Seems Like AI Slop

Earlier this week I made a quote card for LinkedIn. Somebody's real words, a real sentence she'd said about her own organization, set into a simple design. I built it in ChatGPT, which is what I use for most of our graphics.

It came back with a small "CR" in the corner.

That's a Content Credential, the mark that tells anyone looking that AI was involved. And it was. AI laid out that card. What AI had nothing to do with was the sentence in the middle, which a real woman said about her own work, and which is the only part of the card anybody was meant to read.

The mark was accurate and it was useless. It answered a question nobody needed answered, and said nothing at all about the one that mattered.

That's what got me thinking about this.

Everything we've built measures the same wrong thing

While I was doing that, LinkedIn was building the formal version. Tap the three dots on a post now and one of your options is "Seems like AI slop." It went live at the end of July and has been widening since, and the reports feed classifiers that decide what gets recommended past your own network. A post that collects enough of them quietly travels less.

I run a company that builds AI software for nonprofits, so you might expect me to be annoyed about this. I'm not. Most of what that button is pointed at should go. The feed is full of writing nobody thought about before publishing it, and I don't want to read it either.

And to their credit, the people who built it know the difference. Hari Srinivasan, LinkedIn's chief product officer, said plainly that AI and slop are not the same thing, that plenty of people refine their thinking with AI, and that what they want is for real humans to weigh in on what sounds authentic. He also said something I keep turning over: slop is hard to define, and the definition changes.

He's right, and that's the whole problem. Every click on that button is training data. The definition is being assembled in real time out of whatever enough people are willing to flag, which means it will end up describing whatever a crowd can recognize by looking.

So look again at what the badge on my card actually did, because most people have this backwards. LinkedIn isn't inspecting images and judging them. It reads metadata the creating tool signs into the file. An AI image with that metadata stripped, by a screenshot or a re-export, shows nothing at all, and a photograph you took yourself and edited can carry one. The badge records how a file was made. It isn't an accusation, and its absence proves nothing.

The informal version has been running much longer. Somebody posts something they worked on and the top comment is a one-liner about ChatGPT. Punctuation gets treated as a confession, though em dashes stopped being a reliable signal a while ago; what reads as machine-written was never a mark, it's a rhythm, every sentence tidy in the same way and nothing specific to any one person anywhere in it. Meanwhile the accusations land on real people. Allison Rossi, a fractional CMO, said it plainly when the button shipped: "Now LinkedIn has given them another tool to falsely accuse people."

What none of it measures is the thing any of us actually cares about.

Nobody is upset that a machine was involved

They're upset that nobody was.

Read the writing that draws the accusations. It isn't wrong, exactly. It's vacant. Confident sentences about nothing in particular, arranged tidily, with no trace of a person who has ever done the work. You finish it knowing nothing you didn't know before, including anything about whoever posted it.

That's what people are reacting to, and they're right to react. They've just misnamed it. The penalty in that feed isn't for using AI. It's for publishing something with nobody in it. Plenty of that was written by hand, and always was.

The cost of getting this wrong

Here's who I actually worry about, and she isn't on LinkedIn arguing about any of this.

There's an executive director reading all of it right now and concluding that the safe move is to touch none of it. She has a board, a donor list, and a reputation built over fifteen years. She has watched people get called out. She's decided the risk isn't worth it.

That's a rational read. It's also the wrong one, and what it costs her is invisible, which is exactly why it doesn't get counted. She keeps spending three days on the annual report. She keeps spending an evening she doesn't have on the appeal letter. Nobody writes a post about the work she didn't do, and no button flags the four days a year she didn't get back.

A willingness to still try something is not a small thing to hand over to a feed argument.

So let me be plain about what we are

BeMo is AI. We don't hide it, and I'm not going to start now that hiding it would be convenient. What I'd rather do is show what it looks like when a person's idea goes into this kind of software and comes back stronger, tighter, and more meaningful to the people it's for. If we can't demonstrate that, the criticism is right and we should be swept out with everything else.

So here's the demonstration. It belongs to somebody else.

What happened with Jen's appeal letter

Jennifer Allen is the executive director of the Friends of the Saratoga Springs Public Library. Last month she used BeMo to build her annual report, a job that had taken her about three days in past years, in roughly an hour. Three weeks later she came back on her own and wrote the annual appeal letter, the Friends' direct ask to its donors.

She started, in her words, "not entirely sure where to start."

She didn't go looking for a template by name. She described what she needed and got routed to the right one. Then it started asking her questions: about timing, about the goal, about specifically what the money would pay for.

That's the part I'd point at.

"The questions weren't generic. They were insightful, the kind a seasoned development professional would ask, and they pushed me to articulate our case for support more clearly than I might have on my own."

The Friends do not have a development director. What she got wasn't labor. It was a line of questioning she'd otherwise have had to hire, and it made her say out loud what the money was actually for.

Then it did something she hadn't asked for at all. It produced a list of suggested gift amounts, with a description of what each level would cover.

"I hadn't asked for it. It just knew that a strong appeal letter needs that, and the amounts it suggested were spot on for an organization like ours."

That second half is the part that matters. Not that it knew the convention exists. That it knew what an organization the size of the Friends should ask its donors for. That isn't writing. That's knowing whose letter it is.

The letter took about 30 minutes. It used to take a few hours. And when I asked whether it came out better as well as faster, she said yes, and named how: more clear and concise.

Her own description of the result: it "felt personal and purposeful, not like something assembled by a machine."

It was assembled with a machine. She knows that. She wrote that sentence anyway, and both things are true, because the letter has more of her in it than the ones she wrote alone. The questions pulled things out of her that she wouldn't have bothered to articulate to a blank page.

One more thing about the words I've been quoting. Jen wrote that account with BeMo, and told me so in the first line of her email: "I let BeMo write this for me too lol. Finished in minutes!"

Every instinct I have as a marketer said to leave that out.

It doesn't undercut a word of it. Those are her experiences, her judgments, her organization. If any of it had been wrong she'd have told me, because this is a woman who has told me plenty about what our software gets wrong. It helped her say what she thought. It didn't decide what she thought.

Which is true of this essay too, and you should hear it from me rather than guess. I drafted this with AI, the way I draft most things now. I cut a third of it, argued it out of a shape I didn't like, and threw out an opening I was fond of to make room for the one at the top of this piece, which happened to me this week and which no draft was ever going to hand me.

Which is to say I did what Jen did. I treated it as a colleague: something to think with, argue with, and overrule. Nothing here says anything I didn't decide it should say.

Whether I'm in it after all that isn't a question a badge can settle.

So let me give you the test I'd use.

The questions I'd ask instead

"Was AI involved" is unanswerable and getting more so, and even answered, it tells you nothing you can use.

These are the ones I'd ask about any piece of writing.

Did I learn something? Did it make me stop and think twice? Did it connect me to the person who wrote it, or could anyone have posted it? Would I trust the person behind it with something that mattered? And underneath all of them: is there more of that person in this, or less?

Not one of those questions can be answered by a classifier. The answers aren't in the text. They're in the reader. LinkedIn is building machines to catch machines, and the only test that has ever worked is the one already running in your head while you scroll.

It's also a better test than the button, because it catches the vacant AI post and the vacant human one, which is right, since they're the same failure. And it clears work like Jen's letter, which is also right, because the reason that letter sounds like her is that something spent half an hour making her explain herself.

One more thing LinkedIn did

The slop button got the headlines. Something else in the same announcement got almost none.

LinkedIn removed its "enhance your post" tool, the one that would rewrite what you'd written into something smoother. In its place they put a feature that proofreads without changing your voice.

I don't think that's a small detail. The largest professional network in the world looked at what had gone wrong and, as far as I can tell, concluded the problem was never that software touched the writing. It was software replacing the person in it.

That's the whole argument, and they got there from the other direction.

If you're a nonprofit leader who has been avoiding these tools because you're afraid the result will sound like nobody, that instinct is worth keeping. You never needed a badge to tell you what good looks like; you've been able to do that the whole time. Just point the instinct at the right question.

Not whether a machine was involved. Whether you still are.