AI Wrote the First Draft. Do You Have to Say So?

If ChatGPT wrote 70 per cent of a blog post and you edited the rest, do you need to tell anyone? Headlines about AI and trust make it hard to know whether content drafted with AI has become a liability.

It’s a fair question, and most of the advice floating around small business circles doesn’t answer it properly. Half the internet says disclose everything, always, or your customers will feel betrayed. The other half says nobody cares and disclosure labels are just handicapping yourself against competitors who don’t bother. Neither is quite right, and the honest answer depends on what you’re publishing and who’s reading it.

I’ve spent over 20 years in communications. This post is my answer, with the numbers behind it, a workflow you can run in under an hour, and a straight answer on the tools that claim to detect AI writing.

Do customers actually care if AI wrote your content?

Yes, more than they did a year ago, though the clearest data on this is American rather than Australian. Content marketing agency Fractl surveyed 1,008 US consumers in Q2 2026 and found 84 per cent want AI-written text labelled, rising to 90–91 per cent for AI images, video and audio. On the business side of the same survey, only 20 per cent of organisations say they always disclose AI use, and 33 per cent never do.

The trust cost of not disclosing, and getting caught, is also climbing. Fractl’s data shows the share of consumers who’d trust a favourite brand less after heavy AI use roughly doubled between 2025 and 2026, from 20 per cent to 40 per cent. Gen Z respondents were the least forgiving, at 54 per cent, against 33 per cent of Gen X and 32 per cent of Baby Boomers. Women were more likely than men to penalise heavy AI use, 44 per cent versus 34 per cent. I haven’t found an equivalent Australian study, so treat these as directional rather than a precise read on your own customers.

None of that means every Instagram caption needs a disclaimer. It means the businesses getting burned are the ones where AI use is discovered rather than disclosed, usually because the writing reads as generic, the facts are slightly off, or a reader spots the same phrasing across a dozen competitor sites.

Do small businesses in Australia actually have to disclose AI use?

There’s no blanket Australian law requiring you to label AI-written blog posts or social captions. What does apply is the Australian Consumer Law’s existing ban on creating fake or misleading reviews, which the ACCC enforces regardless of whether a human or a tool produced the wording. The ACCC’s own guidance is blunt on this: reviews must reflect a genuine opinion, and a business is responsible for what’s published under its name.

The ACCC has already taken action over exactly this kind of thing, before generative AI made it easier to do at scale. It fined removalist company Citymove for copying and editing testimonials from another site and publishing them as its own, and the Federal Court ordered online platform Service Seeking to pay $600,000 for letting businesses draft their own customer reviews and auto-publishing them if the customer didn’t object within three days. Neither case involved AI, which is the point: the rule against fabricated or edited-to-mislead reviews already exists, and an AI tool that speeds up the fabrication doesn’t create an exemption from it.

Where this bites small businesses isn’t the blog post about EOFY tax tips. It’s the AI-drafted product description that quietly overstates a claim, the AI-summarised testimonial that changes what a customer actually said, or an AI-generated “review” posted to pad out a new listing. Those are already against Australian Consumer Law, disclosed or not. Disclosure won’t save you from a misleading claim; accuracy will.

So here’s where that leaves you. There’s no specific disclosure mandate for ordinary content. There is a real, existing obligation to make sure whatever AI drafts for you is accurate, especially anything that looks like a customer’s own words or a factual claim about your product.

Can AI detection tools actually tell if something was written by AI?

Not reliably enough to stake a client relationship on. Detectors give a probability that text was machine-generated, not proof of who wrote it. Here are three widely used options with their current pricing, and what that means for content drafted with AI and then edited by a person.

Prices below are in USD, as listed on each provider’s own pricing page, so budget a little more once converted and after currency fees.

Originality.ai has a free Basic tier (3 scans a day, 2,000-word cap per scan), then jumps to $14.95 a month for its Pro plan, or $12.95 a month billed annually, covering 2,000 credits a month (1 credit = 100 words, so around 200,000 words of scanning). An Enterprise plan runs $179 a month for 15,000 credits, aimed at agencies rather than solo operators.

GPTZero has a solid free tier reported at around 10,000 words a month, no credit card required. Third-party pricing trackers put its paid tiers starting around $14.99 a month for roughly 150,000 words, up to around $23.99 a month for 300,000 words with batch uploads (GPTZero’s own pricing page uses a seat-based calculator that didn’t return fixed figures at the time of writing, so treat these as approximate and confirm on-site before you buy).

Copyleaks lists a Personal plan at $16.99 a month ($13.99 billed annually) for 100 credits, scanning up to roughly 25,000 words, and a Pro plan at $99.99 a month ($74.99 billed annually) for 1,000 credits, around 250,000 words. It bundles plagiarism checking with AI detection in the one report, which is handy if you’re also worried about content mills or scraped competitor copy.

Keep in mind what a detector is doing: estimating the statistical likelihood that text was machine-generated. Text a person has substantially rewritten, with restructured arguments, specific examples and their own voice, is harder for that kind of estimate to classify. That’s the useful lesson here. The fix for content that “sounds like AI” is genuine editing, and a detector score can’t tell you whether that editing was done well.

A disclosure and QA workflow you can run in under an hour

This workflow suits a business producing one blog post and a handful of social posts a week.

1. Draft with AI, timestamped (5 minutes). Generate the first draft and note which tool and prompt you used. This becomes your paper trail if a customer or the ACCC ever asks how the content was produced. A first-draft prompt you can copy: “Draft a [word count]-word blog post on [topic] for an Australian small business audience. Flag every specific number, date, or claim in [square brackets] so I know what to fact-check before publishing. Don’t invent statistics, case studies, or quotes. Write plainly, no marketing clichés.”

2. Fact-check every specific claim (15–20 minutes). Any number, date, price, statistic, or quote gets checked against a primary source before it survives to the next draft. This is the single most common AI failure mode: confident, specific-sounding claims that are subtly wrong.

3. Rewrite for voice, not just polish (15–20 minutes). Read it aloud. Cut anything that sounds like a LinkedIn infographic. Add one detail only you would know, a specific customer story, a number from your own business, an opinion you’d actually defend in a meeting. This step is also what makes detection tools less reliable, because it’s actually transforming the text rather than disguising it.

4. Decide on disclosure, per piece, not per business (5 minutes). A blog post explaining a process: low stakes, optional disclosure, maybe a line in your author bio saying “drafted with AI, edited by [name].” A testimonial, review response, or “About Us” story: disclose or don’t use AI on it at all, because these are exactly the categories where getting caught costs the most trust per the research above.

5. Run it past a detector only if you’re worried about search penalties, not to prove innocence (5 minutes). Google has said publicly it doesn’t penalise AI content for being AI content, only for being low-quality, unoriginal, or unhelpful regardless of who or what wrote it. A detector score doesn’t change your Search ranking either way, so don’t burn a subscription trying to “pass” one.

Total: 45–60 minutes for a blog post, less for social copy. If you’re paying someone else to do the editing and fact-checking stages, Australian copywriting rate guides put junior freelance rates at roughly $70–100 an hour, so budget $50–90 for the human half of this workflow on a single post. For the wider set of tools around this workflow, see The Ultimate AI Toolkit for Australian Small Businesses.

What not to do

Don’t run AI drafts straight to publish without the fact-check step, even for “safe” topics. An AI draft can easily cite a superannuation threshold or tax rate that’s out of date, and it will go live that way if nobody checks it against the ATO site first.

Don’t use AI to write or summarise customer testimonials and publish them as direct quotes. If AI helped tidy grammar in a testimonial a customer actually gave you, that’s fine. If AI generated the sentiment, that’s the exact conduct the ACCC’s guidance is warning about.

Don’t add a disclosure badge to every piece of content as a blanket policy without thinking about what it signals. Slapping “AI-assisted” on a one-line Instagram caption reads as either try-hard or slightly absurd. Reserve visible disclosure for content where readers are making a judgement based on authenticity: reviews, personal stories, “meet the founder” pages.

Don’t treat a clean detector score as proof of quality. A detector telling you content is “100 per cent human” says nothing about whether it’s accurate, useful, or any good. Human-written copy can get flagged as AI, and lazy AI copy can pass as human. Use your own judgement first.

Where disclosure matters, and where it’s overkill

Split your content into two buckets. Bucket one: informational content where the value is the information itself, a how-to guide, an explainer, a roundup of tools like this one. Readers care that it’s accurate and useful more than they care who typed it. Light-touch disclosure (an author note, not a banner) is enough.

Bucket two: content trading on authenticity, testimonials, founder stories, “why we started this business,” crisis or apology statements, anything where the reader’s trust in your first-person voice is the entire point. AI can help you structure and tighten these, but the sentiment, the specific memory, the actual apology, has to come from you. If it doesn’t, disclose it clearly or don’t publish it as first-person.

Most small business content sits in bucket one. The mistake to avoid is agonising over disclosure for a blog post while quietly letting AI ghostwrite the founder’s “our story” page, which is exactly backwards. If you’re extending this into media outreach, where authenticity counts for even more, How to Pitch the Media Like a Pro (Without Hiring an Agency) is the place to start, and Why Most Small Business PR Fails covers the wider trust argument.

What AI can’t do here

AI can’t tell you which of your claims are still true. It has no way of knowing your current pricing, your latest ABN details, or that the award you mentioned last year has since been discontinued. That’s on you, every time, regardless of how the draft was produced.

AI can’t stand behind a testimonial or an apology. Readers extend trust to a first-person statement because a specific human is putting their name to it. A model has no stake in whether it’s true, so it can’t carry that weight, no matter how well it mimics the tone.

AI can’t tell you when a topic needs a professional. If you’re navigating an actual crisis, a legal claim in your marketing, or a situation where getting the wording wrong has real consequences, that’s the point to bring in a communications professional rather than iterate on a prompt. I say this as someone who sells AI-assisted content services: the tool earns its place for the 80 per cent of content that’s routine, and it’s the wrong tool for the 20 per cent where judgement and accountability matter more than speed.