Made With AI, Not By AI

My wife and daughter are both social media managers for various organizations (as well as for themselves, of course). They both have good marketing instincts, and they know something about how the algorithms work and how to maximize engagement. One thing I've heard them complain about more and more is the prevalence of AI-generated flyers.

One reason my daughter runs about six different school clubs’ social media accounts is to keep AI flyers out of those feeds. Not just because she hates them, but because people scroll right by them. So she wants to make sure that every story or reel about an upcoming school fundraiser or performance matches the right branding and doesn't look like it was made by robots.

And it's pretty obvious when the images were made with AI. There are usually stars or hearts, the font is predictable, and there are brush-stroke underlines and busy backgrounds.

It's a strange moment we're in. Using AI really isn't optional anymore, especially in the workplace, where agents are now just part of the workflow. We're supposed to use AI. Except a lot of effort also goes into making sure that the result doesn't look like we did.

So everybody's using AI, but nobody wants to look like they're using AI. Which means we have to figure out what the identifying characteristics are of "AI slop" so we don't get "caught" using AI. But the tells are different for images than they are for writing. And then there's all the code that agents are producing. Do we need to de-AI-ify that too?

# Find the Flyer

To give my girls some credit, it does feel like the AI flyers are rampant on social media feeds. Though I said that people don't want to look like they're using AI, the people who post these flyers can't possibly think they're hiding it. They are so easy to spot!

As with any new tool, it takes some practice before you can use it well, and until then, it shows. With AI, it's garbage in, garbage out. Or more accurately, nothing in, garbage out.

If you just ask it to “make a flyer for the spring concert,” you’re going to get the most generic flyer possible, because that’s all it has to go on. It’s basically giving you the average of every flyer it’s ever seen. But if you give it something to work with, like the club’s colors, the logo, an actual photo from last year’s show, and some real direction about what you want, the results greatly improve. The more descriptive your prompt is, the better the result will be.

You also don't need to take the first thing it gives you as the final draft. In fact, you shouldn't just use whatever it spits out based on your first prompt.

Instead, use it like you would Canva or Photoshop, and build things out iteratively, trying some things, discarding them, and trying something else. In other words, don't let it build the whole thing and walk away.

You could have it just get you started and spit out a PSD so you can finish it up in Photoshop. Or just keep telling the agent how to fix what it gives you until you get it how you want it. (Though all that iteration isn't free, which is one more reason to give it more to work with up front.)

By the way: none of this is about art, and that's a bigger conversation than I'm having here. I'm talking about everyday personal and business use, like a club flyer, website copy, or a blog post. I have a lot of respect for real artists, and nothing here replaces what they do.

# Does Anyone Care Who Wrote the Code?

In tech and enterprise software, agents are writing most of the code now, if not all of it. But again, it comes down to how you use the tool. When I'm hacking together a prototype or a demo with Claude Code or Copilot, I'm usually fine with whatever it writes, as long as it works.

But code review exists for a reason. Imagine if "it works" were the bar for human-written code! And as I've admitted before, I can't really call myself a software engineer anymore, so I won't be catching the subtle bugs lurking under the surface.

For code that ships to production, there are still (or at least there should be) capable engineers behind the keyboard, designing the architecture and steering the agent where it needs to go. Without that, it's pretty easy to tell when an agent wrote the code. Sometimes it takes shortcuts, like changing a test to match the code instead of fixing the code to pass the test (or just quietly skipping the test altogether). Other times it overcomplicates things, with more defensive checks than you'd ever need and an architectural pattern way fancier than the problem called for.

And agents love comments.

They will write whole paragraphs of them, sometimes explaining something useful in too many words, and sometimes just filling space. Even with an engineer driving, some of these tropes will stick around.

def is_even(number):
    """
    Determines whether a given number is even.

    Args:
        number (int): The number to check.

    Returns:
        bool: True if the number is even, False otherwise.
    """
    # Check if the remainder is zero
    is_number_even = number % 2 == 0
    return is_number_even

But does it matter if code looks AI-written? If agents are the ones maintaining it, and fewer and fewer people ever actually read it, does code quality even matter? For now, I think it still does. Agents have to work with whatever's already there, and messy code can trip them up just like it trips up people. All the usual code smells make the next change harder, no matter who (or what) is making it.

Maybe that matters less over time, as agents get better and humans do less and less of the maintenance. But I don't see architecture and design going the same way anytime soon. Figuring out how a system should fit together, and why, is still likely to need a lot of human input. Then again, maybe the AI overlords will take over completely and prove me wrong at some point.

# Does This Sound Like AI to You?

For me, writing is the trickiest because there are so many different styles, personal preferences, and unique voices. You can have AI write a first draft and then edit it to sound more like a person wrote it, or you can write a stream-of-consciousness draft yourself and ask AI to clean it up. Either way, you don't want it to sound like AI, so you feed it back to the agent and ask whether it sounds too much like AI.

It's the same move as having one agent review the code another agent wrote. That's a real practice, and it's proven useful to a degree. But where does it get us if we're just using AI to solve the problems AI created?

I made this point on LinkedIn with a Project Hail Mary reference. Astrophage is the thing eating the sun, but it's also the fuel that gets Ryland Grace close enough to stop it. The problem and the solution are the same stuff. That tracks.

# Bryan Slop

Here's the part that bugs me most, though. Even when I write something completely from scratch, it's starting to sound like AI.

I'm sure some of that is overexposure. I read so much AI-generated text in a day that its rhythms are leaking into mine. But it cuts both ways, since the things AI does were human habits first. It learned all its favorite writing tricks from us: the list of three, the this-not-that pattern, the em-dash.

Oh, the em-dash! How I loved using the em-dash. Now nobody can use one without someone assuming a robot wrote it. So we're editing our own habits out of our writing to avoid sounding like the thing that copied them from us. And what do you do when an em-dash really is the right punctuation? Use a colon? A comma? We're bending grammar rules in the name of AI avoidance. It's great news for the semicolon though.

"Human Made" by Tom Fishburne, marketoonist.com

Of course, sometimes my from-scratch writing is just bad (which was true before AI too). Instead of AI slop, it's Bryan slop. Maybe it's too wordy (me? never!) or repetitive. Or I've buried the lede somewhere deep in the piece in favor of a clever opening hook that barely works. (Is this getting too meta?) My own writing needs edits and several revisions before it's any good, same as anything an agent hands me. Writing it myself isn't a guarantee of quality. It just means the mistakes are mine.

Which has made me realize we have to be careful about confusing "AI slop" with "AI-written." People use them like they mean the same thing, but I'm not sure that's true. AI can write well, sometimes even better than I do on a first pass. Those brush-stroke underlines on the Instagram flyers aren't slop because AI made them. They're slop because nobody bothered to fix them.

Let's face it, people were making slop long before AI came along. AI just lets us make more of it, faster.

# Give It Something to Work With

In my last post, I wrote about a teacher at my daughter's back-to-school night who has his students do all their writing on paper. His reasoning is that you can't judge what AI produces if you never learn what good looks like. I still believe that. But even when you do know what good looks like, it's still hard.

So which way of writing with AI is better? Do you edit an AI draft, or does AI edit yours? I find that when I'm stuck, an AI draft gets me moving faster than a blank page, but when I already know what I want to say, I start myself and let AI edit. Either way, we're all still figuring out our own AI workflows, and they'll probably keep changing just like the tools do.

Regardless, it's the same lesson as the flyers: nothing in, garbage out. In the case of writing, though, the voice is the branding. When I write with AI, I'll often point it at this blog as a reference so it can try to pick up my tone. That only works if this is the voice I'm going for, though. What about at work, where the writing isn't supposed to sound like us? Website copy or marketing material is supposed to sound like the company. So the agent needs the writing version of the company's "colors and logo," so to speak. Maybe that's a style guide, or past copy to give examples of what you're looking for.

Whatever the workflow, whoever's voice it is, the most important part is that last pass. No matter how you start, AI can get you most of the way there. But at the end of the day, somebody who knows what good looks like still has to review it before it goes out. And even then, it's getting harder and harder to tell who wrote what.

So, was this post written with AI? Some of it, sure. Did I write a rough draft and have AI clean it up, or did AI hand me a draft that I tore apart? My wife and daughter will probably let me know.