Sara Wimmercranz, a founding partner at a Stockholm venture capital firm, caught herself opening a recent work meeting with exactly the line she had feared she'd start using: This is not a morning meeting, it's a reset. That construction — contrast framing — is a chatbot signature, and she knew it. "We just sound like bots," she told The Wall Street Journal's Katherine Bindley in a September 30, 2026 piece on what people have started calling "prompt speak." Wimmercranz also reports catching herself delivering the kind of fawning praise a sycophantic model would generate — telling colleagues they're "brilliant" or "so smart" — which she notes runs against Swedish office norms, where "the culture would just be quietly listen."
Bindley's story is built as a light cultural read: the swear jar, the whiteboard tally, the group chat full of founders whose clipped prompting style has bled into everyday speech, as in "Why no bullets?" It is funny, and the people in it are in on the joke. Deepan Mehta, who runs an AI startup in San Francisco, keeps a whiteboard where colleagues get a tally mark for saying "genuinely," "load-bearing," "smoking gun," "X is the new Y," or any "crazy hyphenated contractions that no one would normally use." Most tally marks at week's end buys happy hour.
Here is the argument I'd make that the story stops short of: this is not a verbal tic, and treating it as one is the mistake. Language convergence at scale is a leadership problem, because the specific thing a manager's voice is for — signaling conviction, marking what matters, making a message feel addressed to a particular person rather than generated for anyone — is exactly what prompt speak strips out. The joke and the whiteboard are a reasonable first response. They are not a sufficient one.
The convergence is measurable, not just anecdotal
The WSJ piece is anecdote-driven, which makes it easy to dismiss as a Silicon Valley quirk. The research says otherwise. A team led by Yakura, Lopez-Lopez and colleagues published empirical evidence of large language models' influence on human spoken communication, finding that LLM-favored vocabulary has migrated into how people actually talk — not just how they write with an assistant open. The Register covered it bluntly as a German research team warning that ChatGPT is changing how people speak, and Newsweek gave it the mainstream treatment. The mechanism is unremarkable: humans mirror the language they are soaked in, and knowledge workers are now soaked in model output for hours a day.
The sharper finding is directional. A Trends in Cognitive Sciences paper on the homogenizing effect of large language models on human expression and thought argues that the convergence narrows the range of expression available — and, its authors contend, the range of thought that expression supports. USC Dornsife's summary puts it plainly: AI may be making us think and write more alike. That is the claim worth sitting with, and it is also the one to hold with some care, since the homogenization thesis is an argument about trajectory rather than a measured effect size on any particular workforce.
Daniel Wolkowitz, who works at an AI startup in San Francisco, offered the WSJ his own version of this: "My hunch is that we're atrophying our brains. As a secondary result, I suspect that more people are starting to naturally communicate that way." He's right about the direction and probably overreaching on the mechanism. The atrophy framing is stronger than the evidence here supports. But the stylistic floor is rising while the ceiling comes down, and that compression is real.
The cost shows up as credibility, not style
If this were purely aesthetic, leaders could ignore it. It isn't. Research covered by EurekAlert on whether writing with AI at work undermines credibility found that heavy reliance on AI-written messages makes managers appear insincere. HR Dive's summary of the same work is the version to put in front of your leadership team: over-relying on AI-written messages costs managers employee trust.
Now combine that finding with the WSJ's most unsettling detail. Wolkowitz says he increasingly finds it difficult to tell whether a Slack or email message was written by a colleague using AI or by one who has unwittingly started writing like AI. If the credibility penalty attaches to the appearance of AI authorship, then it lands on both groups equally. The manager who drafted a careful note himself and the manager who pasted one out of a chatbot read the same to the recipient — and both pay.
Will Barnett, head of recruiting at a Bay Area startup, has already internalized this. He told the WSJ he edits his own writing to make it sound less like AI, a process he notes would be "the exact same workflow if I was actually just using an LLM." He also described the underlying dynamic more precisely than anyone else in the piece: "It's trained me to be more like it, instead of the reverse."
Where the convergence is actually useful
The case against prompt speak can be overstated, and the WSJ story includes its own counterweight. David Bohegan, a nonprofit CEO in Washington, D.C., argues that much of AI-speak is drawn from the same register classically trained business professionals and lawyers have used for years. "I had read 500 business books before AI came out," he said. "And AI has read those same books and 500 more." His practical point is that the models have made a previously specialist vocabulary legible to more people: "Bottom line up front... BLUF was an acronym back in the day. Now people get it more when you say that."
That's a genuine gain, and leaders shouldn't throw it out. Shared structural conventions reduce coordination cost. Allie K. Miller, an AI adviser to Fortune 500 companies, describes teams interrupting a rambling colleague with "What is the goal?" — a prompt borrowed from models that respond well to upfront goal statements. It speeds the speaker up, and Miller notes it also helps the AI agent taking meeting notes.
But notice what the second half of that sentence concedes. The meeting is being partly optimized for the transcription system. There is a difference between adopting a clarifying convention and restructuring human conversation around machine legibility, and "What is the goal?" aimed at a direct person who is thinking out loud is not a neutral intervention. Some of the most valuable things said in a meeting arrive sideways, before the speaker knows what the goal is.
Polished and forgettable is a positioning failure
The line in Bindley's story that should worry executives most comes from Angela Nibs, who has spent 20 years in PR trying to keep executives away from jargon. When she asks founders why they started their company, she reports, they no longer tell a story — they launch into three key pillars, the market opportunity, and their competitive differentiation. Her verdict: "Everyone sounds polished, but no one sounds particularly interesting."
For a founder raising capital or a CEO addressing an all-hands, that is not a style note. It is a positioning failure. Differentiation that is only asserted in the content and erased by the delivery doesn't differentiate. Miller's small lament — that she has never seen an LLM write the word "cacophony," and has new appreciation for anything written before ChatGPT launched in November 2022 — is about more than diction. Idiosyncratic word choice is one of the cheapest signals a listener has that a specific human is present and has chosen to be here.
There is also the sycophancy problem Wimmercranz flagged. Models are tuned to affirm, and leaders who absorb that cadence start handing out "brilliant" and "so smart" at a rate that devalues the currency. Praise inflation is a known management failure; the new thing is that it now arrives through imitation rather than conflict avoidance.
Most organizations have no one responsible for this
The structural issue is that nobody owns the problem. Gartner research found that only 8% of HR leaders believe their managers have the skills to use AI effectively. If fewer than one in ten HR leaders is confident their managers can use these tools well, the odds that any organization has thought carefully about where AI drafting is appropriate in people-facing communication are poor.
This is why Mehta's whiteboard is better than it looks. It is cheap, it is social rather than punitive, and it makes an invisible drift visible without requiring a policy. What it doesn't do is distinguish between contexts. A hyphenated contraction in a product spec costs nothing. The same register in a performance conversation, a layoff announcement, or a retention call with a flight-risk engineer costs trust, and that's where the credibility research bites.
What to change this quarter
A few specific moves, none of which require a tooling decision:
- Draw a no-AI line around high-stakes people communication. Performance feedback, promotion and compensation conversations, departure announcements, and anything delivered to a person about their own standing. Given the HR Dive finding on perceived insincerity, these are where the trust penalty is largest and the efficiency gain is smallest.
- Read your last ten Slack messages to your team out loud. If they contain contrast framing, "it's not X, it's Y," or stacked bullets where a sentence would do, you've drifted. Barnett's self-editing workflow is tedious but it works, and the tedium is a signal about how much you were relying on the pattern.
- Stop prompting humans. "What is the goal?" is fine in a status review. In a one-on-one or an early-stage problem-solving session, let people finish the thought. Interrupting to compress someone's reasoning optimizes for the transcript, not the thinking.
- Audit your origin story. If your answer to "why did you start this" or "why does this team exist" has become three pillars and a market-opportunity slide, as Nibs describes, rewrite it as something that happened to a person.
- Make drift a team norm, not a compliance item. The tally mark works because it's a joke colleagues can make at each other's expense. A policy document on authentic communication would not survive contact with a Tuesday.
The asset that appreciates as output gets cheaper
The convergence is going to continue — the Yakura and Lopez-Lopez study documents it already happening in speech, and the exposure that drives it is increasing, not tapering. What leaders can control is whether their own voice is one of the inputs or one of the outputs. The compression the homogenization research describes creates a straightforward arbitrage: when fluent, structured, confident prose becomes free and universal, the scarce thing is a person who sounds like themselves. Wimmercranz heard it in her own mouth and flinched. The flinch is the useful part. If you haven't noticed it in your own writing yet, assume your team has.
