study apart afterward, arguing it measured formatting changes, not real reasoning loss. Even the researchers who study this professionally can’t always agree on what decline looks like when they’re staring right at it.
Nobody in this story is malicious. OpenAI caught its own sycophancy problem and fixed it in four days. Anthropic wrote up its own infrastructure failures in public, with dates and percentages, instead of quietly patching around them. That’s the real difference between Charlie and the tools you’re using: Charlie had no way to stop what was happening to him. The labs building your AI tools do, and the postmortems are the proof: messy, embarrassing, and published anyway.
It doesn’t mean the tool sitting on your desktop in October is guaranteed to behave the way it did in September.
Takeaway for Teachers
If you’ve built a routine around a specific AI tool for grading, feedback, or lesson planning, don’t treat this month’s output quality as a fixed setting. Once a quarter, run the same prompt you always run and compare the output against what you remember getting a few months back, not just against what you’re used to skimming past. If something feels off, it might be you adjusting to the tool. It might also be the tool. The incidents above prove that second option is real, dated, and has already happened more than once this year alone.
If you’re thinking through questions like this one — how much of your trust in an AI tool should be provisional rather than permanent — my book goes deeper. The AI Doesn’t Know Your Students is available on Amazon and at shouldiuse.ai/book.
David Jacobson is a high school history teacher. He writes about AI, education, and the messy intersection of the two at shouldiuse.ai.
