What’s the stuff I can’t stand about Emily in Paris? Paris as a duty-free display case. A lead whose cultural ignorance keeps getting filmed as pluck. Conflicts that show up and clear out inside one episode like somebody in the room was fretting about viewers’ blood pressure.

Darren Star, he of Sex in the City provenance, did all of that on purpose. Then said he was surprised anyone was offended, that he was only poking at clichés everybody already knows.

Fine. It’s still a choice, and it’s the kind of choice a large language model would’ve walked back from. That’s why I laugh when you see producers talking about their AI-scripted or visualized productions.

Feed that script to any model and ask it to punch things up. You’ll get a more careful Emily. The French characters get inner lives and arcs. The office gets real stakes. Somebody learns something.

Now it’s a show nobody can attack and nobody can remember, and it doesn’t pull 58 million households, because the things I hate are the thing that made it travel.

That’s convergence. Not robots writing badly. Robots writing to the middle.

Nobody complains about the middle

You’ll miss it, because the individual output looks better.

Doshi and Hauser ran this exercise in Science Advances back in 2024. Writers who got story ideas from an LLM turned in stories that reviewers scored more creative, better written, and more enjoyable than what the unassisted writers managed. Then the same paper counted the thing that, well counts. The AI-assisted stories were more like each other. The writing got narrower. The “shape” of the story became the same across each story.

Padmakumar and He at NYU found the culprit. In their ICLR study, essays co-written with a raw base model lost no diversity at all. Essays co-written with the feedback-tuned model did, with different authors sidling toward each other. The polish and the sameness show up in the same box. Same training step. You don’t get one without the other.

And if your plan was “we’ll just prompt it better,” excuse me while I ruin that idyll: Wenger and Kenett put 22 different models up against 102 humans in PNAS Nexus this March. The model answers clumped together way tighter than the human ones, and they stayed clumped after the researchers controlled for how the answers were structured.

Switching AI vendors won’t save you, either. The models look more like each other than any of them looks like us.

They tried prompting their way out, too. Telling a model to be imaginative and bold nudged its individual score and did nearly nothing about the variability. Cranking the temperature widened things up right until the answers stopped being in comprehensible human language.

So there’s your AI amplitude dial: Consensus on one end, gibberish on the other, nothing usable in between.

Five shows that wouldn’t have survived an AI opinion

• Seinfeld. The attempt to sand down the edges actually happened in 1989…and somebody kept the memo. NBC ran the pilot past 400 test viewers who called it “weak”, which an executive later downgraded to a “weak weak”. No part of the audience wanted to watch it again. Nobody liked the supporting characters. Viewers felt Jerry needed a better backup ensemble, still my favorite sentence ever written about Jason Alexander, Julia Louis-Dreyfus, and Michael Richards. Warren Littlefield called it a dagger to the heart and admitted they were scared to move on something research had rejected that hard. Think about it in AI terms: A 400-person panel is just a low-resolution language model. It gave the notes a model gives: make it warmer, get him likable friends, quit filming the laundromat.
• The Sopranos. Nine years, and it ends on a cut to black in the middle of a sentence. Nothing that optimizes for anything tolerates an unresolved ending, and “give the audience closure” is a perfectly good note. Take it, and you hand back the twenty years of arguing that have kept the show top-of-mind since that night.
• Fleabag. The looks at the camera aren’t just a cute conceit. They’re where she goes to not be in the room with others, which is why season two blows the whole thing up by having the priest catch her doing it. Any consensus-driven reviewer flags fourth-wall breaks as a gimmick that holds the audience at arm’s length. Cut it in draft two and the show falls over.
• Breaking Bad, “Fly.” Mid-season, mid-plot, one hour, two guys, one room, one housefly. Plenty of people hated it in real time. Nothing that’s optimizing for retention greenlights that episode. Cut it and the plot loses nothing, which is precisely why it gets cut.
• Twin Peaks: The Return, Part 8. Twenty-odd minutes of black-and-white nuclear abstraction, no dialogue, no plot, dropped into a network revival. That’s as far out as this argument goes. No model and no stakeholder review on earth would permit it to pass.

Emily in Paris and those five flunk the same test from opposite directions. Sand down Emily and she gets less annoying. “Polish” the other five and they’re gone. What’s being removed isn’t quality. It’s variance, but variance is the only part anybody remembers.

Your homepage has blandwidth

The B2B version of this hides in plain sight. That’s because nobody hate-watches a homepage.

Converged B2B copy doesn’t come out bad. It comes out just fine. Category name in the H1, three benefit pillars, a logo bar, trusted by leading revenue teams, request a demo. Legal signs off. Brand signs off. The CEO signs off. Then your buyer opens four tabs, reads four versions of the same page from four competitors, can’t tell you apart, and goes with whoever’s cheaper or whoever G2 mentioned first.

Call it blandwidth. The share of your content that’s alive because nobody in the review cycle had a problem with it.

You can measure the drift into AI-driven blandwidth now, by the way. Kobak’s group at Tübingen combed more than 15 million biomedical abstracts and found at least 13.5% of 2024 abstracts had been run through an LLM, hitting 40% in some subfields. The shift in scientific vocabulary was bigger than the one COVID caused. The tells were the abuse of words like delve, underscore, pivotal, realm. Biomedicine is a field where grown adults fight about comma placement, and it drifted that quickly. Your category page has drifted a lot faster.

Differentiation isn’t the costly part for a company. Citation is.

Generative engines pull from sources that tell them something they don’t have. Copy that recites the category consensus is, definitionally, stuff the model already knows, so there’s no reason to go get it and no reason to name you when the answer comes out. Consensus copy doesn’t just fail to stand out. It deletes the only reason ChatGPT would ever mention your company.

You optimized your way out of the citation.

Which is where clarity quits on you. A page can post a perfect 5 on the Clarity lever of a Messaging Friction Audit and still be swappable with anyone else’s, because clarity only tells you the buyer understood the sentence. Nothing on that page tells you if they could name the vendor an hour later.

For Vendavo, we built a campaign around a number people could argue with: A trillion dollars in profit left on the table due to bad pricing. It brought in 80-plus MQLs and more than 1,200 press mentions. Run that through a consensus review cycle and a trillion softens into significant untapped revenue, because a trillion is the kind of number that starts fights.

Starting the fight was the whole distribution plan. Nobody’s forwarding a link to significant untapped revenue.

So what’s the point, Emily?

Convergence isn’t ruining most B2B content. For most teams it’s the reason the content exists in the first place. Two people covering an entire content function will ship something with these tools and nothing without them, and that’s not a tragedy.

For a good chunk of any portfolio, the standard answer is the right answer. Release notes. Integration docs. Comparison tables. Pricing FAQs. Nobody in recorded history has wanted a distinctive password-reset page. Converge all of it, happily, and go home early.

So the test isn’t which tool you used. It’s what job the piece is doing. If your reader came looking for the standard answer, give them the standard answer and don’t get precious about it. If the piece is supposed to explain why anyone should pick you, then sounding like the standard answer is how you lose the deal. That call gets made asset by asset. In practice, though, one person decides it once for everything.

Whoever owns the content calendar builds an AI production workflow under a volume target (a prompt, a template, or a model they like) and that becomes the default for the blog, the case studies, the category pages, and everything else in the queue. Nothing forces a second look because you’re too busy spitting out content. Two quarters later organic is flat, none of the pipeline traces back to any of it, and the prompt still hasn’t been questioned.

Pull up the last thing you published and find the claim holding it up. Could a competitor slap their logo on that claim and ship it untouched? Then it was never yours. It belongs to the category, and the model already has it.

Here’s a harder one: Could anybody reasonably disagree with it? Because if there’s nothing in there a smart person could fight you about, there’s nothing in there worth citing either.

I still think Emily in Paris is bad. Somebody made a call I get to argue with, which already puts it a few rungs above whatever’s on most B2B blogs.