Hi, it’s Kerry here, back to host The Last Mile for a few weeks while Damian takes a well-earned break after graduating from college.
ChatGPT will be nearly three years old when the next college year starts this fall, so Damian is probably one of the last cohort of students to start college without a generative AI sidekick from day 1.
So, is that good or bad for learning—and what does it mean for businesses that depend on a steady pipeline of curious, capable talent?
The once-in-a-generation opportunity
In 1984, educational psychologist Benjamin Bloom found that one-on-one tutoring vaulted average students two standard deviations above their classroom-taught peers. The famous “2 Sigma Problem” set a gold standard: personalized guidance beats broadcast instruction almost every time.
Generative AI looks like the long-awaited scalable tutor. A tireless bot that can answer questions at 3 a.m., tailor examples to a student’s interests, and loop endlessly until understanding clicks.
A good tutor doesn’t just spit out answers; they coax, challenge, and occasionally push a learner into productive struggle. Large language models weren’t tuned for that.
They were tuned to make the user happy—quick, fluent, confident. That’s fine for drafting email copy. It’s disastrous when “happy” means handing over solutions before the student even tries.
Fortunately, we can point these models in the right direction. Wharton Generative AI Labs have open-sourced a prompt they’ve used in several academic studies to improve how AI interacts with students, aligning it more closely with modern educational approaches: get the AI tutor prompt here
An easy way to use it is to create a Project in ChatGPT and drop the Wharton prompt text into the “Add instructions” box. That will instantly switch the model’s persona from answer-bot to coach for any chat you have inside of that Project.
Try it yourself and encourage your team—and kids—to use it!
What the data says
Wharton researchers put almost a thousand high-schoolers through four 90-minute algebra sessions with two flavors of GPT-4:
- GPT Base – the vanilla chat interface.
- GPT Tutor – the same model, wrapped in a prompt similar to the open-source version linked above.
While the bots were available, both groups aced their practice sets—scores jumped 48% with GPT Base and 127% with GPT Tutor.
The twist came later. When AI was taken away, the GPT Base students crashed, scoring 17% worse than peers who never used AI at all, while the results of students using GPT Tutor held steady.
What this means for you, your team, and your business
I’ve talked before about how we need to Delegate, not Defer to AI.
We need to stay on top of the models because we remain morally—and legally—responsible for their output in professional settings.
As AI becomes more and more capable, so must we. Adults in the workplace need to up-skill and re-skill in order to make the most of AI.
AI can actually help with that, but it doesn’t come naturally. You need to deliberately coach the AI to operate as a learning partner, not an answer machine.
AI doesn’t make work easier; it makes it harder—in a good way!
As we offload the easier cognitive tasks to AI, it means we need to think harder and deeper about more difficult and consequential problems.
To do that, we need to learn faster and embrace the struggle. AI won’t do that naturally and nor—in some cases—will members of your team.
The message is clear: AI can free us to work on bigger, harder, deeper problems. But to do so, we need to become voracious learners and leverage AI to give us the 1-on-1 personal tuition that Bloom found gave that 2 standard deviation edge.
Kerry
PS: If you’re new here, this newsletter brings you the best from Waterfield Tech experts at the frontier of AI, CX, and IT. Also, Kerry posts weekly at The Dualist, and Fish and Dan share their thoughts every other week at Outside Shot and Daichotome.
Here’s what went down this week.
Bleeding Edge
Early signals you should keep on your radar
Claude Code usage up 5.5x. Anthropic’s industry-leading coding assistant is experiencing massive growth, and is bringing reporting dashboards to the product to help engineering managers keep on top of costs – and justify impact. (Venture Beat)
Fair-use fight just got bigger. A Judge has certified a lawsuit that authors are bringing against Anthropic as a nationwide class action, meaning every U.S. writer whose book was scraped can now pursue damages—potentially billions—even though his June order still calls the model-training “spectacularly transformative” fair use. (Reuters)
Leading Edge
Proven moves you can copy today.
Chest-X-ray co-pilot goes open source. Researchers released Ark+, a free, fully open AI tool that spots lung findings on chest X-rays in seconds—early tests cut diagnostic turnaround and error rates, and any hospital can deploy it today. (News-Medical)
Shop-bot built-in. Shopify’s Summer ’25 Edition bakes the Sidekick AI assistant and natural-language catalog search directly into every store—boosting cart conversions out of the box. (Shopify Blog)
Off the Ledge
Hype and headaches we’re steering clear of.
Bias or boost? Depends who’s measuring. A new Warden AI audit of 150 hiring systems finds LLM screeners lower demographic bias on average, but 75 % of HR leaders still list bias as a top worry—proof the fairness debate is far from settled. (TheHRDirector)
Bubble vibes. Economists warn that today’s AI-powered market exuberance may already eclipse the late-’90s dot-com boom—one chart tells the story. (Business Insider)