Write for humans, not algorithms
Why you can’t fight AI writing with AI detection tools
What if you could instantly and flawlessly identify a piece that was written, either fully or partially, by an AI tool at any point? How would you use this detection tool?
For me, and I think for most people, knowing a piece of writing is AI generated is a shorthand for knowing if it’s trustworthy.
What I don’t want is to spend my time reading something that someone else didn’t really write, be it a short story, a work report, a listicle. And I mean “write” as a full-throated process here: living something, thinking about it, talking about it, writing about it, developing the story, polishing the story, shaping the story for where it’s going to land.
Even if two people by complete random chance posted the exact same story, only one was from lived experience, and the other was the result of a perfectly crafted prompt into ChatGPT, I would want to know because I trust the former far, far more than I trust the latter.
But if I think about what passes my bar for trustworthiness, I realize it’s not entirely about using AI.
- There was untrustworthy writing long before AI, and there will be after, too. Some non-AI writing will never be worth a read because it’s written in a deceptive, untrustworthy way.
- Some AI writing is trustworthy for me, because the writer still cares. It’s still useful writing for me. The care comes through in the writing.
I find when readers want to know if something is AI generated, they are not really asking is this AI or not AI. They’re asking: Can I trust this?
That’s the amazing thing about reading great stories. It can change your life. But you need to trust the writing, the writer, the process they took to create this piece and set it before you, to allow it to do that most magical of things that reading a good piece of writing can do: change your life, in some big or small way, for the better.
Why doesn’t Medium offer AI detection
Given some flavor of “please show me less slop” is a common inbound support ticket, you might wonder why we haven’t invested in AI detection.
The truth is that we have. Not through AI detection tools, which are iffy at best. (I have experienced this myself — I asked Claude to generate a few grafs about my cats, and copy-pasted them into Pangram, which classed them as “human written.”)
We don’t want a baseless hunt of people accusing writers of not being genuine, while the savvier AI users get away with using AI tools to rewrite AI writing until it passed detection. (This already happens enough even without a percentage AI attached!)
Even if AI tools were fully reliable, they wouldn’t really answer that underlying question: can I trust this? Is this writing worth reading? Can I let it change me?
Trust matters to Medium. We believe in the power of reading and writing to change lives. Why would anyone pay $5 or $15 a month if they can’t rely on the stories they read?
Not only that, we also didn’t want a future of incentivizing writers to worry that a machine would scan their writing to determine if it should be trusted or not.
Get Zulie @ Medium’s stories in your inbox
Join Medium for free to get updates from this writer.
We’ve been busy investing in the curation era instead. Machines can’t tell if writing is good or not. People, broadly speaking, may not be able to agree on what’s great writing.
But our hypothesis is that if we give readers the tools to find trustworthy voices, curate their feeds, discover new writers, and join communities in the topics they care about, they’ll be more able to find the writing they can trust to change their lives.
The curation era
Here are two problems:
- AI-powered AI detection is not reliable.
- There’s good writing that used AI, and bad writing that didn’t touch AI, and you still want to get the good while avoiding the bad.
How do you solve for that?
There’s a huge demand to find genuine writing nowadays, and so different platforms have gone about it different ways. Pinterest (gen AI labels), Instagram (“AI info” labels), and most recently Substack (Pangram integration), are all offering tools for creators to label their content as AI-generated, or for readers to check if something was generated by AI.
But I find most of us, when we see something in our realm of experience, can detect when something is not quite right; it falls into that uncanny valley. I’m not talking about m-dashes or the word “quietly” or “it’s not X, it’s Y” — every tell has had its moment in the sun and been replaced by another. I mean the writing is boring, repetitive. It doesn’t make sense. Bits don’t ring true to you. It’s too perfect.
You can just tell.
I can’t tell with all writing. With short fiction, for example, I have a hard time identifying AI-generated writing, just because I personally don’t read much of it. (AI novels are another story.) However, regular short story readers and writers tell me they have no trouble picking out genuine versus non-genuine writing.
This is where the demand for some kind of detection comes into play— not AI writing, necessarily, but trustworthy, good, writing. If you can’t do the detection yourself, you want to outsource it. We all want to be able to trust what we read.
Our method was to invest in the curation era. We built (and are building) several different methods to surface the good from the slop:
- There’s Medium’s opinion on what’s good. Our curation team review thousands upon thousands of stories and give them distribution labels to elevate the best while curtailing the things readers have told us they don’t want to see. (Usually, but not always, this is AI slop.) Read more about our distribution guidelines if you’re interested.
- “The best” writing is not universal; what’s great for me is different to what’s great for you. Publications curate and elevate the best for their communities (and reject what’s not a great fit). We’re now paying editors through the Editor Partner Program, and publications can feature stories to their readers. Readers can repost stories (now with a note for extra context!) when they want to really recommend a story.
- Finally, there’s also your individual opinion on what’s good and bad. Follow the writers, publications, and topics you want to see more of. See something you hate? Mute it, be it a writer, publication, or most recently, a whole topic. Love someone’s taste and want to see more of what they’re reading? Check out their activity tab. Want to get an email when a writer you follow publishes something new, so there’s no chance you’ll miss it? Subscribe to their email notifications.
The upshot of this is that for the average reader, if you see a story on your feed, it’s most probably there because a person recommended it. (Over 75% of stories on the average reader’s Medium feed are there because a human recommended them.)
Is this a perfect system? No. Bad writing slips through the cracks. But we’re betting that investing in curation and customization is more sustainable, more effective, and actually delivers on what readers want: Not a way to tell if something was AI generated, but a way to tell if the writing is worth trusting, if the writer is worth following.
Write for the readers, not the machines
Writing started suffering long before AI. When you try to write for machines, not people, you write bad writing.
Writing to rank for SEO is different than writing to speak to a person. But to get that writing seen by a person, writers had to figure out what the right balance of keyword stuffing was so indexers like Google would rank your stories for readers to find.
Writing to go viral is different than writing to speak to a person. But to get your writing seen by as many people as possible, writers were incentivized to post clickbait.
And writing to pass an AI detector is different than writing to speak to a person. But if some machine is sitting on top of all writing and deciding what is and isn’t good enough to pass muster, of course you’re going to want to make sure your writing reads as “100% human” for it, not for your readers.
We want to show readers writing they can trust. We want writers to write for their readers, not Medium, not a detector, and not be afraid to use m-dashes in peace. And we think the best way to get there is to trust people, not machines, to lead the way.
