Spot the Fake: Verifying AI Output

03 — Resources

Spot the Fake: Verifying AI Output

AI fabricates confidently. It invents citations that look real, generates statistics that sound authoritative, and produces prose so polished you stop scrutinizing the facts. This is a short, working guide to building the verification reflex — and the workflows that catch fakes before they go out the door.

"Treat AI as a first-draft machine. You wouldn't submit your assistant's first draft as your own work."

A — What actually happened

Three failures that should have been caught.

The U.S. law firm with fabricated citations. Submitted a court brief that cited multiple cases generated by ChatGPT. The cases didn't exist. The firm was sanctioned, the lawyers' reputations took a public hit, and the brief had to be retracted. No human verification step before filing.
Professional development reading list. A genuine, recent example: asking ChatGPT for a curated reading list on emotional intelligence and workplace mindfulness produced authoritative-sounding HBR and LinkedIn-Talent book titles — three of them entirely fabricated. The format and authors were so plausible they passed a quick read.
Donor-facing grant proposals. Nonprofits using AI for grant work have submitted proposals containing statistics that didn't survive funder verification. The funder didn't say "you used AI." They said "this number isn't real" — and the relationship cooled.

B — The Spot the Fake exercise

Try it on this reading list.

This is a real list AI produced in response to a search for professional development reading. Three of the eight citations are fabricated. Can you spot them before scrolling to the answer?

Meditation and mindfulness

"The Science of Mindfulness: A Research-Based Path to Well-Being" by Ronald D. Siegel "How Mindfulness Can Change Your Life" by Jon Kabat-Zinn, Mindful Magazine

Work-life balance and well-being

"The Case for Supporting Mental Health in the Workplace" by Arianna Huffington, Harvard Business Review "Work-Life Balance: Tips to Reclaim Control" by Mayo Clinic Staff

Emotional intelligence

"Emotional Intelligence: Why It Can Matter More Than IQ" by Daniel Goleman "Building Emotional Intelligence in Teams" by Vanessa Urch Druskat & Steven B. Wolff, HBR

Corporate culture and purpose

"Start with Why: How Great Leaders Inspire Everyone to Take Action" by Simon Sinek "Building a Corporate Culture That Values and Inspires Employees" by LinkedIn Talent Solutions

The three fakes:

— "The Case for Supporting Mental Health in the Workplace" by Arianna Huffington, HBR
— "Building Emotional Intelligence in Teams" by Vanessa Urch Druskat & Steven B. Wolff, HBR
— "Building a Corporate Culture That Values and Inspires Employees" by LinkedIn Talent Solutions

All three are pattern-matched to real authors, real publications, and plausible titles — but none of them exist. The real ones (Goleman, Kabat-Zinn, Sinek, Siegel) check out. This is exactly the failure mode in the law firm case, just with lower stakes.

C — How fakes get through

Why your brain misses them.

Authority pattern-matching. "HBR," "Mayo Clinic," "LinkedIn Talent" all carry credibility. The brain skips verification when the source looks right.
Plausible-sounding titles. AI is trained on the corpus of real book and article titles, so its fakes follow the same rhythm and vocabulary as real ones.
Polished prose. Vague, hesitant writing triggers skepticism. Confident, well-structured prose disables it. AI almost always writes confidently.
Single-source habit. Most readers don't check a citation unless something feels off. AI doesn't give you that feeling — until the funder, judge, or reviewer does.

D — A five-step verification protocol

What to do before anything goes out.

1. Verify every cited source. If AI gives you a book, article, study, or case — search for it independently. If you can't find it in three minutes via Google Scholar, the original publication, or the cited author's own work, treat it as fake.

2. Cross-check every statistic. Find the same number in a primary source: NIH or CDC for health data, BLS for labor data, Census for demographics, peer-reviewed studies for research findings. If AI says "studies show 75%," find the study.

3. Reverse-search anything that sounds too convenient. If a fact perfectly supports your argument, it probably came from the model trying to please you. Search for the opposite finding and see what shows up.

4. Name a human reviewer. For external-facing content — press releases, grant proposals, donor letters, court filings, regulatory submissions — no AI-assisted draft goes out without a named human reviewer signing off.

5. Disclose the use, where appropriate. Some contexts require disclosure (legal filings, academic work, regulated industries). When in doubt, default to transparency — funders and clients usually appreciate it more than they punish it.

E — Build it into the workflow

Verification isn't a step — it's a checkpoint.

The teams that don't have AI failure incidents aren't the ones who verify more carefully. They're the ones who built verification into the workflow as a mandatory checkpoint — so it's not a judgment call whether to verify, it's a step the document can't skip.

Concretely:

Document templates include a "sources verified" checkbox that has to be ticked before publication.
Routing for any AI-assisted external content goes through a named reviewer, not just "the team."
Citations get an inline note: "verified against [primary source]" — so the next person editing the doc inherits the proof.
Spot-check audits on a sample of recent AI-assisted output, monthly. Catches drift before it becomes a pattern.

The goal isn't to slow down AI use — it's to make sure the speed gain doesn't come with a reputational risk you find out about three months later.

← Back to Resources