The Fluent Draft With a Fake Award
A four-person marketing team at a solar-panel installer must write a 700-word product page for a new home battery. They have a tight deadline: copy due in two days, legal review in one. A colleague pastes three real customer reviews, including names and addresses, into a free public chatbot and asks it to 'write persuasive copy in our brand voice and include the 10-year warranty and the MCS certification'. The draft it returns is fluent and confident, cites 'the 2023 Which? award' and a '97% efficiency' figure, and uses a warmer tone than the company's usual plain style. No one on the team has used the tool before. The marketing lead asks you to set a safe, repeatable process for the remaining work.
1. Identify three problems with how the colleague used the chatbot, and rank them by the damage they could cause.
Read the worked answer
Three problems, ranked by worst realistic outcome:
- Personal data in a free public tool (highest). Names and addresses of three customers were pasted into a service whose terms may allow training on inputs. This is a likely data-protection breach, cannot be undone, and may require notification. The customers are identifiable and did not consent to this use.
- Unverified, fabricated claims (high). '2023 Which? award' and '97% efficiency' may be false. Publishing them could breach consumer-protection law and damage trust; the fluency of the text makes them more dangerous, not less.
- Brand-voice drift (lower but real). A warmer tone conflicts with the plain style; this is fixable by editing, but it signals the team did not brief the tool.
The common error: treating a fluent draft as finished work.
2. Write a prompt strategy that would improve the output without repeating the mistakes, and explain what still requires a human.
Read the worked answer
Use a constrained workflow:
- Inputs: de-identify or summarise reviews; no names, addresses, order numbers. Prefer an approved enterprise tool with a data-processing agreement and training opt-out.
- Brief: supply the brand guide, audience, word count, and a short list of approved facts. Instruct: 'Use only the facts listed below. Mark any claim without a source as [VERIFY]. Do not invent awards, statistics or certifications.'
- Structure: ask for an outline first, then a draft, then a plain-English edit.
- Verification: every number, award and certification checked against a primary source; legal review before publication.
Humans still own the factual claims, the legal sign-off and the final tone. The tool can draft; it cannot certify.
3. The tool's output says '97% efficiency'. What steps would you take before that number appears on the page?
Read the worked answer
Treat '97% efficiency' as unverified until traced to a primary source.
- Ask engineering or the manufacturer for the test report: what efficiency (round-trip? cell?), under what conditions, at what temperature and depth of discharge?
- Check whether the figure is from a competitor or a different product; models often blend plausible numbers.
- If no source exists, remove the number. Replace it with a verified figure and its conditions, or with a qualitative statement.
- Ask legal whether the claim needs qualification; consumer law may treat an unsupported efficiency claim as misleading.
- Record the source in the review file so the claim can be defended later.
The model's confidence is not evidence. A single fabricated statistic can undo the page's credibility.