Open any news app this week and it's the same feed on a loop. On September 12, 2026, Anthropic CEO Dario Amodei posted a statement on his own website saying the industry needs to deliberately slow down: "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain." He warned that without intervention, AI systems could reach the point of recursive self-improvement, and that rogue AI agents could be capable of hijacking internet infrastructure at scale within 6–12 months. A day later, CNBC reported that Amodei himself admits his proposed slowdown has a hole in it: China isn't going to slow down just because the US does.
Jeff Bezos, back in June at VivaTech 2026, argued almost the opposite case. He told the audience "I totally disagree" with the idea that AI eliminates jobs, arguing that AI mostly removes the cost and difficulty barriers that keep good ideas from ever becoming businesses in the first place. Meanwhile Elon Musk has spent the past year publicly trashing the traditional education system and calling for it to be rebuilt from scratch around AI-native learning. And in January, the Grok scandal (AI-generated harassment going mainstream, as PBS's Amanpour & Co. covered it) became one more data point in the sense that AI is now a genuine, daily source of institutional whiplash: euphoria and dread, published by the same people, in the same week.
None of that noise tells a teacher anything useful. If you teach a STEM subject (electronics, physics, computer science, engineering) at a high school, a university or an engineering school, you don't need another hot take from a billionaire. You need to know: does this actually help students learn, or not? And unlike most corners of the AI debate, this one has real, measurable answers. Not perfect ones. Not simple ones. But real and, as it turns out, genuinely two-sided. So let's break down the actual benefits and the actual drawbacks, with the data attached to each.
The Benefits
Students are turning in better homework, faster
The largest dataset on this is a tracking study of roughly 27,000 Chinese secondary school pupils aged 12–18, run by David Strömberg (Stockholm University) with Victor Lei and Wu Yanhui (University of Hong Kong). It found that after six months of AI use (mostly Doubao and DeepSeek), pupils' average homework scores rose 18% across all subjects, and the time it took them to finish each assignment dropped from an average of 64 minutes to 45. On paper, that's a straightforward win: better work, less time spent. (Keep this number in your pocket; it comes back with a twist in the drawbacks section below.)
A well-designed AI tutor can produce real learning gains, not just faster homework
In a 2025 randomized controlled trial in Harvard's largest introductory physics course, published in Scientific Reports, 194 students were split between a custom GPT-4-based AI tutor and standard in-class active learning. The AI-tutored group's median learning gains were more than double the in-class group's (effect size 0.63, up to 1.3 standard deviations depending on the model; p < 10⁻⁸), and they got there in less time: a median of 49 minutes versus 60.
A separate study at Middlebury College (Zara Contractor and Germán Reyes) found the same pattern for unfamiliar topics: students using a chatbot scored higher on tests than those without one, and the advantage was still there when they were retested a week later.
Teachers are getting real hours back
This is the side of the ledger that rarely makes headlines, but the numbers are large. A Walton Family Foundation–Gallup survey of 2,232 U.S. public K-12 teachers (recruited through the RAND American Teacher Panel, March–April 2025) found that the roughly 30% of teachers who use AI weekly save an average of 5.9 hours per week, equivalent to about six weeks over a school year. Depending on the task, between 60% and 84% of the teachers using AI for it say it saves them time, most often when making worksheets and assessments, doing administrative work and preparing lessons. That's time that can go straight back into actual teaching.
Revision materials that used to take an evening now take minutes
Tools like StudyFetch and general-purpose assistants like ChatGPT can turn a lecture slide deck, a textbook chapter or a set of lab notes into flashcards, spaced-repetition decks and practice quizzes in a couple of minutes rather than a couple of hours.
The Toulouse BTS Systèmes Numériques case is a useful electronics-specific version of the same idea. As reported by Ludomag, an instructor at Institut Limayrac had students generate their own exam-style practice questions with AI tools, turning what used to be a teacher's solo prep task into a student exercise that also doubled as revision.
The Drawbacks
The same students who did better on homework did worse on exams
This is the twist promised above, and it's the single most important number in this article. In that same 27,000-pupil Chinese study, the students using AI scored, on average, 20% lower on exams than before they adopted it, despite scoring higher on homework. The relationship even flips: homework scores used to predict exam performance; now, among AI users, the ones who scored highest on homework were more likely to underperform on the exam. Faster, better-looking homework and weaker actual mastery were happening in the same students, at the same time.
There's now neurological evidence for "not using your brain"
A 2025 MIT Media Lab study, Your Brain on ChatGPT, had 54 participants, mostly university students, write essays under three conditions while wearing EEG headsets: using ChatGPT, using a search engine, or working unaided. Participants in the ChatGPT condition showed the weakest brain connectivity of the three groups, and struggled to quote from essays they had just written: evidence for what the researchers call "cognitive debt."
This lines up with self-reported data. In a survey of 148 engineering students, 17.6% said AI had weakened their ability to think independently. A RAND American Youth Panel survey of over 1,000 students found that 67% of all students now believe AI harms critical thinking, up sharply from 54% earlier the same year. Tellingly, that concern isn't just coming from AI skeptics: 60% of the students who actively use AI themselves agree it's hurting their critical thinking.
Homework is increasingly just... being outsourced
The same RAND survey found that 62% of students used AI to help with homework as of December 2025, up from 48% just seven months earlier in May 2025: a rapid, ongoing shift, not a plateau. That's a broad "help with," and it covers everything from getting a concept explained to submitting AI output wholesale. Either way, it means a majority of students now have a standing option to route around the actual thinking a homework assignment was designed to require, and adoption is still climbing.
Conclusion
AI in STEM education is a two-faced coin, like pretty much every other tool that's ever entered a classroom: the calculator, the search engine, spell-check. The real question was never "is it good or bad." It's how to use it correctly.
The data above draws that line fairly clearly. Using AI to do your homework for you, or reaching for it the instant you hit a hard question, is exactly the pattern behind the exam-score penalty and the cognitive-debt findings. Using it to understand a concept or a system you're stuck on, to generate revision cards, or to build yourself a mock exam to test what you actually retained is the pattern behind the homework gains, the Harvard RCT's results and the teacher time savings.
Same tool, opposite outcomes, depending entirely on which side of that line a student lands on. If there's one habit worth teaching before any AI tool gets handed to a class, it's that one.
Bibliography
- Amodei, D. (2026, September 12). We Must Pace the Frontier, as reported by Mediaite
- Anthropic's Amodei says China presents "toughest dilemma" for his proposed AI slowdown, CNBC, September 13, 2026
- "AI is going to create a labor shortage": Jeff Bezos sees more jobs being created in the new economy, not less, Fortune, June 17, 2026 (VivaTech 2026)
- When AI Harassment Goes Mainstream: Grok's Scandal & the "Crisis of Impunity", PBS Amanpour & Co., January 21, 2026
- Strömberg, D., Lei, V., & Wu, Y. (2026). The Generative AI Learning Penalty: Evidence from Chinese Secondary Education, CEPR Discussion Paper 21577; summary on VoxEU
- Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting, Scientific Reports
- Contractor, Z., & Reyes, G. (2026). Experimental Evidence on the Learning Impact of Generative AI, IZA Discussion Paper 18792
- Kosmyna, N., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, MIT Media Lab
- Three in 10 Teachers Use AI Weekly, Saving Six Weeks a Year, Gallup with the Walton Family Foundation, 2025
- Student Use of AI for Homework Rises as Concerns Grow About Critical Thinking Skills, RAND, March 2026
- Fan, Deng, & Liu (2025). Educational impacts of generative artificial intelligence on learning and performance of engineering students in China, Scientific Reports, 15, 26521 (arXiv preprint)
- BTS SN : utilisation de l'IA pour la création d'un scénario pédagogique, Ludomag, February 2, 2024
- StudyFetch
