Explain why default AI writing is generic, and spot its common patterns.
Farah runs a modest workwear shop in Singapore (an invented example). She wanted copy for the shop's homepage, so she typed a single line into an AI assistant: "Write a homepage headline and intro for a modest workwear brand in Singapore."
Within a few seconds she had a headline, "Redefine your everyday elegance", and an intro about "timeless pieces crafted for the modern woman who effortlessly balances style and substance". The writing was fluent, with no grammar mistakes. It also sounded exactly like the homepages of three other shops she had browsed that week, and she could not point to one line her customers had ever actually said.
ChatGPT, Claude, Gemini and similar assistants run on a kind of AI called a language model. A language model is trained on a very large amount of text. When you type a prompt, it builds the reply in small pieces, each time predicting what is likely to come next given everything it saw in training.
Fluency comes from this prediction, and so does the generic feel. If your prompt is vague, the likeliest continuation is the most typical text of that kind. Ask for a workwear homepage with nothing else to go on and you get something near the average of every workwear homepage and piece of fashion copy the model learned from, and an average of many homepages does not sound like any one business.
None of this is laziness on the model's part. It carries out the request using whatever information it was given. When the prompt is one line, the only real information is the category, so the model fills every gap with what is typical for that category.
AI drafts have habits in common, and learning to recognise them is the first step to taking them out. Five turn up often.
Stacked adjectives are several adjectives in a row with none of them specific, as in "timeless, versatile and effortlessly elegant pieces".
Lists of three group three items, usually abstract, often in a rhythm that sounds good and means little. "Style, comfort and confidence" and "quality, care and community" are typical.
Dramatic contrasts use a setup and a twist to make an ordinary claim sound deep: "It's not just a blouse, it's a statement", or "More than tuition, a partnership".
Claims with no specific detail could be about anything and cannot be checked, such as "trusted by thousands", "designed with you in mind" and "exceptional quality".
Favourite words appear far more often in AI drafts than in ordinary writing. "Redefine", "effortless" and "timeless" are examples, and Farah's draft used all three within its first two sentences.
Any one of these, used once, is fine. People write lists of three and the odd contrast all the time. When the patterns pile up, readers sense that nobody in particular wrote the text, and they stop trusting it.
To find them, read the AI draft one line at a time, test each line against the five patterns, and mark every generic phrase. You will act on what you mark in the rest of the module.
An assistant has broad knowledge of the world and knows nothing about your business beyond what you tell it. Your customers, the words they use, your prices, your delivery times and your reviews are all unknown to it. So is what sets you apart from the shop down the road, and what went wrong last month.
Missing all of that, the AI can either stay vague or make something up. It usually stays vague, and when it does invent, the result is worse than vagueness.
Farah's draft described the blouses as "made from sustainably sourced fabrics". She has no idea where her fabric supplier sources from. The AI put the phrase in because fashion copy commonly contains it, without anything to show it was true of her shop. If she had published it, the shop would have been making a claim it could not support.
The voice-of-customer file and the benefit and proof sheet you built earlier in the course contain exactly the information the AI is missing. Lesson 8.2 shows you how to give it that information.
A language model can state false things with the same confidence as true ones. It may give a wrong date, invent a statistic, misstate a rule, attribute a quote to someone who never said it, or describe a feature the product does not have. The writing flows just as smoothly whether it is right or wrong, which makes these mistakes easy to miss.
The risk is higher for business copy than for most uses, because a wrong number on your landing page is a claim you are making to your customers. Lesson 3.3 explained that a misleading claim is the business's responsibility, whoever drafted it.
Before anything an AI drafted goes live, list every number, name, date, rule and claim in it. Check each item against a trusted source, such as your own records or an official website, and publish only once every item has been checked.
You can still use AI for writing. Fixing generic output comes down to three things: how you brief the AI, where you use it, and how you edit its output. Lessons 8.2 to 8.4 take these in turn.
For the activity, start by doing what Farah did: ask an AI assistant for a homepage headline and intro with no context at all. Then go through the result line by line, using the five patterns above, and mark each generic phrase. Keep the result, because you will compare it with a properly briefed version in lesson 8.2.
Give an AI assistant no context and ask it for a homepage headline and intro for your business. Then list every generic phrase in what it writes.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).