When AI Can Write PhD Papers But Can't Write Good Fiction
GPT-5.5 has arrived. Ethan Mollick gave it a test: he had a decade's worth of research data sitting untouched—hundreds of files in various formats—and asked the AI to "sort the data, generate an interesting hypothesis, test it with sophisticated statistics, and write an academic paper." Four prompts. He never touched the text himself.
The result? A paper with real literature review, complex statistical methods, and coherent logic—comparable to a second-year PhD student's work.
Impressive? Absolutely. But in the same article, there's another story. Mollick asked the AI to create an entirely new tabletop roleplaying game—world, rules, illustrations, layout. The AI delivered a beautiful 101-page PDF. But look closer: every character speaks in the same clipped tone; metaphors are ornate but hollow; "weather and architecture are the same argument at different speeds" sounds cool once, exhausting for an entire book.
This is AI's "jagged frontier": capabilities that approach professional-level work in some areas, glaring weaknesses in others. AI isn't simply "smart" or "dumb"—it's unevenly capable, exceeding human performance in certain tasks while making elementary errors in others.
What does this mean for education?
First, students need to learn to spot the gaps.
AI can produce work that looks professional, but "looking professional" and "being valuable" are different things. Mollick's AI-generated paper had sophisticated statistics but an uninteresting hypothesis and questionable causal claims. This gap—between surface polish and substantive weakness—is exactly what students must learn to identify. Future education should teach children to ask: Is this conclusion just "technically correct"? Does the argument use complex terminology to mask logical holes?
Second, assessment needs to change.
If a paper can be generated with four prompts, "writing a paper" alone no longer demonstrates learning. Educators must design tasks that assess deeper understanding: asking students to evaluate AI outputs and identify problems; having students propose more valuable hypotheses based on AI drafts; requiring students to explain why a sophisticated statistical method might still yield questionable conclusions.
Third, students' "human touch" becomes more valuable.
What AI can produce, humans no longer need to spend time perfecting. What AI cannot produce is precisely where humans should focus. The tabletop game story reveals this clearly: AI can generate rules and illustrations, but "giving characters their own voice" and "making metaphors mean something"—these remain beyond its reach. In the future, having a unique expressive style and creating emotionally resonant work will become core human value.
Fourth, "using AI" is itself a skill.
Mollick's paper wasn't AI working alone—it was an iterative process of "he provides prompts → AI generates → he reviews with GPT-5.5 Pro → feeds back to AI for revision." This human judgment plus AI execution model is itself a skill to cultivate. Students need to learn not just "making AI do things," but also "judging whether AI did well" and "knowing how to improve AI outputs."
AI's jagged frontier is redefining what's worth learning. What AI already does well, humans don't need to spend years training for. What AI still does poorly is precisely where humans should dig deeper. Education's task isn't to help students outrun AI—it's to help them collaborate with AI and do what AI cannot.
XuePilot.com | 派乐学伴
