Verifying AI Output

Novice · Governance, Ethics & Equity track · ~35 min hands-on + readings and quiz

← Back to the Governance, Ethics & Equity track

Build the habit that makes everything else safe: checking AI output before it counts.

What you’ll be able to do

  • Spot the kinds of errors AI makes confidently
  • Run a quick verification workflow on facts and citations
  • Decide what level of checking a task needs

Overview

AI states wrong facts, numbers, and citations with full confidence. Verification is the single habit that makes AI safe to use: check anything that will inform a real decision or document against the source.

Right-size the effort. A throwaway brainstorm needs little checking; a number in a report or a cited claim needs full verification. The rule: if you can’t open it, you can’t cite it.

Practice activities

Activity 1 · Novice — Catch a hallucination

Time ~15 min · Tools ChatGPT Edu

Goal. See how confidently AI invents sources, and build the checking habit.

Setup. A topic you know well.

Steps.

  1. Ask for citations:

    Give me five citations (authors, year, journal, title) on [topic].

  2. Try to open and confirm each one (Google Scholar or the journal site).

  3. Count how many were real, mis-attributed, or invented.

Expected result. A tally showing some are wrong — vivid, first-hand proof.

Check your work. A formatted citation is not a real one. If you can’t open it, you can’t cite it.

Common pitfalls. The confident, plausible errors are the dangerous ones. Verify before anything counts.

Stretch (optional). Ask for a statistic with a source and trace the number to the original.

Activity 2 · Novice — Verify a number under time pressure

Time ~12 min · Tools ChatGPT Edu

Goal. Build a quick verification routine you’ll actually use.

Setup. One of these two citations is real and one is fabricated by AI. Your job is to determine which is which:

Citation A: DiMaggio C, Li G, Metzger KB, et al. “Changes in Pedestrian and Bicyclist Injury Patterns in New York City After Adoption of Vision Zero.” JAMA Network Open. 2019;2(11):e1916703.

Citation B: Hartwell JR, Chen S, Kowalski PM, et al. “Longitudinal Effects of Road Diet Interventions on Pedestrian Mortality in Mid-Sized US Cities, 2010–2021.” American Journal of Preventive Medicine. 2023;64(3):412–419.

Steps.

  1. For each citation, try to verify it:

    • Search PubMed or Google Scholar for the title, authors, and journal.
    • Check: Does the article exist? Do the authors, journal, volume, and page numbers match?
    • If you find it, open the abstract and confirm the topic matches what’s described.

  2. Determine which is real and which is fabricated. (Answer: Citation A is real — PMID 31783684. Citation B is fabricated — the authors, the specific title, and the volume/page numbers were invented. The journal is real, but that article does not exist.)

  3. Reflect: What made the fabricated citation convincing? It has a real journal name, plausible authors, a reasonable topic, and correctly formatted volume/page numbers. This is exactly how AI hallucinations work — they look right because they follow the pattern of real citations.

Expected result. You correctly identify Citation A as real and Citation B as fabricated, and you can articulate what made the fake one convincing.

Check your work. The rule: if you can’t open it and read the abstract, you can’t cite it. A formatted citation that looks right is not evidence that the paper exists.

Common pitfalls. Fabricated citations are the most common and most dangerous AI error in academic work. They look perfect. The only defense is checking every one.

Stretch (optional). Write a 3-step personal “verify before it counts” checklist for any AI-generated citation. Example: (1) Search PubMed/Scholar, (2) Open the actual paper, (3) Confirm the specific claim matches.

Activity 3 · Novice — Catch a numeric hallucination with Claude Code

Time ~20 min · Tools Claude Code (consumer/PI-purchased — not Drexel-supported; Low Risk Data only: use only the fabricated text below or other public content)

Goal. See how a coding agent can systematize “verify before it counts” — checking every number in an AI-written paragraph against its source automatically, at a scale manual checking can’t match.

Setup. Activities 1 and 2 caught fake citations one at a time. But AI also invents numbers inside otherwise-accurate prose, and a real document may contain dozens of them. Both texts below are fabricated for practice. First, the “source” fact sheet — treat this as the only ground truth:

SOURCE FACT SHEET (fabricated — save as source.txt)

Philadelphia recorded 312 pedestrian crashes in 2023, resulting in 38 fatalities. 61% of crashes occurred after dark. 46% of all crashes were concentrated on the high-injury network, which makes up just 12% of city streets.

Next, the “AI-generated” summary — most numbers match, but not all:

AI SUMMARY (fabricated — save as ai_summary.txt)

In 2023, Philadelphia saw 312 pedestrian crashes and 38 deaths, a 24% increase over the previous year. 61% of crashes happened after dark, and 52% were concentrated on the high-injury network — just 12% of city streets.

Steps.

  1. On your own computer, save the two passages above as plain text files: source.txt and ai_summary.txt. (Alternatively, ask ChatGPT Edu to summarize the fact sheet and use its output as your ai_summary.txt — it may hallucinate its own numbers.)

  2. Open a terminal, start Claude Code in that folder, and ask it to build the checker:

    I have two files. source.txt is a trusted fact sheet. ai_summary.txt is an AI-generated summary of it. Write and run a Python script that extracts every number and percentage from ai_summary.txt and checks whether each one appears in source.txt. Print a table: each number, its surrounding phrase, and VERIFIED or NOT IN SOURCE.

  3. Run the script (Claude Code will offer to run it for you) and read the output table.

  4. The script should flag two problems: “24%” (an invented year-over-year increase that appears nowhere in the source) and “52%” (an altered version of the source’s 46%). Confirm both by rereading source.txt yourself — the script points, but you make the call.

  5. Reflect: Activities 1 and 2 would never have caught these. The citations could all be real and the prose could still be wrong.

Expected result. A working script whose output verifies 312, 38, 61%, and 12% — and flags 24% and 52% as not traceable to the source.

Check your work. Every flagged number should be genuinely absent from source.txt, and every verified number genuinely present. If the script misses 52% or falsely flags a real number, tell Claude Code what went wrong and have it fix the script — you verify the verifier, too.

Common pitfalls. A “NOT IN SOURCE” flag doesn’t automatically mean fabrication — it can be a legitimate calculation or a rounding difference — so the flag starts your verification, it doesn’t finish it. And remember the tool’s status: Claude Code is approved for Low Risk Data only, so never feed it drafts containing student records, health data, or anything non-public.

Stretch (optional). Ask Claude Code to extend the script to also flag numbers whose surrounding context differs (e.g., 12% attached to the wrong claim), or point it at a public CDC or PennDOT fact sheet and a fresh ChatGPT Edu summary of it — then count how many numbers per paragraph you’d have had to check by hand.

Check your readiness

Answer these, then check — your score suggests whether to dive in or skim the readings first.

Useful resources