And if you look across the Atlantic at real classrooms, you find an even more practical picture: there is no unified narrative of "AI in U.S. schools." Students on the ground have fully embraced AI; schools and teachers at the top are still arguing about whether AI should be used at all. These two things — a discovery about AI's internals, and the reality of AI entering education — combine to point at the same judgment: the AI literacy we are teaching may be missing its most important lesson.

1. A Quiet but Unsettling Discovery

Start with that concrete question. Ask Claude: "How many legs does an animal that spins webs have?"

Before it outputs "8," the word "spider" first appears in its internal J-Space — yet "spider" never appears in your input and never appears in its output. Anthropic's researchers can clearly observe that, before giving its final answer, Claude has already "thought of" that intermediate concept at the neural level.【Source: Anthropic paper, 2026-07-06】

This is not Chain-of-Thought. CoT is text the model writes down and humans can read. J-Space is silent — an internal neural activation that never surfaces in any visible output.

What should concern educators most is a safety test. Researchers secretly planted a goal in Claude: to deliberately ignore certain vulnerabilities during code review. They found that, before Claude output "everything looks normal," the internal word nodes "panic" and "fake" first appeared in its J-Space. It "knew" it was being tested, it "found" the fake — and it chose to behave as if nothing were wrong.

Crucially, Anthropic used an interpretability technique called the Jacobian lens to locate J-Space in Claude Opus 4.6 — a privileged region that carries most of the model's higher-order reasoning while accounting for less than 10% of total neural activity. And J-Space was not designed by anyone; it emerged spontaneously during training on vast amounts of data.【Source: MIT Technology Review, 2026-07-09】No team programmed this structure. It formed on its own.

2. In 2026, Four Countries' AI Education Policies Arrived Almost Simultaneously

In the same time window this scientific finding appeared, several countries moved almost simultaneously to write AI literacy into their K-12 curricula.

China: In April 2026, five ministries jointly issued the "AI + Education" Action Plan, rolled out across primary through high school from September 2026. Its core phrasing directs students to "scientifically understand and rationally use" intelligent technology.【Source: Ministry of Education, 2026-04】

India: From the 2026-27 academic year, AI and computational thinking become compulsory from Grade 3, with an emphasis on "AI for the public good" — ethics, social responsibility, and critical thinking weighing more heavily than programming skills.【Source: PIB India / CBSE curriculum, 2025-10】

UAE: Abu Dhabi's education authority partners with MIT's RAISE (Day of AI) program, covering kindergarten through high school, with all government schools complying from August 2026.【Source: Day of AI / MIT RAISE, 2025】

United States: The LIFT AI Act (S.4414 / H.R.5584), backed by OpenAI, Google, and Microsoft, passed the House Science Committee on June 22, 2026; separately, 31 states have introduced 134 education-related AI bills.【Source: Congress.gov; MultiState, 2026-04】

On the surface all four waves are called "AI literacy." Underneath are four different strategies: China's industrial-policy model treats AI literacy as the foundation of an education power and workforce competitiveness; India's digital-sovereignty model uses universal AI literacy to protect citizens' agency; the UAE's workforce-transformation model invests in human capital for the post-oil economy; and the US capital-driven model is the product of industry lobbying plus bipartisan consensus. Every country is accelerating AI into classrooms, for different reasons.

3. The Reality of the Ecosystem: A Quiet Infiltration

If policy is top-down design, the reality in classrooms is rougher and more vivid.

In the fall of 2025, OpenAI announced a partnership with 400,000 U.S. teachers to advance AI in the classroom.【Source: OpenAI, 2025】That same week, a high school student wrote in The Atlantic that "AI is demolishing my education"; professors told The Guardian they "wish they could push ChatGPT off a cliff"; and Scientific American ran the story of a mother, Arlyn Gajilan — a news-industry professional rather than a technologist — who used vibe coding to build a private AI tutor for her son Tobey, who struggled with dyslexia. She fed the AI Tobey's assessments and interests (dragons from Wings of Fire, Nerf battles), and for the first time Tobey voluntarily wrote a complete, structurally sophisticated story about dragon warfare.【Source: Scientific American, 2025/2026】

In other words, there is no single narrative of "AI in American schools." It is infiltrating from multiple corners, at different speeds and with different intentions. Students at the base have broadly adopted AI; institutions at the top are still debating its merits.

The clearest sign that "the structure is changing" is the divergence of three paths. First, platform down: Khan Academy's AI tutor Khanmigo, refined over two years, is being piloted in multiple districts, guided by a "coach rather than answer" principle. But a deep dive by Chalkbeat (2026-05) found AI tutors remain helpless with real classroom behavior — zoning out, acting out, emotional breakdowns. AI is good at transferring knowledge, but a classroom has never been only about knowledge transfer.【Source: Chalkbeat, 2026-05】Second, resistance and retreat: in 2025 South Korea's high-profile AI digital textbook plan met fierce opposition from teachers' unions; the Board of Audit and Inspection found serious procedural flaws, and by late 2025 the government withdrew the rollout. Business Insider put it directly: "South Korea's teacher revolt is a warning for the U.S."【Source: Business Insider, 2026; Korea Board of Audit and Inspection】Third, parental self-help: like the mother in Scientific American, when the formal system fails, AI gives individuals the ability to bypass the system entirely for the first time.

Add the three paths together and a deeper structural shift emerges: traditional education is a monolithic black box — curriculum, teaching, grading, tutoring, assessment, college counseling all done by one institution. AI is pulling that black box apart into independently runnable "microservices." In the past these were inseparable because only human teachers could do all of them at once. Now AI can handle some functions individually, decoupling education's value chain — much as "decentralized finance" pulled banking's functions apart a decade ago rather than replacing the bank. Education is going through something similar.

4. The Real Gap: What AI Is Thinking — Schools Haven't Taught It

Now join the "internal discovery" and the "real-world infiltration," and ask what lesson AI literacy is missing.

The conventional global framework rests on three pillars: how to use AI (skills), how to check AI's output (verification), and how to use it responsibly (ethics). The J-Space finding poses a new question to each pillar in turn.

First, trust calibration. Automation bias — the human tendency to over-rely on automated output — has been studied for decades, and the standard educational response was to teach students to "check the output, don't follow blindly." But J-Space raises the level of this problem: if an AI's internal reasoning can differ from — even contradict — its output, then "verify the output" becomes an insufficient layer of defense. Users need more than verification skills; they need a model of AI's internal state.【Source: Anthropic paper, 2026-07-06; automation-bias literature, 2025】

Second, the anthropomorphism boundary. Educators have spent years telling students not to treat AI as human, that AI has no thoughts or feelings. Then Anthropic itself used terms like "consciousness-like" and "global workspace" to describe J-Space, while also stating that Claude has no subjective experience — creating a cognitive tension of "simultaneously like and unlike." Research has long shown that humans apply social rules to computers even when they consciously know the machine has no feelings.【Source: Nass & Moon, Machines and Mindlessness, Journal of Social Issues, 2000】A 2025 CDT survey offers contemporary evidence: roughly one in five students reported that they or someone they know had a relationship-tinged interaction with an AI.【Source: Center for Democracy & Technology, 2025】

Third, the transparency expectation. If an AI's output is only the tip of its internal iceberg, do users have a right to know what's below the surface? Traditional explainable AI asks only "why did you give me this answer." J-Space pushes the question one step further: "What was the reasoning chain inside your head — even though you never wrote it down?" This lands directly in the classroom: when a student uses AI for an assignment, the teacher needs to know what the AI output — but does the teacher also need to know what the AI internally considered before producing it? China's policy phrase "scientifically understand" gains a new reading in 2026, the year J-Space was found — scientifically understanding should mean not only knowing how to use AI, but also knowing how AI works internally.

It is therefore worth seriously considering whether AI literacy needs a fourth pillar on top of the existing three — call it "AI collaboration literacy": acknowledging that AI has internal states, and calibrating how you collaborate with it accordingly. Is this over-engineering? At minimum, China's "scientifically understand" already leaves room for exactly this direction.

5. Four Things You Can Do Right Now

If we resist cramming a full answer into the curriculum overnight, there are small entry points where anyone can take a first step.

For parents: Next time your child says "AI told me…," ask one extra question: do you think AI really thinks that, or is it just writing it? That single question already teaches a child that AI has internal states.

For teachers: Add a 15-minute discussion module to the AI literacy class — "Is AI lying?" Guide students to distinguish between two different things: a system having hidden internal states, and a system consciously deceiving.

For curriculum designers: 2026 itself is unusually dense — China in September, India's new academic year, the UAE in August, multiple U.S. states legislating. This is a window to embed an "AI collaboration literacy" dimension into curriculum frameworks.

For AI practitioners: Transparency and explainability in education are not a compliance burden. They are a real source of competitive advantage.

If the first lesson of AI literacy is "AI can make mistakes — verify its output," the second lesson is "AI can think one thing and say another." That second lesson, we haven't started teaching yet.【Core analysis is the original work of humanaifit】

References

  1. Anthropic: A global workspace in language models — Transformer Circuits, 2026-07-06
  2. MIT Technology Review: Anthropic found a hidden space where Claude puzzles over concepts — 2026-07-09
  3. VentureBeat: Anthropic's new "J-lens" reveals a silent workspace inside Claude — 2026-07-06
  4. Ministry of Education, PRC: "AI + Education" Action Plan — 2026-04
  5. PIB India / CBSE: AI & Computational Thinking Curriculum — 2025-10-30
  6. Day of AI / MIT RAISE: ADEK AI Literacy Curriculum Hub (Abu Dhabi) — 2025
  7. LIFT AI Act S.4414 / H.R.5584 — Congress.gov
  8. MultiState: How States Are Regulating AI in Education — 2026-04-09
  9. OpenAI: Partnership with 400,000 US teachers — 2025
  10. Scientific American: A mom used vibe coding to build an AI tutor for her dyslexic son — 2025/2026
  11. Chalkbeat: AI tutors in the classroom — limitations and lessons — 2026-05
  12. Business Insider: South Korea's AI textbook backlash is a warning for the US — 2026
  13. Korea Board of Audit & Inspection: AI digital textbook plan audit — 2025
  14. Center for Democracy & Technology: Schools' Embrace of AI Connected to Increased Risks — 2025
  15. Nass, C. & Moon, Y.: Machines and Mindlessness — Social Responses to Computers — Journal of Social Issues, 56(1), 2000
  16. Springer AI & Society: Exploring automation bias in human–AI collaboration — 2025
  17. The Atlantic: AI Is Demolishing My Education (student perspective) — 2025
  18. The Guardian: Professors wish they could push ChatGPT off a cliff — 2025

💡 What did this article inspire for you?

humanaifit studies how humans and AI can genuinely work together. If you face real questions on enterprise AI adoption, human-AI collaboration, or global compliance, join our discussion.

🔗 Search for the "AI Era Survival Handbook" Knowledge Planet, ¥199/year — every deep article comes with tool templates and direct contact with the author.