How AI Writes a Product Description (and Where It Goes Wrong)
An AI writing tool doesn't look up your product. It predicts words. That single fact explains most of what's good and most of what's bad about a draft.
What the tool is doing
The tools behind many modern chatbots are called large language models. A large language model is a program trained on a vast amount of text, so it can generate language1. One common type, the generative pre-trained transformer, is first trained to predict the next word and is then often fine-tuned to follow instructions and act as an assistant1.
So when you type "write a description for my linen tea towel", it doesn't open a catalogue. It builds the text that usually follows a request like that. Most tea towel descriptions mention softness, absorbency and a washing tip. Yours may too, whether or not they're true.
The quality of that training matters as well. Biased or inaccurate training data can make a model's output less reliable1.
Where it makes things up
Researchers call this a hallucination. It's a response generated by AI that contains false or misleading information presented as fact2. Chatbots can embed plausible-sounding random falsehoods in what they write2.
"Plausible-sounding" is the trouble. A made-up detail reads exactly like a real one. Nothing flags it.
In a product description, the usual guesses are:
- Material (pure cotton, when yours is a blend).
- Size, weight or capacity.
- Where the item was made.
- Care steps, such as "machine washable".
- Features the product never had, borrowed from similar products.
One source notes that in chatbots, customers tend to prefer rapid, overconfident answers over cautious, uncertainty-aware ones2. So a confident tone is not evidence that a detail is correct.
Name only versus facts supplied
Here's how the two ways of asking compare.
| Product name only | Your facts pasted in | |
|---|---|---|
| What the tool has | A guess at what such products usually are | Only the details you gave it |
| Gaps | Filled with typical wording | Fewer gaps, though it can still add extras |
| Risk | High: invented specifics read as real | Lower, but still needs a line-by-line check |
| Suits | Brainstorming rough ideas only | Drafts you plan to publish |
Even with facts pasted in, a tool can add a sentence you never gave it. Treat every draft as unchecked.
Why the shop carries the risk
The tool isn't the one selling. You are. In Australia, the ACCC says business information about products must be accurate, truthful and based on reasonable grounds. That includes images and descriptions of what is offered3. It adds that it makes no difference whether a business intends to mislead or not3.
That's Australian consumer law, and rules differ in other countries. Still, the habit travels well: a sentence you publish is a sentence you stand behind.
The next pages cover what to paste in ([page:what-to-give-the-tool-before-you-ask]) and how to keep one reference sheet ([page:product-fact-sheet-template]) to check against. This site gives general information only. For a specific legal question, ask a lawyer or your regulator.
Cited material
- Large language model, Wikipedia
- Hallucination (artificial intelligence), Wikipedia
- False or misleading claims, Australian Competition and Consumer Commission (ACCC)