AI for Seniors

Why AI Sometimes Makes Things Up (and How to Spot It)

By AI for Your Day · May 5, 2026

Share

---

Some links below are affiliate links — I may earn a small commission, at no extra cost to you, if you sign up for a tool I recommend. More on how I choose what to recommend.

A real story: in 2023, a New York lawyer filed a legal brief in federal court that cited six prior court cases. The judge looked them up. None existed. The cases were entirely made up. The lawyer had used ChatGPT to write the brief and trusted what it gave him.

He was sanctioned. The story made national news. It became the cautionary tale every law firm now uses to explain why you can’t just hand AI work to an associate to file.

This is not a rare problem. It is the single most important thing to understand about AI.

What “hallucination” actually means

When AI is asked something it doesn’t actually know, it doesn’t say “I don’t know.” It generates a plausible-sounding answer that may or may not be true. The industry calls this hallucination.

The reason it happens is built into how AI works. AI doesn’t store facts the way an encyclopedia does. It learned patterns of how language flows — which words tend to follow others, how citations look, how a doctor’s note is structured, how a recipe is written. When you ask a question, it generates the most plausible sequence of words. If real facts happen to fit that plausible sequence, great. If not, it generates plausible-sounding facts that don’t exist.

The result: an AI will often write something that sounds completely correct but contains made-up details. A made-up court case. A made-up source. A made-up quote attributed to a real person. A confidently wrong year for a historical event. The AI doesn’t know it’s wrong because it doesn’t actually “know” anything — it pattern-matches.

What this looks like in real life

The most dangerous hallucinations are the ones that look most innocent. Some real categories:

• ⚖️Citations — court cases, academic papers, news articles. Often plausible-sounding case names, plausible-sounding paper titles, plausible-sounding URLs that don't exist.
• 📅Dates and years — confidently wrong years for historical events, especially less famous ones. (Anything from before 1850 or outside US/UK is high-risk.)
• 💬Quotes — quotes attributed to real public figures who never said them, often with the right "voice" so they sound authentic.
• 🔢Statistics — confident percentages and numbers attributed to real-sounding studies that may not exist.
• 💊Medical specifics — drug interactions, dosages, diagnoses based on symptoms. Even when the AI is mostly right, the details that matter most are the ones most likely to be off.
• ⚖️Legal specifics — what your specific state's law says, what the deadline for a specific filing is, what a contract clause actually means.
• 💰Financial specifics — tax rules, investment specifics, plan benefits.
• 👤Biographical details — schools attended, jobs held, dates of birth/death, especially for people who aren't ultra-famous.

When hallucination actually matters

Here’s the part most articles miss: hallucination is only a real problem in some uses of AI. For others, it almost never causes harm.

High risk — verify everything:

• 🚨Anything you'll send to a professional (lawyer, doctor, accountant) as if it were research
• 🚨Anything that will be published or shared publicly
• 🚨Anything where being confidently wrong would harm someone
• 🚨Anything involving health, legal, or financial decisions
• 🚨Specific facts you'll repeat as if you knew them yourself

Low risk — relax:

• 🟢Brainstorming ideas (you'll evaluate them anyway)
• 🟢Drafting an email or a message that you'll edit before sending
• 🟢Summarizing something you can spot-check by reading the original
• 🟢Asking for explanations of concepts (you'd notice if it were way off)
• 🟢Talking through a decision (you're the one deciding)

The same AI that confidently invents court cases is also great at writing your dinner-party invitation. The category of task is what matters, not the tool.

The 30-second sanity check

When the task is high-risk, you don’t have to fact-check the entire response. Run this small set of checks on the parts that matter:

The whole loop takes about 30 seconds for the parts that matter. It’s faster than the time saved by using AI in the first place.

Why AI sometimes admits it doesn’t know — and sometimes doesn’t

A pattern worth noticing: ChatGPT, Claude, and Gemini have all gotten better at saying “I’m not sure” when they don’t actually know something. But they’re not perfect at this, and the gap between confident-correct and confident-wrong looks the same to you. There’s no visual cue.

Some things that increase the chance of hallucination:

• 🌐Niche topics — AI is great on topics with lots of training data; it gets shakier on obscure ones
• 📜Anything historical or pre-internet — citations and details for older events are often patchy
• 🌍Non-Western contexts — Western, English-language sources dominated training data; quality varies for other regions
• 🆕Recent events — anything more recent than the AI's training cutoff is invented unless the tool can search the web in real time
• 🔢Specific numbers — exact percentages, exact dates, exact dollar amounts. If precision matters, double-check

Three ways to lower the hallucination rate

You can’t eliminate hallucinations, but you can reduce them noticeably with three simple habits.

A useful mental model

Think of AI as a brilliant intern who has read everything but has zero personal experience. They can write you anything you ask for, articulate, well-structured, fluent. But when you ask them about a specific case file or a specific lab result or a specific policy, they will sometimes invent details that sound right because their job is to sound right, not to know.

You wouldn’t have an intern file court documents without checking them. Don’t have AI do it either.

For everything else — the brainstorming, the drafting, the explaining, the summarizing, the thinking-out-loud — the intern is great. Use them every day. Just don’t trust them with the details that matter without a 30-second check.

---

Caught a hallucination that nearly burned you? We collect reader stories — email help@aiforyourday.com.