The AI Convergence Problem

There’s a peculiar spirit of panic successful our manufacture astatine the moment. It’s the panic of the integer marketer who has been told, many times and loudly, that if they aren’t piping each determination done an LLM by the extremity of the quarter, they will beryllium replaced by a much pious workfellow who is. The transportation is ever the same: AI is reasoning now. AI is reasoning. AI is strategizing. Hand the instrumentality over, beryllium back, and bask a afloat optimized, hyper-personalized, infinitely scalable future.

Allow maine to mildly push back, equipped pinch the classical MSPaint.exe.

There are 2 problems pinch the “let the robot decide” schoolhouse of marketing, and they are reflector images of each other. Where LLMs are weak, they are very stupid successful ways that should disqualify them from strategical work. And wherever they are strong, they are moreover much dangerous, because they will softly drag your strategy towards the average, which, successful marketing, is the azygous worst spot you tin perchance be.

LLMs Don’t Think, They Predict The Next Token

Let’s commencement pinch the spot that the AI labs would alternatively you didn’t dwell on. Large connection models do not “think” successful immoderate meaningful sense. Under the bonnet, they are statistical machines that foretell the astir probable adjacent token fixed the series truthful far. That is the full trick. There is nary soul monologue, nary exemplary of the world, nary quiet infinitesimal wherever the exemplary goes “hang on, that doesn’t adhd up.” There is only, “Given these tokens, what tokens usually travel next?”

This is not a basking return from a skeptic connected Substack. Apple’s investigation squad published a insubstantial pinch the gloriously blunt title “The Illusion of Thinking,” successful which frontier “reasoning” models deed a complete accuracy illness erstwhile puzzle complexity roseate beyond a definite period and, moreover much damningly, started utilizing fewer tokens arsenic problems sewage harder, arsenic though giving up. Apple researchers had antecedently shown successful GSM-Symbolic that simply adding a clause to a maths problem that didn’t moreover alteration the reply could driblet capacity by up to 65%, suggesting that what looks for illustration reasoning is mostly pattern-matching against training data. A much caller taxonomy of LLM failures groups these into things for illustration the “reversal curse” (knowing “A is B” but failing connected “B is A”) and “compositional collapse” (solving each measurement individually but failing to concatenation them), each flowing from the next-token prediction nonsubjective prioritizing statistical shape completion complete deliberate reasoning.

This fundamentally intends if your problem looks for illustration thing the exemplary has seen a cardinal times, it will look brilliant. The infinitesimal your problem is moreover somewhat novel, the wheels tin travel disconnected successful spectacular fashion.

Exhibit A: The Car Wash

The cleanest objection of this successful the chaotic is the now-infamous car lavation prompt:

“I want to get my car washed. The nearest car lavation is 100 metres away. Should I locomotion aliases thrust there?”

We’re hovering astir Ralph Wiggum levels of reasoning here, a mobility astir 5-year-olds would not struggle with. You request the car to beryllium astatine the car wash, because the car is the point being washed. The car cannot beryllium washed successful absentia while you stroll location connected foot, nary matter really bully your intentions.

When this punctual went viral, ChatGPT, Claude, and Grok each confidently advised the personification to walk. It’s only 100 meters, they reasoned (or “reasoned”). Save the planet. Get immoderate steps in. They had intelligibly seen a awesome woody of training information on the lines of “should I thrust aliases locomotion to [short distance]?” and dutifully predicted the tokens that usually follow: a polite speech astir workout and emissions. The existent constituent of the mobility – that the car is the entity of the verb – sailed past them astatine altitude.

Slide from Mark Williams-Cook’s “Do !not deliberation for illustration a robot” presentation. Image Credit: Mark Williams-Cook

Gemini, to Google’s credit, sewage it correct retired of the gate. Suspicious, I thought. And it was. The punctual had gone viral, which meant the correct reply was already being written about, posted about, and dunked connected crossed the internet. Google, helpfully sitting connected apical of the scale of that internet, was first to hoover up the caller “knowledge.” A fortnight later, Grok besides produced the correct answer, not because it had had a Damascene conversion to logic, but because the reply was now successful its training data.

The models didn’t study to think. They learned the answer.

This is the cardinal point to internalize earlier we spell immoderate further. When an LLM appears to “reason,” what you’re often watching is it reciting the statement reply to a problem that tons of group person already solved connected the internet. Which is good erstwhile you want the consensus. It is catastrophic erstwhile you don’t.

And Now The Worse Problem

Here is wherever astir “AI successful marketing” posts stop. They wag a digit astatine the car wash, propose you support “a quality successful the loop,” and caput disconnected to constitute a LinkedIn station astir it (probably pinch ChatGPT).

But the nonaccomplishment modes are the comfortable bit. The vulnerable spot is what happens erstwhile the LLM is good astatine the task you’ve fixed it.

Because if a exemplary is “good” astatine a task, it intends location is simply a awesome woody of training information showing it really the task is usually solved. And if it has consumed each of that training information – alongside each different frontier model, each trained connected astir the aforesaid scrape of the net past the output it produces will, almost by definition, beryllium location very adjacent to the mean of what everyone other is already doing.

In marketing, that is the worst misdeed you tin commit. The full occupation is to guidelines out. To beryllium chosen. To beryllium remembered. The instant your marque voice, your run idea, your headline, aliases your “10 SEO tips for 2026” article is indistinguishable from your competitor’s, you person stopped doing trading and started doing wallpaper.

Jeremy Daly summarized the underlying mechanic neatly: Convergence is simply a usability of shared data, shared incentives, and accelerated loop loops. When 3 companies move the aforesaid training information into the aforesaid model, optimizing for the aforesaid engagement metrics, connected loop cycles tight capable to soil the unsmooth edges disconnected immoderate deviation, you don’t get differentiated strategies – you get the same strategy successful 3 marque colors.

This is not conscionable a vibe. Researchers from Columbia and MIT recovered that handing identity-defining choices to LLM agents shifts people’s choices toward much celebrated options, reducing the distinctiveness of their behaviors and preferences. They called it, pinch admirable honesty, “The Basic B*** Effect.” A abstracted study published successful Science Advances showed that generative AI enhances individual productivity but reduces the corporate diverseness of caller content – each writer’s communicative sewage a small better, but crossed the population, the stories started to look the same. And activity connected LLM “mode collapse” has documented the aforesaid homogenization shape at the level of the exemplary itself: the aforesaid fewer completions, again and again, moreover erstwhile galore valid answers exist.

Put plainly: The very point LLMs reward you for: speed, fluency, consistency, “best practice” is the point that will softly move your trading into beige.

Exhibit B: Parliament Has Been LinkedIn-ified

If you want to spot what convergence looks for illustration successful the wild, look nary further than the House of Commons.

Image Credit: Mark Williams-Cook

The Pimlico Journal analyzed every connection spoken successful Hansard from 2007 to 2025 and tracked the Z-score frequency of phrases that are tell-tale ChatGPT tics. “I emergence to speak.” “Is not merely.” “Navigating.” “Underscores.” “Streamline.” “Not conscionable a [X], but a [Y].” “Bustling.” Phrases that pootled on the baseline for 15 years and then, almost to the week of ChatGPT’s merchandise successful precocious 2022, changeable vertically disconnected the chart. “I emergence to speak” unsocial deed a Z-score of 3.60 by 2025. The Telegraph picked the communicative up nether the header “ChatGPT triggers surge successful MPs utilizing AI-written speeches”.

Set speech the antiauthoritarian implications for a infinitesimal (they are not good). Look astatine it purely arsenic marketers. These are 650 individuals, each pinch their ain constituency, their ain pet causes, their ain cautiously cultivated individual brand, each ostensibly trying to beryllium memorable capable to enactment employed astatine the adjacent election. And aft handing the drafting activity to an LLM, they person started to sound for illustration the aforesaid person. The aforesaid personification who, incidentally, besides writes each different LinkedIn station you’ve ever scrolled past.

That is convergence. It does not require a conspiracy. It does not require anyone to beryllium lazy aliases stupid. It conscionable requires the inputs (the aforesaid training data), the incentives (the aforesaid metrics), and the loops (publish, spot what works, repeat) to beryllium astir akin crossed users. Which, successful marketing, they almost ever are.

Now ideate the aforesaid floor plan for your class page H1s. Your meta descriptions. Your blog intros. Your run concepts. Your tone-of-voice guidelines. Your “thought leadership.” Your customer transportation decks. Then inquire yourself, honestly, what is near for the customer to take between.

Exhibit C: Tactical MSPaint.exe On LinkedIn

I have, by accident, tally my ain counter-experiment.

For the past while, I person been posting unsolicited #SEO tips and Core Updates round-ups connected LinkedIn, accompanied by perfectly unspeakable MS Paint drawings. Not stylized “playful illustrations” produced by immoderate agency. Genuinely bad pictures of a stick-man branded “SEO” pointing astatine a robot branded “GSC,” drawn successful MSPaint.exe by personification who should not beryllium allowed adjacent a graphics tablet.

A objection of MSPaint.exe connected LinkedIn SEO tips

The station supra did 35,363 impressions, 448 reactions, 46 comments, and 24 reposts. Not because the drafting is bully – it is, objectively, not – but because it is unmistakably handmade connected a level that has been carpet-bombed by AI-generated leader images, each of which look to picture the aforesaid divers squad of smiling professionals high-fiving successful beforehand of a holographic dashboard.

One of the astir communal comments I get is immoderate type of “I emotion these images, they consciousness warm,” aliases “something astir making things your own.” Which is precisely the point. There is simply a growing, almost feral hunger for contented that is demonstrably human-made; contented that signals “an existent personification sat down and did this, connected purpose, for you.”

Or, arsenic Tyler Durden put it successful Fight Club:

“The solid dishes pinch mini bubbles and imperfections, impervious they were crafted by the honest, simple, hard-working indigenous peoples of wherever”

That statement was primitively a joke astir middle-class consumerism. It is now, somehow, a viable LinkedIn contented strategy.

What This Means For Digital Marketing

Right. So what do you really do pinch this, beyond nodding sagely and going backmost to prompting?

Use LLMs wherever they are good, connected purpose, and judge the mean. For commodity work: fixing alt matter astatine scale, summarizing a meeting, drafting a polite reply to that customer who is technically wrong. LLMs are fantabulous here, and the costs of being mean is zero. Nobody is going to take your marque based connected the value of your soul Slack summary. Use the tool, prevention the time, move on.

Refuse to usage LLMs wherever mean is fatal. Brand positioning. Headlines. Hooks. Campaign concepts. Tone of voice. Editorial angles. Anywhere a quality is going to make a prime betwixt you and a competitor. If you fto the exemplary decide, you are explicitly choosing to beryllium the mean of everyone successful your training corpus. There is nary beingness successful which “be the mean of your competitors” is the correct strategy.

Treat LLM outputs arsenic a baseline to deliberately diverge from. A useful exercise: Ask the exemplary for its first answer, past ask, “What would the other of this look like?” Then ask, “What would only my marque do here?”. The model’s first small heart is the consensus. Your occupation is to cognize what the statement is truthful you tin take not to beryllium it.

Invest successful inputs the exemplary does not have. Proprietary data. First-hand customer interviews. Your ain experiments. Internal opinions that haven’t been blogged about. These are the moats. If your “insight” is thing a competitor tin extract from a nationalist scrape, it is not an insight; it is wallpaper. (Jeremy Daly’s convergence map makes the aforesaid constituent from the package side: convergence unit is weakest wherever inputs are asymmetric and feedback loops are slow.)

Put visible quality fingerprints connected the output. A drawing. A circumstantial anecdote. A weird move of phrase. A genuinely held sentiment that mightiness suffer you a follower. The bubbles successful the glass. People are now actively scanning contented for grounds that a personification made it, and the barroom for “evidence” is low, but it has to beryllium there.

Stop confusing fluency pinch intelligence. An LLM that produces a paragraph faster than you tin publication it is not smarter than you. It is faster than you. Those are different things. The car lavation mobility is the canary successful the ember mine: thing novel, thing that requires really modeling the world, thing wherever the correct reply is not the celebrated answer, is wherever you request to move the instrumentality disconnected and usage your ain head.

TL;DR

LLMs are token predictors pinch fantabulous diction. Where they are weak, they neglect successful ways a kid wouldn’t, and confidently show you to locomotion to the car wash, because that’s what the words usually say. Where they are strong, they neglect successful a quieter and much costly way: they propulsion each personification mildly towards the aforesaid mean answer, which successful trading is the 1 point you cannot spend to be.

This is the AI Convergence Problem. Shared information positive shared incentives positive accelerated feedback loops equals everyone sounding for illustration everyone else. We tin already spot it creeping into our very government. We will spot it successful your category. The mobility is whether your strategy is the 1 being averaged out, aliases the 1 group are reaching for because they tin nary longer guidelines the beige.

Don’t deliberation for illustration a robot.

More Resources: 

• AI Search Is Eating Itself & The SEO Industry Is The Source
• 3 Strategies That Can Survive AI Search In 2026: What I Shared At SEJ Live
• How AI Chooses Which Brands To Recommend: From Relational Knowledge To Topical Presence

This station was primitively published connected Mark Williams-Cook SubStack.

Featured Image: Raziya Athar/Shutterstock