Strengthen, Don't Substitute: Four Stories and One Test for AI in Learning
By the GenHax Team · October 30, 2026 · Categories: AI in Education, Experiential Learning, Knowledge Graphs, NSF, Media Literacy
On Monday, NSF renewed a national institute to study whether learners can take a skill somewhere new. On Tuesday, Microsoft's chief data scientist offered the cleanest test for AI in education we've read this year. On Wednesday, a youth team from Benin won a UNESCO hackathon by teaching older adults to spot fakes themselves. On Thursday, a survey from Peru counted more than half a million theses that hardly anyone can search. Four stories from four corners of the world, all asking the same question.
Here's the question: after using the AI, can the learner do more without it?
It sounds simple. It isn't. Most AI tools in education are judged by what they produce: the essay, the summary, the answer, the time saved. Very few are judged by what's left in the person once the tool is switched off. That's the difference between a set of training wheels and a taxi. Both get you somewhere. Only one leaves you able to ride.
We spent this week following that thread. Here's what we found, and what we think it means for anyone designing learning experiences right now.
Monday: the hardest test in learning got $18.4 million
On October 20, Georgia Tech announced that NSF had renewed two of its national AI institutes. One of them, AI-ALOE (the National AI Institute for Adult Learning and Online Education), received $18.4 million over five years. The new phase focuses on what researchers call adaptive expertise, which Georgia Tech describes as "the ability to transfer learned skills to new situations."
Ashok Goel, AI-ALOE's director, framed it this way: "Our goal is not simply to use AI to deliver more content or automate instruction." His colleague David Joyner described the aim as studying AI support "without replacing the human elements of education."
Transfer is the hardest test in learning because you can't cram for it. A student who memorized how to solve a problem in chapter 7 often freezes when the same idea shows up in different clothes. A model can recognize the pattern for them, but then the student never builds the habit of recognizing it themselves.
We paired this news with one of our own live challenges, because it's literally a transfer problem. NASA flies mice and other model organisms into orbit. Sergio Baranzini's lab at UCSF built SPOKE GeneLab, a knowledge graph that maps those animals' genes to human orthologs and then links them to pathways and drug targets. When we checked the Open Knowledge Network registry on Monday, it held 140 studies from 23 missions and about 13.2 million gene-expression measurements, all public.
The challenge asks one question: does what biology knows on Earth still hold in space? Learners can't answer that by asking a chatbot. They have to write the queries, find where the results shift, and then explain to a mentor why the shift matters. The AI helps them move faster through the data. It can't do the transferring for them.
Tuesday: a one-line test
On October 26, Juan Lavista Ferres, Microsoft's Chief Data Scientist, introduced Learning Curves, a new book he co-edited with Frank Nagle. One sentence from his post has stayed on our desk all week:
"AI tends to do the most good when it strengthens human capability and the most harm when it substitutes for it."
He followed it with another line worth printing out for every curriculum committee: "The children in our schools deserve tools that serve their learning rather than stand in for it."
The timing matters. Earlier this month, EDUCAUSE released its 2027 Top 10 IT Issues, voted on by 882 people, the largest pool the list has ever had. Number one: "Determining where AI adds real value." Mark McCormack, EDUCAUSE's senior director of research and insights, was careful about what that ranking means: "This doesn't signal consensus as much as it signals urgency."
Urgency without consensus is exactly when a simple test helps. Lavista Ferres's line turns into a question any program manager can ask about any tool: after a student uses it, can they do more on their own than before? If the answer is "they can do more with the tool," that's productivity. That's useful, but it isn't learning. If the answer is "they can do more without it," you have something that strengthens.
This is how we try to design GenHax challenges. AI is a build tool, not an answer machine. Learners ship something real and then defend every choice with a mentor. That defense is where the strengthening shows up, because you can't explain a decision you didn't make. In KGC BUILD, 90% of participants did better on their next AI project. The next project is the right place to measure, because it's the one where the previous tool isn't holding your hand.
Wednesday: teach grandma, don't flag for her
On October 23, UNESCO named the four winners of its 2026 Youth Hackathon on Media and Information Literacy. The numbers alone are remarkable: 2,028 project submissions from 129 countries, judged by 130 international experts. Teams of young people aged 18 to 30 went through a mentoring programme, then submitted a proposal and a three-minute pitch.
One winner, Lucide from Benin, pairs young people with older adults to help them recognize AI-generated content and online scams, with support delivered through WhatsApp. "This victory highlights the important role young people can play in building a safer digital environment," said Prosper Charmel Affoukou of the Lucide team.
Look closely at the design. Lucide could have built a detector that labels videos as fake. Instead, it teaches people to spot fakes themselves. A detector fails the moment the fakes improve. A grandmother who knows what to look for, and who knows a young person she can message, keeps getting better. That's the strengthen-versus-substitute test, passed by a youth team before most institutions have even written it down.
The other three winners deserve credit too. PEKA from Indonesia (Darlene Verica) gives children aged 6 to 12 and their parents a critical-thinking toolkit. Beyond the Bubble from India (Arshad Mahalkari) shows people how their own echo chambers form. I Thought I Knew You, a team spread across Germany, Guatemala, Japan and Ecuador (Carlos Fernando Leon Martinez), turned cloned voices and fake videos into a mystery game played in a school group chat.
Not one of them wrote an essay about disinformation. Each picked a real person, built something for them, got mentored and defended it. The winners will present at the Voices festival in Thessaloniki, November 26 to 28. And in each case, the people who learned the most were probably the builders themselves.
Thursday: 522,334 theses and the shelf problem
Thursday's story was quieter, but it may be the most important one for universities.
Donia Ruelas, Elmer Coyla and Miguel Aceituno of the National University of the Altiplano in Puno, Peru, together with Elwin Huaman of the University of Cambridge, posted a survey on arXiv titled "Building Knowledge Graphs for Theses." It's a preprint, so treat it as a careful literature review, not a final word. They reviewed 46 peer-reviewed studies published between 2021 and 2025 and concluded that while research papers now have rich scholarly knowledge graphs, graph construction for theses "has received comparatively limited attention."
The scale is striking. Peru's national registry, RENATI, "consolidates more than 522,334 theses." The Altiplano's own repository "hosts more than 12,624." Each one represents months or years of a student's work, and most of it sits as PDFs that are hard to search, link or build on.
The authors are honest about the limits of the obvious shortcut. Large language models can help extract information, but "limitations such as hallucination and factual inconsistency persist." That's why they argue for hybrid approaches that combine ontologies with LLMs, and why they name "the lack of end-to-end methodologies specifically designed for thesis collections" as an open gap.
Read it alongside the rest of the week and a pattern appears. A thesis is supposed to show that a student can do research. But if the work goes onto a shelf, nothing builds on it: not the field, not the next student, often not even the author. The learning stops at submission. Last week we asked whether a thesis still proves anything once AI can write it. This week's quieter question is that even the honest ones mostly go unread.
Our NSF challenges try to flip that. Through the EduGate Proto-OKN partnership, more than 600 participants have worked with 8 research teams, building on knowledge graphs that researchers actually query. The work doesn't end at the deadline. It feeds a graph someone else will use tomorrow.
Four design moves for the strengthen test
If you run a course, a bootcamp, an upskilling program or a research group, here's how we'd turn this week into practice.
Measure the next project, not this one. The real signal is how a learner performs on the following task, ideally one that looks different. Build a "transfer check" into your calendar: a short, unfamiliar problem two weeks after the main project ends.
Make every learner defend one decision. A five-minute conversation in which a student explains why they chose one approach over another tells you more than a polished deliverable. It's also nearly impossible to outsource.
Build for a real person. Lucide built for grandparents, and the SPOKE GeneLab challenge builds for researchers. A named user turns "is this good enough to grade?" into "does this actually help?", and that question is much harder to fake.
Give the work a second life. Before a project ends, ask where it will live afterwards: a shared graph, a public repo, a partner's dataset. Work that others build on teaches the builder to care about getting it right.
Try it this weekend
On Friday we closed the week with a small game: six numbers from these four stories to match. (Spoiler: 2,028 is the UNESCO projects, and 522,334 is the theses.) But the number we care about most wasn't in the game. It's the answer to one question you can ask about any student, in any program, this weekend: what can they do now, on their own, that they couldn't do in September?
If the answer is hard to find, the next project is a good place to start. Browse live challenges, including SPOKE GeneLab and our NSF Proto-OKN knowledge-graph challenges, at app.genhax.com/explore, or turn your own idea into a mentored Personal Challenge at app.genhax.com/personal. Researchers who want to bring an open problem to a team of learners can start at app.genhax.com/government. The next GenHax Chronicles newsletter lands on LinkedIn on Monday, November 9.
Learn by building. Prove your skills. And then prove you can do it without the training wheels.
Sources
Georgia Tech Research News, "Georgia Tech's NSF AI Institutes Enter New Era With $38.4 Million in Renewed Support," Oct 20, 2026. news.research.gatech.edu
Open Knowledge Network registry, SPOKE GeneLab statistics (checked Oct 26, 2026). registry.okn.us; NSF award #2333819; BaranziniLab/spoke_genelab
Juan Lavista Ferres, "Learning Curves: The future of education, from classrooms to careers, in an AI-driven world," Microsoft On the Issues, Oct 26, 2026. blogs.microsoft.com
Rhea Kelly, "AI Leads 2027 Educause Top 10," Campus Technology, Oct 1, 2026. campustechnology.com; Abby Sourwine, "Educause '26: Top 10 Reflects Age of Perpetual Change," GovTech, Oct 1, 2026. govtech.com
UNESCO, "Meet the winners of UNESCO's 2026 Youth Hackathon on Media and Information Literacy," Oct 23, 2026. unesco.org
Donia Ruelas, Elmer Coyla, Miguel Aceituno & Elwin Huaman, "Building Knowledge Graphs for Theses: A Survey," arXiv:2610.28648 [cs.DL], preprint, Oct 2026. arxiv.org
GenHax: EduGate Proto-OKN partnership figures, app.genhax.com/government; KGC BUILD outcome data (internal).