The Great Flattening, Part 2: The Data Is Worse Than the Anecdotes

Earlier this year, we published The Great Flattening. Our argument was straightforward: AI is making every company sound the same. We pointed to data from Barron's and AlphaSense showing that the phrase "not just X, it's Y" appeared in 73 corporate documents in a single quarter after barely registering across the previous two decades. Progressive, Citizens Financial, Synopsys, Royal Caribbean, A.O. Smith. All reaching for the same construction. All sounding like the same company.

That piece got a reaction. Readers shared it. CMOs forwarded it to their content teams. We heard from brand leaders who said they'd already noticed the problem but hadn't been able to name it.

But we'll be the first to admit: we were working mostly from pattern recognition and corporate filings. We could see the convergence happening. We couldn't prove the mechanism behind it.

Since then, peer-reviewed science has landed. Multiple studies, from multiple institutions, all examining the same question we raised. And the findings are worse than what we originally described.

The Artificial Hivemind

A research team led by Liwei Jiang at the University of Washington, Carnegie Mellon, and the Allen Institute for AI published what became the Best Paper at NeurIPS 2025. They gave their finding a name: the "Artificial Hivemind."

The team tested over 70 large language models using a dataset of 26,000 real-world, open-ended queries, the kind of things people actually type into ChatGPT, Claude, Gemini, and other tools every day.

When they asked 25 different models to write a metaphor about time, each generating 50 responses, the outputs didn't spread across the conceptual space the way human answers would. They collapsed into two clusters. One group wrote variations of "time is a river." The other wrote variations of "time is a weaver." Different architectures, different companies, different continents, two ideas.

The specifics are even more telling. When asked to write a product description for iPhone cases, DeepSeek-V3 (built in China) and OpenAI's GPT-4o (built in San Francisco) produced identical phrases: "Elevate your iPhone with our," "sleek, without compromising," and "with bold, eye-catching." The measured similarity between these two models sat at 81%. DeepSeek and Alibaba's Qwen hit 82%.

These are not the same model with different branding. These are independent systems developed on opposite sides of the planet, arriving at the same words.

When we wrote The Great Flattening, we said every company was starting to sound like ChatGPT. The Artificial Hivemind study revealed something we hadn't considered: ChatGPT sounds like Claude sounds like Gemini sounds like DeepSeek sounds like Qwen. The convergence isn't just happening at the brand level. It's baked into the tools. The infrastructure itself is homogeneous.

That realization changed how we think about this problem at The State of Brand. We'd been telling our readers that the flattening was a choice, that companies were being lazy with their AI usage. The Hivemind data suggests it's closer to a default. You have to actively fight the convergence, because the tools are pulling everything toward center on their own.

Better Prompts Won't Save You

This is the section we wish we didn't have to write, because it undermines the most common advice in B2B marketing right now.

In March 2026, Emily Wenger and Yoed N. Kenett published "Large Language Models Are Homogeneously Creative" in PNAS Nexus. They ran standardized creativity tasks across multiple LLMs and compared the diversity of outputs at a population level. LLM responses mirrored other LLM responses far more than human responses mirrored other human responses, even after controlling for structure and other variables.

Then they tried to fix it. They adjusted the "temperature" setting, the parameter that controls randomness in text generation. Higher temperature should, in theory, produce more varied outputs.

It didn't. As Wenger told PsyPost, she "was surprised by the degree" of homogeneity. Raising the temperature didn't produce more creative outputs. It produced gibberish. The models went from repetitive to incoherent with nothing usable in between. There was no setting to toggle that restored genuine diversity.

A separate study published in ScienceDirect, analyzing 2,200 college admissions essays across three preregistered experiments, landed on the same conclusion. The diversity gap between human-written and AI-written essays didn't just hold steady. It widened as volume increased. More output, more sameness. And the researchers specifically tested the interventions that AI consultants are currently selling to marketing teams: better prompts, adjusted parameters, more specific instructions. None of them closed the gap.

We want to be direct about what this means for the brands we cover. The most popular advice in B2B content strategy right now is "use AI but make it sound like you." The research suggests that's much harder than anyone selling that advice is willing to admit. The tools are architecturally inclined toward sameness. Fighting that inclination requires more than a brand voice doc uploaded to a custom GPT.

The Flattening Goes Deeper Than Language

If homogeneous outputs were the whole story, this would be a content quality issue. Annoying but manageable. The newer research says it goes further.

Recently, a review in Trends in Cognitive Sciences, one of the most respected journals in cognitive science, argued that LLMs promote "stylistic and conceptual homogenization while downplaying alternative voices" and that left unchecked, this risks "flattening the cognitive landscapes that drive collective intelligence and adaptability."

The researchers aren't saying AI makes bad content. They're saying AI narrows the range of ideas that circulate. Over time, as people read and internalize AI-generated text, the boundaries of what feels normal or expected tighten and the conceptual space shrinks.

The review also flagged something our original piece didn't touch: cultural homogenization. Research cited in the paper found that AI suggestions push writing specifically toward Western, English-language styles and flatten cultural nuances. For any B2B brand operating internationally, this has real implications. The AI isn't just making your content generic. It's making your content generically American, even when your audience isn't.

Then in April 2026, a meta-analysis pulling from over 130 studies concluded that while AI boosts individual productivity, it "subtly shapes and homogenizes human expression, thought patterns, and group creativity."

Your Buyers Already Know

When we wrote the original Great Flattening, we framed the opportunity in brand strategy terms: sounding different is a competitive advantage. Since then, consumer research has put numbers on what happens when brands sound the same. The cost is measurable and it shows up as lost trust.

Klaviyo's 2026 AI Consumer Trends Report, drawn from a survey of 8,000 consumers across eight countries, found that only 13% completely trust AI. When consumers notice AI-generated content in brand marketing, they are four times more likely to trust the brand less than more: 31% reported decreased trust, versus just 7% who said trust increased.

SmythOS research found that half of consumers can now correctly identify AI-generated content. When they do, 52% disengage. Not occasionally. As a pattern.

A 2026 survey cited by the New York Post found that 54% of Americans report experiencing "AI fatigue." Separately, 43% of users say they no longer trust most online content at all, per O'Dwyer's PR News.

The number that stuck with us the most came from Klaviyo's "AI Enthusiasts" segment: people who actively use AI, like it, and integrate it into their daily routines. Even among that group, 39% said they would trust a brand less for using AI-generated content, and 38% said they run into what they called "AI slop" from brands multiple times a week.

If the people who love AI are turned off by AI-generated brand content, the rest of your audience has already moved on.

We've been saying this to the brands we work with privately for months: the distance between "we use AI internally" and "our audience can tell" is shrinking fast. The Klaviyo data confirms it.

This Gets Worse Over Time

One thing we didn't fully appreciate when we wrote Part 1 is how the Great Flattening compounds.

Research on what scientists call "model collapse" shows that as AI models train on increasingly AI-generated content (which now constitutes a growing share of what's published online), their outputs become more homogeneous with each generation. The models ingest their own outputs. The variance narrows. The center of gravity pulls tighter.

For B2B brands, this creates a cycle we've started describing to our readers as a sameness spiral. Your competitors use the same AI tools to produce content that sounds the same. That content floods the internet. The next generation of models trains on that content. The next round of brand communications converges further. And 74.2% of new web pages already contain detectable AI-generated content, so the training data is getting less diverse, not more.

The floor keeps rising toward identical. Every quarter that passes without a deliberate intervention makes the problem harder to reverse.

Where Our Original Thesis Holds, and Where We'd Update It

We've been doing some honest reflection at The State of Brand about what the original Great Flattening got right and where it fell short.

The diagnosis holds. Corporate language is converging and AI is the primary driver. The brands that refuse to participate in the convergence have a genuine strategic advantage.

Where we'd update our thinking: we framed the flattening primarily as a discipline problem. We implied that companies were being careless, running their communications through ChatGPT without enough thought and accepting whatever came back. The fix, we suggested, was better prompting, stronger brand guidelines, and tighter editorial oversight.

The research published in the months since tells us the problem is more structural than that. The Artificial Hivemind study demonstrates that different models built by different companies produce strikingly similar outputs no matter how you prompt them. The PNAS Nexus study shows that adjusting model parameters doesn't restore diversity. The college essay research shows that the gap widens with scale. And the Trends in Cognitive Sciences review makes the case that the homogenization reaches past text and into how people think after sustained exposure to AI-generated language.

Better prompts help at the margins. Brand voice guidelines help more than that. But neither one addresses what's actually happening: the tools converge toward the statistical center of all language ever written. That convergence is a feature of the architecture, not a bug in anyone's workflow.

Which brings us back to the core argument we made in Part 1, only now we'd state it more forcefully. If you want to sound different, you cannot get there by using AI differently. You need to have something different to say before the AI ever touches it. The point of view has to exist first. The brand has to have a voice worth preserving, or there's nothing for any amount of prompt engineering to protect.

The Window Is Open. It Won't Be Forever.

We ended the original Great Flattening with a line we still believe: the companies that sound like themselves will own the next decade.

Six months later, with a NeurIPS Best Paper, a PNAS Nexus study, a Trends in Cognitive Sciences review, a 130-study meta-analysis, and a wave of consumer trust research all confirming what we saw in those AlphaSense filings, we'd say it with more conviction.

Most B2B companies are still running their content through the same tools with the same defaults and landing on the same outputs. The brands choosing to sound like themselves stand out more each week, because the background noise is getting more uniform, not less.

But the advantage has a shelf life. As more companies recognize the flattening and start investing in distinctive voice, the early movers will have already built the recognition and credibility that latecomers will have to fight for. The research we've laid out in this piece makes the timeline clear: AI homogenization is structural, it compounds, and audiences are already responding to it with less trust and less attention.

We've been covering brand strategy long enough to know that competitive advantages rarely announce themselves this clearly. If your company has real clarity about who it is, what it believes, and why it exists, every piece of flattened content from your competitors is making your voice louder by contrast.

That was true when we wrote Part 1. The science now says it's even more true than we realized.

We'll keep tracking this. If you're a brand leader working on this problem, we want to hear how it's going.