What happened when a professor forced students to use AI

While most professors are still scrambling to catch students using ChatGPT, Wharton professor Ethan Mollick decided to flip the script entirely.

This semester, he made AI tools mandatory across three undergraduate and masters-level entrepreneurship classes. Students had to use AI for everything from brainstorming and essay writing to generating images and building app concepts. After several weeks of experimentation, Mollick shared what actually happens when you force an entire classroom to use AI.

His primary takeaway: without proper guidance, almost everyone uses AI poorly at first.

The three types of prompters

To demonstrate this, Mollick assigned a project where students were explicitly told to cheat. The task was to produce a five-paragraph academic essay and an illustration using AI, testing at least five different prompt iterations and writing a reflection on the process.

Predictably, almost everyone began by copying and pasting the prompt directly into ChatGPT. The initial results were universally mediocre C- grade papers. However, forcing students to iterate five times split the room into three distinct camps.

The first group made tiny wording tweaks, basically hoping the model would magically fix itself. The results remained vapid. The second group piled on constraints, industry jargon, and assigned personas, which improved quality but felt impossible to fine-tune reliably.

The clear winner was the third path: co-editing. Instead of demanding a finished product from a single prompt, students treated the AI like an intern. They critiqued specific paragraphs, corrected factual hallucinations, demanded better storytelling, and asked for multiple options before assembling the final piece. This back-and-forth produced the best writing and forced students to engage deeply with the material to spot errors.

Students caught the lies, but stopped asking questions

Educators often worry that students will blindly accept AI hallucinations, but Mollick observed the opposite. Because the grading policy held students strictly accountable for factual accuracy, they audited the output closely.

Students easily spotted fabricated citations and made-up anecdotes. More importantly, they caught subtle conceptual gaps, noting where the AI understood a basic framework but missed its real-world nuance. They also quickly recognized bias, particularly in image generators that defaulted to depicting all business executives as men, and adjusted their prompts accordingly.

Overall project quality surged compared to pre-AI semesters. But Mollick noted an unexpected cultural downside: students raised their hands far less often during lectures. When asked why, one student admitted it was simply easier to ask ChatGPT for an explanation after class than to speak up publicly.

Why it matters

We think Mollick's experiment provides the blueprint for how knowledge work will actually evolve with generative AI. Banning these tools is futile; what matters is shifting the user's role from a consumer of text to a critical editor. Treating AI as an all-in-one generator delivers mediocrity, but treating it as a collaborative drafting partner forces humans to exercise higher-level discernment.

For professionals and learners, the lesson is clear: stop looking for single magic prompts. The highest quality work comes from conversational co-editing, where you challenge the model, correct its hallucinations line by line, and inject domain knowledge.

At the same time, the drop in classroom participation is a real warning sign. When an infinitely patient AI is always waiting on your laptop, the social friction of asking questions in public disappears, along with the spontaneous peer debates that make live environments valuable. Tech can accelerate individual productivity, but keeping the shared room alive will require deliberate design.