In a much-discussed Washington Post column, journalist Sally Jenkins recounts how she asked ChatGPT to analyze one of her own recent articles — a profile of tennis rising star João Fonseca.
The column (‘ChatGPT couldn’t answer my questions about tennis, so it made things up‘) describes how she was getting suspicions about ChatGPT’s truthfulness after wrong answers and reading about the experiences of other writers. So Jenkins
“began by asking Sage to do a quality analysis on a recent piece I had written about Fonseca,” she writes.
What followed startled her. The AI returned quotes she had never written, including made-up lines it claimed were hers. Her conclusion: ChatGPT “doesn’t just make mistakes. It lies.”
I tried it too — and got fake quotes
Out of curiosity, I replicated Jenkins’s experiment. I gave ChatGPT the headline and link to her Fonseca article and asked for a quality analysis.
Sure enough, I got fabricated quotes too — vivid, plausible-sounding phrases that were not in the original article.
But when I pasted the full article text directly into the prompt and specified that it should quote only what was present, the hallucinations stopped entirely.
This highlights two crucial — and often misunderstood — causes of AI hallucination.
Two root causes of AI hallucinations
1. Software-level: ChatGPT is a language model, not a fact engine
ChatGPT is designed to complete patterns in language. If you ask it for an analysis of a text without actually giving it that text, it fills in the blanks with what sounds right — not what is right.
Jenkins, like many first-time users, likely assumed that by referencing her article, the model would “look it up.” But unless she pasted the full article or uploaded it in a usable format, the model couldn’t access it.
Lesson: Without specific, visible input, the AI will generate content based on style and probability — not truth.
2. Technical: input format matters more than most users realize
There’s a major difference between input methods — and each comes with its own risks:
- A link: Even when browsing is enabled, links are an unreliable source. Many modern web pages are dynamically structured — using JavaScript, modular loading, cookie prompts or paywalls. This means the content may be:
- 🔄 Fragmented across different sections (title, body, metadata),
- 🔐 Hidden behind consent walls or pay-per-view barriers,
- 🧩 Rendered in a way that breaks reading order or excludes key elements
Bottom line: even when technically accessible, webpages often cannot be reliably parsed by AI tools. The risk of distortion, omission or hallucination remains high.
- A PDF: Under the right conditions, PDFs can work — but they must be:
- ✅ Clean and OCR-readable (not a scan or image file),
- ✅ Within token limits (roughly under 100 pages, often less),
- ⚠️ Free from layout issues like multi-column text, headers, or footers, which can disrupt reading order and mangle content.
- Full plain text in the prompt: This remains the most reliable method. It ensures the model sees exactly what you see — without truncation, distortion, or guesswork.
Use: “Here is the full article. Only reference direct quotes.”
What happened when Jenkins challenged the response
Jenkins didn’t just receive a hallucinated quote — she pushed back. After ChatGPT attributed a description to her article that included Fonseca “punching his strings,” she replied: “I never described him punching his strings.” And rightly so — the quote was fictional.
But instead of correcting itself, the model offered another polished but entirely made-up explanation. It praised Jenkins for portraying a duality: Fonseca as fiery on court and charming off it, referencing invented quotes like “yelling at himself” and “talking about liking to smile and do crazy stuff.”
This wasn’t the model being dishonest — it was doing what it’s designed to do: maintain conversational coherence, not factual accuracy.
When confronted, ChatGPT doesn’t access memory or validate against a source. It builds a new reply from prior conversation tokens, using stylistic probability. The result: layered hallucinations. Jenkins saw it as the AI “squirming,” but in reality, it was just confidently continuing down the wrong path.
How to prevent hallucinations — four practical strategies
Whether you’re a journalist, lawyer, researcher, or analyst, here’s how to get reliable, verifiable results:
- ✅ Provide the full text — and beware of format traps Paste the full article into the prompt as plain text. PDFs can work if they are clean, OCR-readable, and within size limits — but links alone are unreliable.
- ✅ Separate tasks into clear steps Split your query into summary, citation, and interpretation. Use: “Step 1: summarize. Step 2: list direct quotes. Step 3: analyze tone.”
- ✅ Require citation discipline Ask the model to flag unverifiable content. Use: “If a statement is not found in the text, label it as unverifiable.”
- ✅ Avoid broad, open-ended prompts Don’t ask “What do I say about X?” Ask: “In paragraph 4, what is said about X?”
This mirrors legal experiences
Jenkins’s experience is not unique to journalism. Lawyers using ChatGPT to summarize decisions or respond to court filings increasingly report:
- Misquotes from rulings,
- Invented case law,
- Or errors in legal reasoning that sound persuasive but aren’t grounded in fact.
Like journalists, legal professionals must learn to use AI as a tool, not an oracle. That means:
- Providing clean, structured input,
- Controlling the format,
- And demanding traceable output.
Final thought
Sally Jenkins was right to be concerned. But ChatGPT didn’t lie — it improvised. And like many first-time users, she didn’t realize that AI can only analyze what it’s explicitly shown — and only in formats it can properly read.
The fault isn’t in the machine — it’s in the assumption that the machine sees what you see.