AI copywriting versus the human: does it work and can you spot the AI text?

Can a language model write your ads, product copy and newsletters? Yes, and faster than you can. The question that matters is what happens between the first draft and the moment the text goes online. We write with AI ourselves and see that step in between as the real craft. Below you will read what AI copywriting is, what research says about its quality, when a human copywriter wins, what a workflow with fact-checking and final editing looks like, and what Google and the EU AI Act make of it. Along the way, you can take a test with two opening lines.

What is AI copywriting?

AI copywriting is writing commercial copy, such as ads, product descriptions, emails and blogs, with a language model such as ChatGPT, Claude or Gemini. A person provides the briefing and the context, the AI writes a first draft and an editor checks facts, tone and brand voice. By now it is commonplace: 60% of Flemings who use generative AI at least monthly write texts with it.

That figure comes from the imec.digimeter 2025, the annual survey of digital life in Flanders. 43% of Flemings use generative AI at least monthly, compared with 28% in 2024 and 18% in 2023. Of those active users, 60% use it to write emails, articles or social media posts, and 79% say it saves them time. It has taken hold at work too: 51% of working Flemings use generative AI at least monthly for their job. Agreements follow more slowly, because only 27% say their employer has a clear policy and guidelines on generative AI.

It is measurable in business writing as well. Researchers from Stanford, among others, analysed corporate press releases from January 2022 to September 2024. Their estimate: by the end of 2024, roughly 24% of corporate press releases had been written with LLM assistance (Liang et al., Patterns, 2025).

The tools range from general assistants such as ChatGPT, Claude and Gemini to specialised writing tools such as Jasper. The superpowers of ChatGPT, Claude, Gemini and Perplexity shows which assistant is strongest for which task. ClickForest writes blogs, social posts and newsletters for SMEs in Flanders this way: with AI as the production engine and human editing on every text.

Can you tell which opening line is written in AI style?

Probably less well than you think. Below are two opening lines for a B2B landing page. We deliberately wrote one in the style that rolls straight out of a language model, the other as an experienced copywriter would approach it. Both lines were made up for this test. Pick the one you think is the AI style and read on: the answer is at the bottom of this article.

Variant A: “Transform your business processes today with our innovative cloud solutions and experience the difference immediately.”

Variant B: “Your competitor banked their first million yesterday. You’re still wrestling with Excel sheets from 2019.”

Does AI copywriting work?

For speed and a good first draft, yes. In an experiment in Science, professionals using ChatGPT needed 40% less time on average, and their text scored 18% higher on quality. We found no independent study that compares AI copy and human copy on real sales. And trust needs attention: half of American consumers say they would rather buy from brands without generative AI in their content.

The experiment by Noy and Zhang (Science, 2023) gave 453 college-educated professionals a writing task that suited their occupation, and randomly let half of them use ChatGPT. That group needed 40% less time on average, and their text scored 18% higher on quality. Note what was compared: a person with AI against a person without AI. An AI writing entirely on its own was not part of the test.

A study in Science Advances saw the same pattern in short stories, with a downside. Writers who received ideas from an AI wrote stories that were rated as more creative, better written and more enjoyable, especially when they were less creative themselves. The AI-assisted stories were, however, more similar to each other than stories by people alone (Doshi and Hauser, 2024). For copy, that is the heart of the matter. A brand that sounds like everyone else does not stand out. The distinctive example, the angle and the line that sticks have to come from a person.

Then there is trust. In October 2025, Gartner asked 1,539 American consumers how they view brands that use generative AI in their messaging, advertising and content. Half say they would rather buy from brands that do not, and 61% often doubt whether the information they base everyday decisions on is reliable. This is a survey of what people say; it does not measure how they actually buy afterwards.

“Marketers should treat GenAI as a trust decision as much as a technology decision.”

— Emily Weiss, Senior Principal Analyst at Gartner (March 2026)

When do you choose a human copywriter?

Choose a human copywriter when the text positions your brand, when timing and cultural sensitivity count, and in a crisis. There every nuance matters and a mistake costs more than the time you save with AI. For product descriptions at scale, variants for A/B tests and first drafts of blogs or posts, AI is strong, provided a person checks the text.

Strategic brand positioning

A language model writes fluently about your product, as long as you provide the context. Which place your brand takes in the market and which promise you can keep requires knowledge of your customers, your competitors and your margins. Think of the customer group you deliberately leave aside, or the price you hold while a competitor goes on sale.

Cultural nuances and timing

Timing and a feel for what moves an audience remain human strengths. Think of Spotify’s annual “Wrapped” campaign, which presents listening data while tapping into collective nostalgia, into what is on people’s minds at that moment and into the urge to share something.

Crisis and reputation communication

During a PR crisis, every word counts. The difference between “we take responsibility” and “we acknowledge our role” can decide whether a brand recovers or suffers further damage. For standard commercial copy AI does fine; in crisis communication the risk of the wrong tone is too great to use a first draft unchecked.

AI or human? The split per task type

Task type	Strongest option	Why
Product descriptions at scale	AI + human	AI delivers volume, a person checks the product facts
Email variants for A/B tests	AI, a person checks	AI quickly produces many variants, the test decides which one wins
Blog drafts for SEO	AI + human	AI structures, the human adds expertise, examples and sources
Social media posts	AI + human	AI delivers volume, the human guards the tone
Strategic brand positioning	Human	Requires market knowledge and choices a model does not make for you
Crisis and reputation communication	Human	Every nuance matters, a mistake is expensive
Emotional storytelling	Human	Evoking an emotion requires your own observation and experience

For product copy that also needs to rank, read more in SEO for e-commerce. An A/B test decides which variant of an email or ad wins, as described in conversion optimisation for your webshop.

How do you write copy with AI, from briefing to final edit?

Work in five steps: a briefing with context, a prompt that passes on that context, a first draft from the AI, a fact-check and a final edit for tone and brand voice. A person does the last two. A language model can fill in missing information with details that sound convincing and are still wrong, and you only catch those with a deliberate check.

1. Briefing. Record who the text is for, what it needs to achieve, which channel it appears on and which facts must be in it: prices, specifications, dates, names. Add examples of your own tone. At ClickForest, every project starts with an analysis of the client’s existing content, tone of voice and brand values, and that input goes into the prompts.
2. Prompt. Give the model a role, a goal, the context from your briefing and the desired form of the output. Writing better prompts explains how to build such a prompt. If you want to go further, you will find techniques such as few-shot in prompt engineering.
3. First draft. Have several variants made and pick the strongest as a starting point. This is where the time saving is.
4. Fact-check. Hold every figure, every name, every product claim and every source against your briefing or the original source. Language models can hallucinate: fluent sentences with made-up details. According to the imec.digimeter, only 54% of active AI users in Flanders always check the result against other sources. Google asks for that same care for accuracy with automatically generated metadata, such as titles.
5. Final edit. Read for tone, brand voice and rhythm. Watch the word choice for your market: in Dutch, for example, terms from the Netherlands stand out to a Flemish reader, and the formal or informal “you” is a brand choice. Cut the sentences any competitor could also have published.

An example: what context changes in a prompt

This illustrative example shows how the prompt steers the result, and why step 4 remains necessary.

General prompt: “Write an email for customer retention”

AI output: “Thank you for your trust in our company. We appreciate your loyalty and would like to keep you informed of our latest developments.”

Prompt with context: “Write a personal email to Marie, an account manager who purchased our software three months ago. She recently expanded her team from 5 to 12 people. Use a warm but professional tone, subtly reference her growth and proactively offer support. Write as if you have known her for years.”

AI output: “Hi Marie, I saw on LinkedIn that your team has grown significantly, from 5 to 12 people in three months, impressive! That growth naturally brings new challenges. How is the onboarding of your new colleagues going in our system? If you run into anything, let me know. We recently launched a team dashboard that is a perfect fit for teams of your size.”

The second prompt produces a much more personal email. Do look at what was made up along the way: the mention of LinkedIn and the new team dashboard were nowhere in the prompt, and neither was the growth “in three months”. Those are exactly the details you remove during the fact-check.

ClickForest works this way itself. The articles on this blog, in Dutch, English and French, are written with AI and edited by a human. For Flemish SMEs in the Mechelen and Antwerp region, ClickForest writes blogs, social posts and newsletters with AI, and Frederiek Pascal checks every text for facts, flow and brand voice before it is published. If you would rather hand that work over, you will find the approach on the page about outsourcing AI copywriting with human editing.

Can you recognise AI copy?

Barely, and certainly not reliably. In six experiments with 4,600 participants, people could not tell which self-presentations an AI had written, and with poems the score of 46.6% correct answers was even below chance. In 2023, OpenAI withdrew its own detector for AI text because it was not accurate enough. The signals below are therefore mainly useful for improving your own texts.

The research in PNAS explains why. People rely on rules of thumb that seem intuitive and are still wrong: they see the word “I”, contractions or a story about family as typically human. A language model can play on that and write text that the researchers call “more human than human” (Jakesch, Hancock and Naaman, 2023). Something similar happened with poems: readers without poetry expertise gave AI poems higher scores for rhythm and beauty, and therefore more often took them for human work (Porter and Machery, Scientific Reports, 2024).

Few people think they can do it themselves. In a representative survey by the Nuremberg Institute for Market Decisions (NIM) in the US, the UK and Germany, only 25% think they can recognise AI content. At the same time, concern is growing: 80% of Flemings are worried they can no longer tell the difference between human and machine, compared with 64% in 2023 (imec.digimeter 2025).

“The strongest increases in concerns relate to authenticity.”

— Lieven De Marez, research director for society and innovation at imec and UGent (March 2026, translated from Dutch)

Signs of unedited AI text

People sometimes write like this too. For an editor, these patterns are the places to step in:

• Every paragraph follows the same structure: heading, explanation, benefit, call to action
• Colons and bullet lists, even where an ordinary sentence would do
• The same thought twice, in different words
• General statements without an example, figure or name
• No point of view, even where your reader expects one
• Emotions that are named (“impressive!”) without the text evoking them
• A flat, neutral tone on sensitive topics
• Words that fit on any website, such as “transform” and “innovative solutions”

Does AI copy rank in Google?

That depends on the text. Google judges the quality of content, regardless of whether a person or an AI wrote it. What does break the rules is using AI to mass-produce pages that add nothing for the reader. Google calls that “scaled content abuse” and treats it as spam. For a webshop with hundreds of product descriptions, that line matters.

“Appropriate use of AI or automation is not against our guidelines.”

— Danny Sullivan and Chris Nelson, on behalf of the Google Search Quality team (February 2023)

Google confirms this in its guidance on generative AI, updated in December 2025: according to Google, AI is especially useful “when researching a topic, and to add structure to original content”. In the spam policies, updated in August 2026, the first example of scaled content abuse is using generative AI “to generate many pages without adding value for users”. So the facts, your own examples and insights come from you, and every product page you write with AI deserves its own check. For how to use AI in your SEO work, see combining SEO and AI. If you also want to show up in the answers of ChatGPT or Perplexity, read more in GEO strategy.

Do you have to label AI texts?

Not always. Since 2 August 2026, AI text that informs the public on matters of public interest must carry a disclosure, unless a person reviewed the content and someone holds editorial responsibility for it. According to the European Commission, advertising copy and product descriptions are in principle outside this, except for claims about, for example, health, consumer safety or sustainability.

That rule is in Article 50 of the EU AI Act, and Article 50 has not been postponed. The Digital Omnibus only gave providers of AI systems that were already on the market before 2 August 2026 until 2 December 2026 for the machine-readable marking. For anyone who writes AI copy, these details matter most:

• Reviewing means checking the content. According to the Commission, human review requires someone with knowledge of the subject to look at the content. A spelling or grammar check does not count for this.
• No retroactive labelling. Content that was created and published before 2 August 2026 does not need a label. An older AI text that you only publish after that date does fall under it.
• Hiring an agency does not automatically make you a deployer. If you have an agency make an ad without deciding yourself whether and how it uses AI, you are not a deployer according to the Commission’s guidelines.

For what Article 50 requires for chatbots, AI images and deepfakes, see the EU AI Act for SMEs.

Then comes the trade-off the law does not make for you. In thirteen experiments, people trusted someone less as soon as that person disclosed using AI, whether the disclosure was voluntary or mandatory. Those who were caught using AI by someone else lost even more trust than those who disclosed it themselves (Schilke and Reimann, 2025).

“We find that, despite being touted as a practice of ethical transparency, AI disclosure paradoxically erodes trust.”

— Oliver Schilke (University of Arizona) and Martin Reimann, in Organizational Behavior and Human Decision Processes (April 2025)

The NIM saw something similar in advertising. Participants were shown the same product ad, and those who were told the image had been made with AI rated the ad more critically. So disclose AI use where the law requires it or where your audience expects it, and above all make sure the text itself is right.

How do you start copywriting with AI?

Start with one type of text you write often, such as product descriptions or social media posts. Write a fixed briefing, save the prompts that work and have a person check every text before it goes online. If you work with a team, agree who does the final edit and which texts never go out without human review.

As a beginner

• Choose one assistant and get to know it well. An overview per task is in the best AI marketing tools for marketers.
• Build a prompt library. For each type of text, save the prompts that produce good first drafts. A fixed template such as the RACE model for ChatGPT prompts helps with this.
• Compare with your own version. Now and then, write the same text yourself and see what the AI does better and what it misses.

As an advanced user

• Capture recurring tasks. If you often make the same kind of text, you can capture the workflow, for example in Claude Skills.
• Test variants. Have the AI create variants of subject lines or ad copy, check them and let an A/B test decide which one wins.
• Make agreements within your team. Only 27% of working Flemings have an employer with a clear policy on generative AI. Set out which tools are allowed, who reviews and what never goes out without a check.

Learning goes faster with guidance. ClickForest provides AI training for teams at Flemish SMEs in Mechelen and Antwerp: practical work with ChatGPT, Claude and Gemini, effective prompting and using AI in marketing and content, applied to your team’s own work.

Which opening line is the AI style?

Variant A is the AI style. It sounds professional, but “transform”, “innovative cloud solutions” and “experience the difference immediately” would fit on any company’s landing page. Variant B is the human style: a concrete image (“Excel sheets from 2019”) and a confrontational tone. Were you unsure? Then you are in good company, because in research, readers barely recognise AI text.

With a good briefing, a language model also writes lines like variant B, and a copywriter on a bad day writes variant A. The difference lies in the choices behind the text: which detail you pick, which tone you strike and which reader you write for.

Who wins: AI copywriting or the human?

A person with AI. Research shows that professionals with AI wrote better texts in less time, but also that in an experiment with short stories the AI-assisted texts grew more alike, and that people trust someone less once they disclose AI use. The gain lies in the combination: AI for speed and variants, a person for the facts, the brand voice and the choices that make your texts your own.

This is how we work ourselves, and how we work for clients. ClickForest writes blogs, social posts and newsletters with AI and human editing for Flemish SMEs in the Mechelen and Antwerp region. Want AI copy that sounds like your own brand? Take a look at content production with AI and final editing, teach your team to write with AI through AI training or get in touch.