A burning question for anyone in the content business is this: Will AI replace me?
In the short term, it’s certainly having a sharply negative effect on writers’ careers.
Freelance writing work is disappearing, according to a study done by Upwork that Christopher Mims reported on in June. Since ChatGPT first appeared, the number of freelance jobs posted on sites like Upwork and Fiverr has declined by 21%, and freelance writers’ fees have declined by 19%, worse than any other category of freelancer.
But we've also seen that AI-generated content can hurt companies. Highwire (my former employer) reported earlier this year that one popular platform published 130 AI-generated blogs in two weeks and saw its overall traffic drop by 30%, while another company that published 500 articles in 30 days saw its traffic cut in half.
And Neil Patel found that human-generated content generated 5.44X more traffic than AI-created posts. So maybe human writers are valuable after all!
As time goes on, I believe more and more companies will realize this basic fact: If you want to communicate with humans, you can't outsource that work to robots.
Leaving aside SEO and traffic enhancement, AI-generated copy can hurt you in another way: It makes your copy too generic. If you’re trying to stand out as an original thinker, AI will never work: It will only make you sound like everyone else.
For these reasons, I’m not too concerned about being replaced by an AI.
However, I am eager to learn ways to use AI better, because it’s a powerful tool that can improve my writing.
To understand why — and what areas of writing AI is actually good at supporting — it helps to understand a little bit about how today's AI tools work.
How generative AI works
Generative AI platforms, like ChatGPT or Claude, operate by producing coherent-sounding, readable text based on inferred language patterns.
An AI model is a predictive text machine — like the type-ahead suggestions you might see in Google search or on your iPhone — but for much longer text snippets.
The language patterns that GenAI uses to generate text are not fixed laws derived from years of theoretical analysis into how language works. Rather, they are so-called transformer models (click that link! it goes to a really good explanation by Luis Serrano) built in an automated, iterative way using neural networks, which simulate the behavior of neurons through probabilities.
AI researchers train these models by feeding them billions of source documents, usually scraped from the Internet. By refining their outputs over and over again, the networks are eventually able to produce text that appears coherent and legible to human readers.
These large language models (LLMs) are effective because language is relatively predictable and structured. When you give an LLM a substantial volume of text, such as the entire internet, it gets pretty good at creating text that resembles human writing found on the internet.
The text is often factual, and that’s not surprising because the source texts are usually factual. But LLMs’ output is not in any way constrained by “the truth” or any underlying understanding of facts. They are merely generating plausible-sounding text based on the inherent patterns of language.
Don't ask a dream machine to write a research paper
Understanding how AI works can help you know why tools like ChatGPT will never be very good at creating original copy. Everything they have been trained on already exists, and all they can ever do is generate strings of words based on the patterns that have emerged across that training corpus.
It also explains why AI frequently produces “hallucinations”— it has no model of the truth and no way to connect the words it is generating with any real-world referents. It’s not even fair to call it a hallucination because that would imply that the AI is sometimes generating something “real.” As Andrej Karpathy wrote a couple of months ago, "In some sense, hallucination is all LLMs do. They are dream machines."
In fact, everything an AI generates is pure language pattern. It’s just that sometimes those language patterns correspond to facts, and sometimes they don’t.
Knowing these two things, it’s clear why it’s not a good idea to ask an AI to write a blog post for you. What it produces is going to be generic, and more or less unhinged from any reference to the truth.
21 ways to use AI to improve your writing
But those two attributes make generative AI models powerful writing assistants in other ways.
What writer wouldn't want to have a "dream machine" on their desk, right next to the keyboard, a pile of fresh Moleskines or Leuchturms, and an endlessly self-refilling cup of coffee or tea?
In fact, I’ve found several AI tools — particularly ChatGPT, Claude, and Grammarly — very helpful in a variety of supporting roles:
- Brainstorming topics
- Providing reports to start my research
- Rephrasing sentences or paragraphs
- Providing additional examples to underscore a point
- Coming up with alternative headlines
- Providing feedback on rough drafts
- Expanding outlines with more detail, or filling in gaps
- Generating additional questions to ask in an interview/sourcing call
- Simplifying copy I’ve written
- Getting me started on writing in a particular style
- Checking copy for grammatical errors
- Checking copy for possible issues with voice or tone
- Suggesting more direct and powerful ways to phrase things
- Drafting formulaic copy, such as cover letters
- Helping me prepare for interviews
- Listing competitors to a company
- Summarizing information from a survey
- Analyzing survey results for common themes or phrases
- Providing possible objections to statements
- Summarizing publicly available information (eg about a company or person)
- Prompting me with additional questions to consider
With AI-generated copy, you should always check the facts (even if it provides a citation, it often turns out that the source doesn’t actually exist).
You should always double- and triple-check the writing and revise it substantially before submitting it to a client or for publication.
When I use AI, I never, ever trust it to give me something complete or accurate.
It’s especially important to check everything, because AI-generated copy generally looks so good. It’s easy to think, “Wow, that was way better than I expected!” and then fail to do anything to improve or even verify its facts. But don’t let yourself be hoodwinked. Plausibly good copy is AI’s forte. It’s up to us humans to keep it in check.
And remember, while the AI might have helped you create the rough draft, you’re ultimately responsible for the final draft.
In an upcoming post, I’ll share details on some of the AI prompts that I’ve used successfully to help me with writing and editing.
Do you use AI in your writing? Comment below or write back to let me know. I want to hear what works for you!
Bonus recommendations
Good stories worth reading:
Inside the Secret Negotiations to Free Evan Gershkovich – The Wall Street Journal's reporter was released in a prisoner swap last week after 16 months in Russian prisons. The entire time he was imprisoned, the WSJ worked tirelessly to make sure the world knew about his case – and they were deeply involved in facilitating the work that led to his release. This is a long story, but worth reading because it provides a rare level of detail into what goes on behind the scenes in these high-stakes international diplomatic dances.
Emma Carey: The skydiver who survived a 14,000-foot fall – You may not ever want to go skydiving after reading about this mishap. Miraculously, Emma survived. Even more miraculously, she has clearly retained her optimism and love of life. This story tells a terrifying but ultimately uplifting tale, and paints a portrait of a woman who's making the most of life after her accident – without recriminations.
