AI has a measured gain on routine copy, and none yet on the work that wins awards

Six practices the evidence supports

1. Let the model draft routine text, and check it.
2. Feed it your past test results when you have them.
3. Write your own first draft when you are the expert, and ask the model to attack it.
4. Ask for five options, pick one, and add the fact it did not have.
5. Check each fact yourself.
6. Test two versions on readers when the piece matters.

A model now matches staff writers on routine marketing email

A US wine retailer ran three randomised trials on about 27,500 customers. Emails written by a model sold as much as emails written by staff (Dubé and Xu, 2025). In two of the trials the model had been trained on five years of the company's most successful emails. In the third, Claude with a simple prompt did as well or better. A human edit of the AI emails added no measurable sales (Chicago Booth Review, July 2026).

On other routine text the gain is in time. Professionals with ChatGPT finished professional writing tasks 40% faster, and graders rated the output 18% higher. The weaker writers gained most (Noy and Zhang, Science, 2023, 453 people).

Past test results beat a bare prompt on subject lines

Across 36 email campaigns and 283 million impressions, a model tuned on the results of past A/B tests lifted subject-line click-through by 32% over human experts. ChatGPT with a prompt and no such data did no better than the experts (Angelopoulos, Lee and Misra, working paper; one author works for a vendor).

So data about what worked is the input with a proven gain. A voice profile is a different input, and nobody has tested it.

It helps a non-expert as a critic, and it can hurt an expert as a ghostwriter

In an experiment scored on clicks on social ads, non-experts who used the model for feedback on their own draft wrote better ads. Experts who let the model write the first draft did worse. The authors' explanation is that the writer anchors on the model's draft (Chen and Chan, Management Science, 2024).

Practice with a model can improve the writer. One group practised a cover-letter rewrite with an AI tool. Afterwards, with no tool, they wrote better than people who had practised alone or had feedback from human editors (Lira and others, 2025).

Five options give better ideas, and more similar ones

Writers who saw five AI ideas wrote stories that were 8.1% more novel and 9.0% more useful than writers who saw none. The less creative writers gained most, and the stories became more alike (Doshi and Hauser, 2024).

A person must check the facts

• 758 consultants at BCG used GPT-4. On tasks the model could do, they worked faster and better. On one task outside its ability, those with AI were 19% less likely to produce a correct solution (Dell'Acqua and others, 2023).
• AI first drafts of job advertisements made employers 19% more likely to post. The posts were "more generic and less informative" and led to no more hires (Wiles and Horton, 2025).

The error rates for facts and sources are on The issues.

Readers pick the winner

• Models alone could not reliably predict which headline would win. The best model-only method was right about 47% of the time, which the authors call marginally better than random (Ye, Yoganarasimhan and Zheng, 2024).
• Among 1,042 content marketers, those who use AI and those who do not are "equally likely to report 'strong results'". The 8% who do not use AI were much more likely to report disappointing results (Orbit Media, September 2026).
• We found no award-winning effectiveness case that states AI-written copy with results. None of the 46 cases on the 2026 IPA Effectiveness Awards shortlist is described that way in its title (LBB, May 2026).

Disclosure lowers trust, unless you show your sources

Thirteen experiments found that people who disclose AI use are trusted less than people who do not (Schilke and Reimann, 2025). In the cases the university reported, trust fell by 16% to 20% (University of Arizona). Being found out by someone else was worse than disclosing.

One study found no penalty when a news article labelled as AI-made also listed its sources (Nieman Lab on Toff and Simon, 2023).

If you disclose, show the sources and name the person who is responsible.

Practices with no measured result