Linguistic Murmuration
Photo by James Wainscoat on Unsplash
Watch a murmuration of starlings and you'll see thousands of birds move as one — no leader, no signal, just local alignment propagating through the flock until the whole thing turns in the same direction at the same time.
The AI industry does the same thing with language.
The pattern
A term appears. Maybe in a paper, maybe in a product launch, maybe in a tweet that gets 40,000 impressions. Within weeks it's in every pitch deck, every landing page, every LinkedIn post. Not because everyone agreed on what it means, but because everyone agreed it sounds right.
Agentic. Orchestration. RAG. Copilot. Memory. Context window. Guardrails.
Each word starts precise. Each word ends vague. The flock converges on the vocabulary before it converges on the meaning underneath.
I've started calling this linguistic murmuration: the rapid, leaderless alignment of an industry around shared terminology that hasn't been shared in definition.
The term first appeared in this reply to Tracy Chou on May 18, 2026, the day before this post was published.
Why it matters
When a customer reads "AI memory" on five different product pages, they're reading five different claims. One means a vector database. One means a conversation log. One means a fine-tune. One means a prompt cache. One means whatever the marketing team thought sounded closest to what engineering actually built.
The words are identical. The products are not.
This isn't dishonesty — it's convergence without coordination. Each company picks the term that most closely approximates what they've built, and because the approximation is close enough, the term sticks. After enough adoption, the word belongs to the flock, not to any single bird.
The problem is that customers make purchasing decisions based on words. And when words don't resolve to consistent meaning, the customer pays the cost of disambiguation.
How it happens
Murmuration in starlings works because each bird follows three rules: match the speed of your neighbors, stay close, and don't collide.
Linguistic murmuration in AI works similarly:
-
Match the velocity of your neighbors. When a competitor ships "agentic workflows," you ship "agentic" something within the quarter. Not because you've built the same thing, but because the absence of the word reads as the absence of the capability.
-
Stay close. Use the same terms the analyst reports use. The same terms the RFPs ask for. The same terms the customer has learned to search. Diverging from consensus vocabulary is expensive even when your product genuinely differs.
-
Don't collide. Avoid directly contradicting what the flock says a term means, even when your implementation is meaningfully different. Instead, stretch the definition. "Memory" can mean many things. Let it.
The result is a field where terminology compounds faster than understanding.
The cost
For builders, the cost is specificity. You can't explain what makes your approach different when every differentiator has already been absorbed into shared vocabulary. "We do RAG" tells you nothing about architecture. "We have memory" tells you nothing about persistence. "We're agentic" tells you nothing about what actually runs without human approval.
For buyers, the cost is trust. If every product sounds the same, the decision defaults to brand, price, or whatever demo went last. The actual differences — the ones that determine whether the product solves the specific problem — get buried under identical language.
For the industry, the cost is precision. The more a term is stretched, the less it communicates. Eventually the word is just a flag that means "we participate in this market." It no longer carries technical or product information.
What to do about it
I don't have a policy prescription. Language is a commons and nobody governs it.
But I do think builders owe their customers clarity, even when clarity is harder to write than consensus vocabulary.
At Achiral, we try to say what we mean:
- When we say shared memory, we mean a persistent, organization-scoped knowledge layer that accumulates from your team's actual work — not a conversation log, not a prompt cache, not a context window.
- When we say compounding intelligence, we mean the system gets more useful over time because it retains and connects what your team does — not that the model gets retrained on your data.
- When we say private, we mean your data lives in your own isolated tenant and never leaves it — not that we pinky-promise to anonymize it.
Every company should be able to do the same: state the specific claim behind the shared word.
If the specifics sound different, that's not a problem. That's the point.
This post is a companion to Why Achiral?, where the name itself gets the same treatment.