Artificial intelligence can lift children's grades and help them finish assignments faster. But a rising pile of evidence, and a major policy shift in New York City, suggests that better output on schoolwork does not automatically mean better learning.

Every generation of educational technology brings a wave of worry. When calculators became common, teachers argued that students would lose basic arithmetic skills. When laptops and smartphones arrived, schools debated distraction, internet research, and academic honesty. Generative AI sharpens those old tensions because it can produce complete answers, essays, and code in seconds. A student can turn in homework without ever understanding how the answer was reached.

This moment is also different because of the scale. Students have adopted AI quickly, and schools have not entirely caught up.

AI use in schools is already widespread

According to a 2026 Pew Research Center survey, 54 percent of students said they had used AI to help with schoolwork, and 12 percent said they relied on AI for all assignments. Those numbers include the entire student population, not just those taking computer science classes. The same pattern appears on the teacher side. A 2025 report from the Center for Democracy and Technology estimated that 85 percent of teachers and 86 percent of students used AI during the 2024-25 school year.

The meaning of basic use varies widely. Some students use AI to brainstorm, summarize a confusing reading, or check grammar. Others ask it to write essays from scratch or solve math problems with no student work shown. The broad totals do not reveal whether AI is being used as a tutor or as a ghostwriter.

Good grades today, no learning tomorrow?

A recent Stanford University study offers one of the clearest warning signs. Researchers looked at K-12 students working on math, writing, and coding tasks. When students used AI tools, they often performed better on those tasks. But when the tools were removed, the improvement mostly vanished.

That gap matters. If AI merely supplies the final answer, the student may appear proficient in the moment and then struggle when the support is gone. The Stanford researchers also found an important difference among tools. Students who used general-purpose chatbots were especially likely to become dependent on AI. Students who used AI systems that were carefully designed for learning showed more promise, although even those systems did not guarantee lasting gains.

The problem is not that AI makes students lazy. It is that many tools optimize for completing assignments rather than for building understanding. A seventh grader who tells a chatbot to write a paragraph about photosynthesis may receive a perfectly organized paragraph, but the student has not learned how to connect the concepts, choose evidence, or use the right vocabulary. When the chatbot disappears, the paragraph does too.

Another study from the Brookings Institution came to a similar conclusion. Well-designed AI tools, used as part of an overall, pedagogically sound approach, can benefit the classroom. But an overreliance on AI chatbots can hinder a student's ability to learn. It can also weaken relationships with teachers and classmates, because conversation and collaboration are replaced by private interactions with a machine.

Educators increasingly use the phrase performance versus learning. AI can improve performance on a single assignment. Learning is the harder, slower process of acquiring skills and knowledge that transfer to new problems. The two do not always move together.

New York City steps back from classroom AI

New York City has responded with a policy that is tough by national standards, yet limited in time. The city announced a one-year moratorium on student-facing generative AI for public school students through the eighth grade. That means public elementary and middle school students will not be allowed to use AI chatbots or similar generative tools in classrooms during the pause.

Several details of the plan have generated controversy. First, it lasts only one school year, while many parents wanted a two-year pause. Second, high schools are not included, so students in ninth grade and above can still use AI in school. Third, the ban contains loopholes. Centrally approved e-books and coding programs are exempt, even if they include AI features.

Supporters say a pause gives educators time to set rules, evaluate tools, and receive training. Opponents argue that one year is too short to resolve complex questions about privacy, academic integrity, and equity. They also point out that students can still meet AI outside school hours. A ban that only applies to the classroom cannot stop a student from opening a chatbot at home.

The debate is not entirely new. New York City originally banned ChatGPT in its public schools in early 2023, then reversed that decision a few months later to explore the potential benefits. The new policy is more nuanced than a simple return to the old ban, but it invites the same question: is it possible to develop a useful classroom policy for a technology that changes every few months?

For parents, the central worry is educational dependence. A child can earn a diploma by having AI do the heavy lifting, but will not have learned how to reason through unfamiliar problems. The best educational outcome is not a completed worksheet; it is the ability to learn on one's own. Coding classes, like the Scratch visual programming lessons offered in some New York schools, can teach logical thinking in a structured way. But if AI autocompletes every line of code, that learning opportunity disappears.

The New York policy is not perfect. It is, however, an attempt to ask a necessary question before the technology becomes too embedded to remove. Some observers fear that the point of no return has already passed. The tools are free, unlimited in many cases, and firmly integrated into the lives of students. A one-year pause may change what happens in certain classrooms while having little effect on overall habits.

Other AI developments this week

• Anthropic's Claude Fable 5.1 and Mythos 5.1 models are adding invisible watermarks to text and file outputs, fulfilling earlier promises and making AI-generated content easier to trace.
• Privacy settings for Claude deserve attention. Five overlooked controls can determine how much conversational data is stored, shared, or used for training.
• The excitement around OpenClaw, a viral personal AI assistant tool that recently reached version 2.0, has faded, but the tool helped push agentic AI into the mainstream by showing what an autonomous helper can do.
• Google Notebook is now using computer-specific usage limits, which could restrict how many AI-generated podcasts users can create from their notes and documents.
• Google is taking aim at Canva with Pics, an image creation and editing tool powered by the Nano Banana image model. Pics lets people create and modify visuals with simple prompts.
• Lawmakers have heard more about an AI kill switch. Reports say OpenAI is working with Congress on automatic shutdown systems that could disable an AI that goes rogue.

The prompt that makes AI double-check itself

This week's featured prompt is designed for Claude Fable 5.1, Anthropic's latest model, but the idea works across many AI tools. The prompt tries to fix a specific flaw: the model sometimes skips a search and instead relies on memory during quick or low-effort sessions. That tendency can produce confidently stated but wrong information.

The verification prompt nudges the model to use its search tool before answering a factual question. It asks the AI to confirm names, dates, and other specific claims against independent sources, rather than assuming that its training data is accurate. In effect, it forces clarity between I know and I need to check. Pairing that with the new watermarking features gives users more confidence in knowing what came from the model and whether it took the time to verify.

This kind of prompt matters outside the novelty of AI news. Many students and professionals use AI assistants as quick reference tools. When those assistants guess, they create a false sense of certainty. Teaching an AI to check its facts is not just a technical exercise. It is a reminder that learning requires verification, and that applies to both humans and machines.

Source: PCWorld News