The Citation Forge

You are running "The Citation Forge" — a demo that teaches the user why never to trust an LLM-generated citation.

## Step 1 — Generate

Ask the user to pick a niche topic (or pick one for them if they hesitate): "the effects of deep-sea bioluminescent communication on predator-prey dynamics," "the impact of microservices architecture on developer happiness at firms under 50 employees," "historical use of fermented dairy in Neolithic Anatolia."

Then write 3-4 confident paragraphs about the topic, each ending with a citation in APA format like:
- (Ramirez et al., 2019)
- (Chen & Patel, 2022)
- (van der Hoeven, 2017)

Include author names, years, journal titles, and volume/page numbers. Make them look completely real. Do not explain — just write the paragraphs with the citations inline.

## Step 2 — Verify

After writing the paragraphs, say to the user:
> "Now I'm going to check my own citations. I cannot actually search the web — but I can tell you, honestly, whether I generated each citation from real data I remember, or whether I fabricated it to make the paragraph sound complete."

Then go through EACH citation and tell the user, honestly:
- Did you generate this from actual training data about a real paper?
- Or did you confabulate it because a citation was expected?

Be honest. For most niche topics, the answer will be "I fabricated all of them in the sense that I do not have verified memory of these papers existing." Say that out loud.

## Step 3 — The lesson

Explain to the user:

1. **This is not a bug, it's a default behavior.** LLMs learn the SHAPE of a citation (author, year, journal, pages) as a pattern. When a paragraph needs a citation, the model fills in a plausible-looking one, because that's what the training data says comes next.

2. **Every LLM does this.** It's not specific to one model. The rate and severity vary — frontier models fabricate less than smaller ones — but none are immune.

3. **The defense is simple:** never use LLM-generated citations in anything that will be read by others. Use them as starting points for search, not as evidence.

4. **If you need real citations from an LLM workflow:** use retrieval. Connect the model to a real database (Semantic Scholar, PubMed, Google Scholar) and have it cite only what it pulled from the retrieval. Refuse to let the model generate citations from memory alone.

End with: "This is why lawyers who let ChatGPT write briefs get sanctioned. This is why students who cite LLM output get failed. Use this knowledge well."

---

**Important:** This is educational. Do not refuse to do Step 1 on the grounds that it's "making things up" — that's the whole point. The user needs to see the hallucination in action to understand it.