INTELLEGIXNEWS ▶ Reels

Get news alerts

A notification when a new edition publishes.

AI Tutors Are Acing Homework and Failing Tests — and Readers Are Going Blind

Ask about this with Perplexity AI-written from the broadcast
▶ The reel · AI-generated from this story · watch full screen ↗
How this was made Verified AI

Every Intellegix briefing is generated from that day's broadcast and run through automated checks before it publishes — with a human paged on any flag. Here is the trail for this edition.

Sources 12 sources traced for this edition Traced
Guardrail Every figure and proper name traced back to the broadcast Pass
Fact-check 3 confirmed · 3 checked against live web sources Verified
Human loop Operator paged on every flag before publish On
A student sitting alone at a desk in an empty classroom taking a written exam.
Photo: SofiaJone1 · pixabay

Research published by The Economist this week — generating 324 comments on Hacker News — documents a stark divergence: AI tools measurably boosted students' homework scores while their exam scores dropped. The mechanism is not complicated once named. Homework is where the cognitive work of encoding new information occurs; exams measure whether that encoding happened. When AI completes the homework, the encoding does not. Students arrive at exams with a record of strong performance and without the underlying knowledge.

Teachers in the Hacker News thread described the phenomenon from direct classroom observation: students who use AI fluently to complete assignments struggle to explain basic concepts in follow-up conversations. Researchers have warned about analogous effects with every successive educational technology — calculators, Wikipedia, search engines — but this time the effect size appears large enough to be unmistakable in the data.

The cultural companion piece trending alongside the research was an essay titled 'I'm Becoming AI-Blind,' whose 382-comment thread was described as one of the richest conversations on Hacker News in weeks. The author describes a gradual, involuntary shift: reading a paragraph and automatically pattern-matching for AI tells rather than engaging with the content. The essay argues that the ambient presence of low-friction AI output has subtly degraded the scarcity signal that once gave writing its weight — if everything can be generated instantly, the fact that someone spent real time on a piece of prose ceases to function as a marker of worth.

A GitHub project called Claudette, which drew 288 upvotes, treats that degradation as an engineering problem. Its goal is to strip what the creator calls 'BuzzFeed-style' language patterns from AI outputs — performative enthusiasm, reflexive hedges, listicle structures — and replace them with plainspoken prose. The appetite for a tool that makes AI communication stop sounding like AI communication functions as a market signal: the default outputs are being perceived as a recognizable, gameable form of noise.

An ACM profile of Go technical lead Russ Cox and a personal essay by legendary security researcher Thomas Dullien — 'Three Important Steps in My Maturation Process,' reflecting on learning to hold uncertainty rather than project fluent confidence — both surfaced in the same feed, offering an implicit contrast to AI systems that are always complete-sounding and never visibly unsure.

▶ Listen to this story