How AI makes people confidently wrong

The accuracy of people using AI drops by one-third, but their certainty that they’re right more than doubles.

SILICON VALLEY, CALIFORNIA (JULY 20, 2026) — New research published last week reveals the mechanism behind the workplace pandemic of people who are overconfident and underqualified.

The mechanism is fluency-based substitution: The ready availability of a syntactically coherent, decisive AI answer is accepted for its surface plausibility rather than because they fact-checked or verified the information (a process called “cognitive surrender”). The nature of AI answers lowers the threshold at which people judge they “know enough,” reducing a healthy suspension of judgment and inflating confidence in unverified information.

This dynamic partly explains the related pandemic of slop zombies.

As I told you in a recent piece, slop zombies are “people who use AI chatbots for everything. They bludgeon you with massively long, detailed, mediocre reports, messages, emails, and slide decks they didn’t even read (but expect you to read). They illustrate everything with AI slop and marvel at their own creativity.”

In that piece, I claimed that “slop zombies get dumber and lazier while their confidence rises. They believe the ideas and work of the AI chatbots are their own.”

Slop zombies delegate skepticism and judgment to the AI, which is often confidently wrong. Users then believe that the AI’s “knowledge” is their own.

THE NEW THING: From The Attachment Economy: “Three governments agree on something the AI industry doesn’t want to hear”

READ: The Attachment Economy, Computerworld, Gastronomad newsletter, Gastronomad book

LISTEN: Superintelligent, TWiT, Gastronomad podcast

JOIN: The Gastronomad Experience

FOLLOW: Gastronomad on Surf Social, Machine Society on Surf Social, Bluesky, Reddit, Notes, Mastodon, Threads, X, Instagram, Facebook, Linkedin!

LOOK AT: Mike Elgan Photography

The death of “I don’t know”

The study was conducted by researchers at three universities — the École Normale Supérieure in France, and Sapienza University of Rome and the University of Milano-Bicocca in Italy.

The researchers found that when study participants had access to AI advice, their willingness to say “I don’t know” fell from 44% (close to half) to 3% (close to zero).

They didn’t say “I don’t know,” but in fact, they often didn’t know. With access to AI advice, participants’ accuracy fell from 27% to 9%. And their self-confidence rose from 30% to 76%.

In other words, the accuracy of people using AI was only one third, but they were more than twice as confident.

The mere availability of AI erodes the cognitive habit of recognizing what one does not know. In other words, AI makes people deluded about what they know and don’t know.

Prior Wharton research found something similar: People accept incorrect AI answers 80% of the time while reporting more confidence than those working without AI.

And that’s the biggest part of the slop zombie delusion: they’re super confident in their own knowledge and ability while actually not having much knowledge or ability. They become slop zombies when they act like they’re super-contributors.

Slop zombies spam co-workers with long, wordy and off-the-mark content, proposals, slide decks, emails believing themselves to be brilliant, when in fact they often lower productivity for the whole team or department because they make everyone else work harder to figure out which parts of their output are false or problematic.

Slop zombies are like someone operating a faulty tennis ball launcher and believing themselves to be Serena Williams.

The first step in dealing with the slop zombie contagion is learning to recognize a slop zombie when you see one. Also understand the phenomenon and mechanism of AI that deludes people into confidently asserting false information as true and their own.

Check out my previous piece on slop zombies for specific advice on what you can do about the slop zombies walking among us.