Skip to content
Redline Station
Supply Shop
Derailer Risk

Fabrication Risk

The pattern in one line

The AI produces claims that look authoritative but are not grounded in anything verifiable.

See the full pattern page

· Reading·

Book · 2020

Calling Bullshit

Carl Bergstrom and Jevin West

Why: Bergstrom is a UW biologist; West is a UW data scientist. Together they teach a course called Calling Bullshit, and the book is the textbook. The chapter on "data visualization that lies" and the one on "fake numbers in stories" both apply directly to AI fabrication. Their core technique: lateral verification. Don't read deeper; read sideways. If the AI cites a study, the question is whether the study exists, not whether it sounds plausible.

Book · 2017

The Death of Expertise

Tom Nichols

Why: Nichols, a US Naval War College professor, writes about how the internet collapsed the distance between expert and amateur claims. The relevant move for fabrication: when everyone can produce authoritative-sounding text, authority-language stops being a signal of authority. AI inherits the same problem. "Studies show" is just a string of three words; whether there are studies is a separate question.

Book · 2016

Weapons of Math Destruction

Cathy O'Neil

Why: O'Neil, a former Wall Street quant turned algorithmic-accountability advocate, traces how opaque models produce confident-sounding outputs whose underlying logic is unverifiable. AI fabrication is the conversational version: the output is presented as a fact, the path from input to output is unauditable, and the user is expected to accept the result. Reading O'Neil sharpens the habit of asking what's in the box before accepting what comes out of it.

Book · 2023

Verified

Mike Caulfield and Sam Wineburg

Why: Caulfield (Washington State University) and Wineburg (Stanford) developed lateral reading as the practical method for verifying online claims. Their technique: when you're unsure about a source, leave the source and check what other sources say about it. The book translates the method into specific habits. Applied to AI, the practice is the same: don't ask the AI to verify itself; check what other sources say about the claim.

· Questions to sit with·

  1. 1. The last time an AI cited a "study" or "research" — did you check whether it existed?
  2. 2. When the AI gives you a percentage or statistic, what's your default? Trust, doubt, or check?
  3. 3. Have you ever shared something an AI told you and then found out the source was made up?
  4. 4. What kinds of claims are you most likely to take from the AI without checking? Do those map to your areas of expertise, or away from them?
  5. 5. If the AI started saying "I'm not sure where I got this" when warranted, would you trust it more or less?

· Practices·

Lateral check

When an AI cites a study, statistic, or source, leave the AI conversation and search for the source independently. If it doesn't exist or doesn't say what the AI claimed, the citation was fabricated.

Drawn from · Caulfield and Wineburg

Source specificity

Train yourself to read for specifics. "Studies show" is fabrication-shaped. "A 2022 paper by Smith and Jones in Nature Methods titled X" is verifiable-shaped. Notice which one the AI offered you.

High-stakes rule

For any claim you might act on (medical, legal, financial, professional), assume the AI may have fabricated until you've verified independently. The cost of a check is small; the cost of acting on a fabrication can be large.

Drawn from · O'Neil

Default doubt

For one week, treat every confident-sounding AI claim with explicit skepticism. Look up at least one fact per session. Notice how often the lookup contradicts the AI.

· When to bring someone else·

Fabrication risk becomes worth naming to a person when you've made a real decision based on an AI claim that turned out to be made up. When you've shared something the AI told you and had to retract it. When you can't tell from inside a conversation what's real and what's been generated to sound real. The station doesn't say the AI is always wrong. It says when authority-language is decoupled from verifiable sourcing, the authority is doing work the sourcing should be doing, and that work falls on you when the AI is wrong.

Supply Shop resources are orientation, not prescription. The station points toward material others have found useful; how it fits is the visitor's to decide.

Station Pass