On 30 July, LinkedIn gave its members a button that says “Seems like AI slop”. Reports feed its classifiers. Flagged posts get less reach outside your network. And in the move I find far more interesting than the button, LinkedIn switched off its own AI post writer and replaced it with a tool that proofreads your words instead of changing them. Hari Srinivasan, LinkedIn’s chief product officer, announcing it: AI slop is a top priority for all of us.
Nobody who has scrolled a feed this year needed telling. We have all been inundated. LinkedIn is the worst of them, and not because its members are worse people. It is the worst of them because it runs on text, and text is the cheapest thing these models make.
Here is where I part company with most of the commentary about this. It is not an argument against using AI, and the people reaching for it are not the problem. Pointing a model at a mass of material, pulling the signals out of it and putting a shape on the patterns is real work, done in minutes, that was too slow to bother with a year ago. I use it every day and I would not run a business without it.
The machine is very good at surfacing what might matter. It has no idea what is true for your business, and no stake in whether it is right.
Which is why the next step is not optional, and it is the one everybody skips. Somebody with judgement and something to lose has to look at what surfaced, decide what is actually true here, and put their name on it. The machine does the surfacing and a person does the deciding. Skip the second half and you have not saved an hour. You have moved the error downstream, at speed, at volume, with a confident tone on it.
The trade nobody priced
The slop everyone complains about is the obvious failure, and the boring one. The failure worth your attention is what happens when the AI writing is good.
Researchers at UCL and Exeter ran this properly. Around 300 writers produced short stories, some with access to AI-generated ideas, and 600 people rated the results. The AI-assisted stories came back rated more novel, more useful and better written. Individually, it worked exactly as advertised.
Put those two findings side by side and you have the trade. Each piece of writing gets better. All of the writing gets more alike. And the people who gain the most are the ones who had the least style of their own to begin with, which is exactly how everybody ends up in the same middle.
We have collectively agreed to sound slightly better in exchange for sounding like each other. It is a terrible trade, and almost nobody has priced it, because the cost never lands on the post you are writing today. It lands across the year, on all of them at once.
That middle is competent. Well structured, correctly punctuated, briskly paced, utterly interchangeable. You have read four hundred posts out of it this year and cannot name one of them now.
Your voice is the asset being averaged away
Tone of voice is not decoration laid over the substance. When anyone can produce competent text on demand, it is close to the whole of your differentiation as a person. It is how a reader decides in four seconds whether there is somebody here worth listening to. And it carries the things a model cannot supply on your behalf: what you have actually seen, what you got wrong, what you believe that your competitor does not.
Hand that over to be smoothed out and you are trading the one thing nobody else has for polish that everybody else is getting too. Once, it is a good edit. Every week for a year and your writing is indistinguishable from the writing of people who know a fraction of what you know. You have spent twenty years earning a point of view and then published it in a voice that belongs to nobody.
Most of what gets filed under thought leadership now is neither. It is a summary with a confident tone bolted on. The tell is simple: if the piece could have been written about any company by anyone, then that is exactly what happened.
Which is a strange thing to do voluntarily, because the platform has started paying for the opposite.
The feed has already started pricing this
Three signals point the same way. Reach is being cut on content the classifiers judge to be slop, so generic writing now carries a distribution cost and not merely a reputational one. LinkedIn switched off its own AI writer, which is the organisation holding the best data in the world on AI-written text deciding that protecting the author’s voice beats automating it. And the formats that are hardest to fake keep climbing: chief executive Ryan Roslansky has reported three straight quarters of double-digit growth in video uploads, with comments up 24% across the year.
My own feed says the same thing, and that part is observation rather than proof, so take it as such. What rises now is people. Faces, voices, somebody describing a thing that actually happened to them, with the detail you only get from having been in the room. What sinks is the sourceless, nobody-in-particular piece that reads like it was assembled rather than written.
People engage with people. That was always true. It has just become measurable, now that the alternative costs nothing.
The half nobody is policing
So the feed gets a button, a classifier, a reach penalty and a public conversation about standards. Good. Now look at where the identical defect is running with none of that.
Your team is already using AI to make commercial decisions. Sizing a discount. Picking which segment to chase this quarter. Drafting the answer that goes to the board. Working out why churn moved. Daily, privately, in whatever tool happens to be open. That was settled by behaviour long before anybody wrote a policy about it, and no policy is going to unsettle it.
Ask a model that knows nothing about your business and it answers anyway, because answering is what it is built to do. It fills the gap with plausible generality and hands it over with total confidence. That is confident garbage, and the danger is precisely that it does not look like garbage. A visible gap invites a question. A fluent wrong answer ends the search and goes straight into the decision.
Nobody reports that one, because it was never a post. No classifier, no reach penalty, no comment section pointing out that it reads like a machine. It just becomes what your company believes. Multiply it across a commercial team and you have a hundred private, confidently argued, mutually contradictory answers to the same question, and a strategy nobody agreed to.
LinkedIn gave the feed a slop button. Your commercial decisions do not have one.
Your feed is being cleaned up for you. Your pipeline is not.
What grounding actually means
The fix for both halves is the same, and it is a specification rather than a sentiment. Your commercial context has to be:
- Structured, organised around how this company actually wins and grows customers
- Up to date, true this week rather than as of the last strategy offsite
- Verified, evidenced and traceable to something real rather than remembered
- Locked in, with a named human who decided it is true and can be asked why
Give a model that and what comes back is built out of your evidence, your positioning and your own hard-won conclusions. It sounds like your business because it is made of your business. Withhold it and you get the average of everybody else’s, which is exactly what the feed has learned to discount and exactly what makes a decision expensive.
AI with your context is leverage. AI without it is a very fast way to be wrong in your own name.
The irony, which I’m not going to hide
I used AI to build this article. It would be absurd to pretend otherwise in a piece about honesty, and worse to make the argument I’ve just made and then obscure how the thing in front of you was produced.
So, the division of labour, exactly. The machine went and found the LinkedIn announcement, the Pangram study, the UCL and Exeter research and the platform numbers, then set out what each said and where they contradicted each other. It caught one source dating a figure to the wrong year, which is why that number is not in this piece. Hours I did not have. That is the good use, and it is the same one I argued for: sweep the material, surface the pattern.
It did not decide anything. The opinions are mine and I’ll own each of them. That AI in the loop is right and AI unattended is not. That the voice cost is the worse problem and almost nobody is talking about it. That the decisions half is where the real money leaks. The judgement about what is true and worth saying, the willingness to be wrong about it in public, and the voice it is said in are the parts I cannot outsource, because they are the only parts that were ever mine.
Which is the argument, demonstrated rather than asserted. Use the machine to see further. Keep the judgement, and keep the voice.