An LLM Is Not a Search Engine

Most people first try an LLM by asking it a factual question — "who won the league last season?", "what's the latest version of React?", "what is the capital of X?". Sometimes the answer is right. Sometimes the answer sounds right but isn't. That second case is dangerous, and it's worth understanding why.

What the model is, and isn't

A search engine looks up live web pages. An LLM doesn't. The model has been trained on a snapshot of text up to some cutoff date, and it generates the most likely continuation from that frozen knowledge. Three implications:

1. Recent facts will be wrong or missing. Anything that happened after training is invisible.
2. Niche or local facts may be outdated. "Who's the head roaster at a Cairo coffee shop founded in 2024?" — the model may have never seen that.
3. The model never says "I don't know" by default. Generating fluent text is its strength, even when the underlying knowledge is shaky.

A capable model like Claude Sonnet 4.5 will often hedge correctly when asked about a recent event:

"I don't have access to real-time information, so I can't tell you who won the most recent Egyptian Premier League title. My knowledge was last updated in April 2024, and league seasons continue beyond that date."

That's the good failure mode — explicit hedging. The bad failure mode is when the model confidently invents a name, a date, or a statistic. That's what people call "hallucination", and it's the single biggest reason early prompters lose trust.

Three rules to stay safe

When you actually need a fact, follow these:

Rule	Why
Use a tool that can search the live web when freshness matters.	The model's training cutoff is real.
Ground the model with the source text if you have it. Paste the doc; ask it to answer only from that.	Removes guessing.
Ask the model to say "I don't know" when the answer isn't in the source.	A trained model can hold that line if you tell it to.

Module 8 will go deep on the third rule (it's called grounding). For now, the lesson is simpler: when you ask the model a factual question, treat the answer as a draft you have to verify, not as the final word.

What LLMs are great at instead

Use them for the things they're actually good at: writing, rewriting, summarising, classifying, extracting structure, drafting code, brainstorming, translating, restructuring messy notes. These are tasks where the model's language skill matters more than its memory.

A working rule of thumb: if the answer depends on truth that lives in the world today, verify it. If the answer depends on shaping or transforming text you already have, the model is your best tool.

Next: the four sub-skills you're actually building across this course. :::