When the Right Answer Stops Counting: Four Stories About Proof in the Age of AI
By the GenHax Team · November 6, 2026 · Categories: AI in Education, Assessment, Experiential Learning, Research, Accessibility
On Monday, a hurricane expert admitted that nobody knows why AI forecasts get things right. On Tuesday, a Fields Medalist stopped grading homework for being correct. On Wednesday, a student team in Auckland built an AI tool that lets you check its work. On Thursday, an audit found that the research saying AI boosts learning doesn't hold up. Four stories, one problem.
The problem is this: a correct answer used to be decent evidence that someone understood something. It isn't anymore.
That's not news to anyone who has graded an essay since 2023. What surprised us this week was how far the problem reaches. It shows up in weather forecasting, in university math, and in the research literature that is supposed to tell us whether AI helps students at all. In each case, the answer looks right, and the reasoning behind it is missing or can't be checked.
And in one case, a team of students designed their way out of it. Here's the week, and what we think it means for anyone building learning experiences right now.
Monday: right answer, no explanation
On October 26, WUSF reporter Jessica Meszaros published a story for NPR on how AI is changing hurricane forecasting. One quote from hurricane expert Michael Lowry has stuck with us:
"We don't know why they might get a forecast right. We also don't know why they might get a forecast wrong."
The example was Hurricane Isaias. Conventional models first sent the storm toward southeastern Louisiana. Google DeepMind's AI model had the Florida Panhandle path right from the start. On intensity, though, it trailed the National Hurricane Center's own blended forecast. So NHC does what a careful teacher does with a talented but unexplained student. It takes the AI's input seriously, blends it with "several proven models," both AI and traditional, and keeps a human reading the output.
We paired the story with our Climate forecast challenge, run with Temple University and NSF. Eduard Dragut and Longin Jan Latecki lead an NSF Proto-OKN project (award #2333789) that built a public knowledge graph of climate science. When we queried it on Monday, it held 10,910 papers, 394 climate models and 481 experiments. The project explicitly covers both physics-based and AI-based models, and it names "teaching and training of students at various levels, from high school to doctoral programs" as a goal.
For learners, that graph is the opposite of a black box. Every claim links back to a paper, a model or an experiment. You can't just say a model is good at hurricanes. You have to show which experiments say so.
Tuesday: correct is cheap now
The next day we turned to UC Berkeley, where Richard Borcherds, a Fields Medalist, stopped grading homework for correctness. Students now get credit for handing it in. He explained why by email to Newsweek:
"AI can now solve many undergraduate mathematics problems, often better than the students themselves."
He isn't giving up on standards. "Students still need to understand the material to do well on examinations," he wrote, adding that "in-person exams still work fine for assessing students." He hopes that taking grades off correctness will "reduce the incentive to just ask AI for answers and increase the chance that they do the homework themselves."
MIT's Eric Klopfer went further in the same piece: "At this point, at least for some students, homework is no longer a valid evaluation of their learning." Cornell's Rene Kizilcec put it in six words: "Assessment reform is necessary and urgent."
We agree with all three. Still, an in-person exam is only one place where proof can live. The other is a thing you built, explained out loud to someone who knows the field. A chatbot can write a proof. It can't stand next to you while a mentor asks why you picked this data source over that one, or what you'd change if the results had come out the other way. That conversation is the part you can't paste in.
That's how GenHax challenges end. Teams ship a working project and walk mentors through it. As one participant, Ankita B., put it: "I learned more in 2 weeks than in my first year as a professional developer." Two weeks of building something real, defending it, and fixing it is a very different kind of evidence than a correct worksheet.
Wednesday: show the original
Midweek, we went looking for someone who had solved the problem instead of describing it. We found Kindred.
Kindred was built by a team of ten at Auckland International Campus (Otago Polytechnic) for the Aotearoa AI Hackathon 2026, run by Unitec with Seen Ventures. The idea came from team lead Daasha Fernando. The app helps Deaf people understand everyday documents, like bills, tenancy notices and benefit letters, by translating them into New Zealand Sign Language with plain-language captions. It won third place and an honourable mention.
Here's the design choice we keep coming back to. Kindred displays "the original document alongside the translation." The AI doesn't ask to be trusted. It puts its source on the same screen, so anyone can check one against the other.
That small decision answers Monday's and Tuesday's problem better than any policy memo we read this week. When you can't see the reasoning, show the source. When you can't verify the answer, make checking it part of the product. And it came from students working on a real problem for real people. Credit to Daasha Fernando, Reign Beato, Reggie Ragot, Sujal Gajera, Ayush Subedi, Yax Avaiya and their teammates, with Senior Lecturer Kathiravelu Ganeshan supporting them.
It reminded us of something Marie Francesca Berrouet, Program Manager at C-STEP at Queens Community College, said about GenHax's own social-impact work: "I love that this is NOT destructive AI. You just connect the dots for what already existed and made it easier." Kindred doesn't replace the letter. It makes the letter readable, and the letter stays right there.
Thursday: the research has the same problem
If you've sat through a conference talk in the last two years, you've probably seen a slide that says "studies show AI improves learning." On November 2, Education Week's Stephen Sawchuk reported on a researcher who decided to check.
Patrick O'Neill, an associate professor at Ivy Tech Community College in Bloomington, Indiana, looked at 14 meta-analyses drawn from nearly 40 that all concluded AI had positive effects on learning. According to EdWeek, all 14 "fail basic standards for this type of research." One put an elementary physical-education study in the same pool as medical-student simulations. "A few included studies that didn't even use AI tools." Knowledge was measured "alongside motivation, engagement, mental health, and perceived learning all in the same bucket." Some references were likely hallucinated. These papers have been cited more than 3,000 times, and one, published in Nature, was retracted a year later.
O'Neill's diagnosis: "There is a hunger for this evidence which is driving people to publish quickly and overstate the solutions." MIT's Justin Reich, who hosted a webinar on the audit, called it "a perfect storm of factors to motivate and incentivize problematic research."
A separate review from Stanford's AI Hub for Learning found that of more than 800 studies, just 20 had a strong cause-and-effect design, and none of those involved K-12 students. Stanford's Susanna Loeb said that getting this right "requires that each of us are really honest in what papers and research studies can do and what they can't do."
To be clear about what this means: O'Neill didn't find that AI hurts learning. He found that we don't yet have good evidence either way. That's an uncomfortable but honest place to stand. So in our Thursday post we didn't add our own effect size to the pile. Instead, we pointed at what a challenge produces that anyone can inspect: a built artifact, reviewed by a mentor, on knowledge graphs that NSF Proto-OKN research teams actually use. More than 600 participants have worked with 8 research teams that way.
NSF seems to be asking for the same thing. Its FINDERS FOUNDRY program for K-12 AI education (NSF 26-507) requires every leadership team to include at least one K-12 educator, one technologist, one researcher and one parent or guardian. Evidence that's built with the people who'll use it is harder to overstate.
What we're taking into next week
Put the four stories side by side and a pattern shows up. Each one gives a practical move for anyone designing a course, a program or a challenge:
Grade the explanation, not just the answer. Borcherds moved the stakes to exams. You can also move them to a short oral walkthrough: five minutes where the learner explains one decision they made, and one they'd change.
Keep the source on screen. Kindred's rule works in any classroom. If a student used AI, ask them to show what it was working from and where they checked it. A claim that comes with its source can be checked.
Blend, don't hand over. NHC doesn't fire its forecasters when an AI model gets a track right. It blends the model in and keeps a human responsible. Treat AI output in student work the same way: as one input the learner has to reconcile with others.
Be honest about what you can measure. If the evidence on AI and learning is still thin, don't pretend otherwise in your program reports. Collect the artifacts instead: projects, mentor feedback, the next build. They're slower to summarize and much harder to fake.
None of this needs a new tool. It needs work that is real enough to be worth explaining, and someone in the room who knows enough to ask "why?"
Try it this weekend
We closed the week with a small game: five quotes, five speakers. (Spoiler: the "hunger for this evidence" line is O'Neill's.) The best answer of the week wasn't a quote at all, though. It was a design choice by a student team in Auckland: show the original.
If you want your learners to build something they can explain, browse live challenges at app.genhax.com/explore, including the Climate forecast challenge with Temple University and our NSF Proto-OKN knowledge-graph challenges. 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 keep the original on screen.
Sources
Jessica Meszaros, "How AI is transforming hurricane forecasts," WUSF/NPR, Oct 26, 2026. kuow.org
NSF award #2333789, "Proto-OKN Theme 1: Knowledge Graph to Support Evaluation and Development of Climate Models," Temple University (PI Eduard Dragut, co-PI Longin Jan Latecki). nsf.gov; Climate Models KG statistics from a live query on the Open Knowledge Network, Nov 2, 2026. registry.okn.us
Amanda Greenwood & Hollie Silverman, "Berkeley professor stops grading homework over AI: how teachers are adapting," Newsweek, Oct 2026. newsweek.com. Originally reported by The Daily Californian.
Auckland International Campus, "Students develop AI tool for Deaf community," Aug 10, 2026. auckland.op.ac.nz
Stephen Sawchuk, "The Research on AI in Education Is a Mess," Education Week, Nov 2, 2026. edweek.org; Patrick O'Neill preprint, EdArXiv
Stanford SCALE / AI Hub for Learning, "The Evidence Base on AI in K-12," 2026 (as reported by Education Week). scale.stanford.edu
NSF 26-507, FINDERS FOUNDRY solicitation. nsf.gov
GenHax: Proto-OKN participation figures, app.genhax.com/government; community testimonials from genhax.com.