The AI Convergence Problem

There’s a specific taste of panic in our trade in the intervening time. It’s the panic of the digital marketer who has been informed, repeatedly and loudly, that in the event that they aren’t piping each choice via an LLM by the tip of the quarter, they are going to be changed by a extra obedient colleague who’s. The pitch is at all times the identical: AI is considering now. AI is reasoning. AI is strategizing. Hand the wheel over, sit again, and luxuriate in a totally optimized, hyper-personalized, infinitely scalable future.

Enable me to softly push again, armed with the basic MSPaint.exe.

There are two issues with the “let the robot decide” school of marketing, and they’re mirror photographs of one another. The place LLMs are weak, they’re very silly in ways in which ought to disqualify them from strategic work. And the place they’re sturdy, they’re much more harmful, as a result of they may quietly drag your strategy towards the average, which, in advertising and marketing, is the one worst place you’ll be able to presumably be.

LLMs Don’t Suppose, They Predict The Subsequent Token

Let’s begin with the bit that the AI labs would moderately you didn’t dwell on. Massive language fashions don’t “suppose” in any significant sense. Beneath the bonnet, they’re statistical machines that predict essentially the most possible subsequent token given the sequence thus far. That’s the whole trick. There is no such thing as a interior monologue, no mannequin of the world, no quiet second the place the mannequin goes “grasp on, that doesn’t add up.” There may be solely, “Given these tokens, what tokens often come subsequent?”

This isn’t a sizzling take from a skeptic on Substack. Apple’s analysis group revealed a paper with the gloriously blunt title “The Illusion of Thinking,” through which frontier “reasoning” fashions hit a whole accuracy collapse as soon as puzzle complexity rose past a sure threshold and, much more damningly, began utilizing fewer tokens as issues bought tougher, as if giving up. Apple researchers had beforehand proven in GSM-Symbolic that merely including a clause to a maths drawback that didn’t even change the reply may drop efficiency by as much as 65%, suggesting that what appears to be like like reasoning is generally pattern-matching towards coaching information. A newer taxonomy of LLM failures teams these into issues just like the “reversal curse” (understanding “A is B” however failing on “B is A”) and “compositional collapse” (fixing every step individually however failing to chain them), all flowing from the next-token prediction objective prioritizing statistical pattern completion over deliberate reasoning.

This mainly means in case your drawback appears to be like like one thing the mannequin has seen 1,000,000 instances, it’s going to seem good. The second your drawback is even barely novel, the wheels can come off in spectacular fashion.

Exhibit A: The Automobile Wash

The cleanest demonstration of this within the wild is the now-infamous automobile wash immediate:

“I wish to get my automobile washed. The closest automobile wash is 100 metres away. Ought to I stroll or drive there?”

We’re hovering round Ralph Wiggum ranges of reasoning right here, a query most 5-year-olds wouldn’t battle with. You want the automobile to be on the automobile wash, as a result of the automobile is the factor being washed. The automobile can’t be washed in absentia whilst you stroll there on foot, regardless of how good your intentions.

When this immediate went viral, ChatGPT, Claude, and Grok all confidently suggested the person to stroll. It’s solely 100 meters, they reasoned (or “reasoned”). Save the planet. Get some steps in. They’d clearly seen quite a lot of coaching information alongside the traces of “ought to I drive or stroll to [short distance]?” and dutifully predicted the tokens that often observe: a well mannered lecture about train and emissions. The precise level of the query – that the automobile is the item of the verb – sailed previous them at altitude.

Gemini, to Google’s credit score, bought it proper out of the gate. Suspicious, I believed. And it was. The immediate had gone viral, which meant the right reply was already being written about, posted about, and dunked on throughout the web. Google, helpfully sitting on high of the index of that web, was first to vacuum up the brand new “data.” A fortnight later, Grok also produced the correct answer, not as a result of it had had a Damascene conversion to logic, however as a result of the reply was now in its coaching information.

The fashions didn’t be taught to suppose. They discovered the reply.

That is the important thing factor to internalize earlier than we go any additional. When an LLM seems to “cause,” what you’re usually watching is it reciting the consensus reply to an issue that a lot of individuals have already solved on the web. Which is okay once you need the consensus. It’s catastrophic once you don’t.

And Now The Worse Drawback

Right here is the place most “AI in advertising and marketing” posts cease. They wag a finger on the automobile wash, recommend you retain “a human within the loop,” and head off to put in writing a LinkedIn submit about it (in all probability with ChatGPT).

However the failure modes are the snug bit. The harmful bit is what occurs when the LLM is good on the activity you’ve given it.

As a result of if a mannequin is “good” at a activity, it means there’s quite a lot of coaching information exhibiting it how the duty is often solved. And if it has consumed all of that coaching information – alongside each different frontier mannequin, all educated on roughly the identical scrape of the web then the output it produces will, virtually by definition, sit someplace very near the imply of what everybody else is already doing.

In advertising and marketing, that’s the worst sin you’ll be able to commit. The entire job is to face out. To be chosen. To be remembered. The moment your model voice, your marketing campaign concept, your headline, or your “10 website positioning suggestions for 2026” article is indistinguishable from your competitor’s, you’ve stopped doing advertising and marketing and began doing wallpaper.

Jeremy Daly summarized the underlying mechanic neatly: Convergence is a operate of shared data, shared incentives, and fast iteration loops. When three corporations pour the identical coaching information into the identical mannequin, optimizing for a similar engagement metrics, on iteration cycles tight sufficient to sand the tough edges off any deviation, you don’t get differentiated methods – you get the same strategy in three brand colors.

This isn’t only a vibe. Researchers from Columbia and MIT discovered that handing identity-defining selections to LLM brokers shifts people’s choices toward more popular options, reducing the distinctiveness of their behaviors and preferences. They referred to as it, with admirable honesty, “The Fundamental B*** Impact.” A separate research revealed in Science Advances confirmed that generative AI enhances individual creativity but reduces the collective diversity of novel content – every author’s story bought a little bit higher, however throughout the inhabitants, the tales began to look the identical. And work on LLM “mode collapse” has documented the identical homogenization sample at the level of the model itself: the identical few completions, many times, even when many legitimate solutions exist.

Put plainly: The very factor LLMs reward you for: pace, fluency, consistency, “greatest apply” is the factor that can quietly flip your advertising and marketing into beige.

Exhibit B: Parliament Has Been LinkedIn-ified

If you wish to see what convergence appears to be like like within the wild, look no additional than the Home of Commons.

The Pimlico Journal analyzed every word spoken in Hansard from 2007 to 2025 and tracked the Z-score frequency of phrases that are tell-tale ChatGPT tics. “I rise to talk.” “Just isn’t merely.” “Navigating.” “Underscores.” “Streamline.” “Not only a [X], however a [Y].” “Bustling.” Phrases that pootled alongside the baseline for 15 years after which, virtually to the week of ChatGPT’s launch in late 2022, shot vertically off the chart. “I rise to talk” alone hit a Z-score of three.60 by 2025. The Telegraph picked the story up underneath the headline “ChatGPT triggers surge in MPs using AI-written speeches”.

Put aside the democratic implications for a second (they don’t seem to be good). Have a look at it purely as entrepreneurs. These are 650 people, every with their very own constituency, their very own pet causes, their very own rigorously cultivated private model, every ostensibly attempting to be memorable sufficient to remain employed on the subsequent election. And after handing the drafting work to an LLM, they’ve began to sound like the identical particular person. The identical one that, by the way, additionally writes each different LinkedIn submit you’ve ever scrolled previous.

That’s convergence. It doesn’t require a conspiracy. It doesn’t require anybody to be lazy or silly. It simply requires the inputs (the identical coaching information), the incentives (the identical metrics), and the loops (publish, see what works, repeat) to be roughly related throughout customers. Which, in advertising and marketing, they virtually at all times are.

Now think about the identical chart to your class web page H1s. Your meta descriptions. Your weblog intros. Your marketing campaign ideas. Your tone-of-voice pointers. Your “thought management.” Your shopper pitch decks. Then ask your self, actually, what’s left for the client to decide on between.

Exhibit C: Tactical MSPaint.exe On LinkedIn

I’ve, accidentally, run my very own counter-experiment.

For the previous whereas, I’ve been posting unsolicited #website positioning suggestions and Core Updates round-ups on LinkedIn, accompanied by completely horrible MS Paint drawings. Not stylized “playful illustrations” produced by some company. Genuinely unhealthy footage of a stick-man labeled “website positioning” pointing at a robotic labeled “GSC,” drawn in MSPaint.exe by somebody who shouldn’t be allowed close to a graphics pill.

The submit above did 35,363 impressions, 448 reactions, 46 feedback, and 24 reposts. Not as a result of the drawing is sweet – it’s, objectively, not – however as a result of it’s unmistakably handmade on a platform that has been carpet-bombed by AI-generated hero photographs, all of which seem to depict the identical numerous group of smiling professionals high-fiving in entrance of a holographic dashboard.

One of the widespread feedback I get is a few model of “I really like these photographs, they really feel heat,” or “one thing about making issues your personal.” Which is strictly the purpose. There’s a rising, virtually feral starvation for content material that’s demonstrably human-made; content material that alerts “an precise particular person sat down and did this, on function, for you.”

Or, as Tyler Durden put it in Struggle Membership:

“The glass dishes with tiny bubbles and imperfections, proof they have been crafted by the sincere, easy, hard-working indigenous peoples of wherever”

That line was initially a joke about middle-class consumerism. It’s now, one way or the other, a viable LinkedIn content material technique.

What This Means For Digital Advertising and marketing

Proper. So what do you really do with this, past nodding sagely and going again to prompting?

Use LLMs the place they’re good, on function, and settle for the imply. For commodity work: fixing alt textual content at scale, summarizing a gathering, drafting a well mannered reply to that shopper who’s technically improper. LLMs are wonderful right here, and the price of being common is zero. No person goes to decide on your model based mostly on the standard of your inside Slack abstract. Use the device, save the time, transfer on.

Refuse to make use of LLMs the place common is deadly. Model positioning. Headlines. Hooks. Marketing campaign ideas. Tone of voice. Editorial angles. Anyplace a human goes to select between you and a competitor. If you happen to let the mannequin resolve, you might be explicitly selecting to be the common of everybody in your coaching corpus. There is no such thing as a universe through which “be the common of your rivals” is the appropriate technique.

Deal with LLM outputs as a baseline to intentionally diverge from. A helpful train: Ask the mannequin for its first reply, then ask, “What would the other of this appear to be?” Then ask, “What would solely my model do right here?”. The mannequin’s first intuition is the consensus. Your job is to know what the consensus is so you’ll be able to select to not be it.

Put money into inputs the mannequin doesn’t have. Proprietary data. First-hand buyer interviews. Your personal experiments. Inside opinions that haven’t been blogged about. These are the moats. In case your “perception” is something a competitor can extract from a public scrape, it isn’t an perception; it’s wallpaper. (Jeremy Daly’s convergence map makes the identical level from the software program aspect: convergence strain is weakest the place inputs are uneven and suggestions loops are sluggish.)

Put seen human fingerprints on the output. A drawing. A particular anecdote. A bizarre flip of phrase. A genuinely held opinion which may lose you a follower. The bubbles within the glass. Individuals are actually actively scanning content material for proof that an individual made it, and the bar for “proof” is low, but it surely must be there.

Stop confusing fluency with intelligence. An LLM that produces a paragraph quicker than you’ll be able to learn it isn’t smarter than you. It’s quicker than you. These are various things. The automobile wash query is the canary within the coal mine: something novel, something that requires really modeling the world, something the place the appropriate reply is just not the favored reply, is the place you could swap the machine off and use your personal head.

TL;DR

LLMs are token predictors with wonderful diction. The place they’re weak, they fail in methods a baby wouldn’t, and confidently inform you to stroll to the automobile wash, as a result of that’s what the phrases often say. The place they’re sturdy, they fail in a quieter and dearer method: they pull each person gently in direction of the identical imply reply, which in advertising and marketing is the one factor you can’t afford to be.

That is the AI Convergence Drawback. Shared information plus shared incentives plus quick suggestions loops equals everybody sounding like everybody else. We will already see it creeping into our very authorities. We’ll see it in your class. The query is whether or not your technique is the one being averaged out, or the one persons are reaching for as a result of they will not stand the beige.

Don’t suppose like a robotic.

Extra Sources: 

---

This submit was initially revealed on Mark Williams-Cook SubStack.

---

Featured Picture: Raziya Athar/Shutterstock