AI Milk [2026]
Dr. Ian Malcolm said it in 1993 and nobody in enterprise technology has improved on it since: “Your scientists were so preoccupied with whether they could, they didn’t stop to think if they should.”
Jurassic Park’s failure mode was not that the technology didn’t work. The dinosaurs were magnificent. The park was a genuine technical achievement. The problem was that nobody in the room had seriously entertained the question of what happens when it goes wrong at scale — because the fact that it could be done was treated as sufficient justification for doing it. Capability was mistaken for authorization. The board meeting that should have asked the hard questions never happened.
The enterprise AI deployment story of 2025–2026 is Jurassic Park in a conference room. The tool is impressive enough that questioning whether it is working feels like questioning whether electricity is a good idea. And so the board meeting doesn’t happen here either.
There is a man in a workshop right now explaining that his goal is to have three AI agent employees running by the end of the year. Someone asked him who would review the output — especially if it did not align with his domain expertise.
The room went quiet.
That silence is the InGen board meeting that never happened. And the thing being unleashed is not a dinosaur. It is something subtler and harder to see coming: the slow homogenization of how organizations think.
What AI Milk Is
AI Milk is what you get when enterprise AI tools process human communication at scale without human judgment applied to the output. It is the email that sounds professional but carries no one’s voice. The meeting summary that captures every agenda item and misses the decision that actually mattered. The Slack message that is grammatically impeccable and operationally empty. The strategic document that is indistinguishable from the one the competing team generated with the same prompt, the same tool, and the same absence of friction.
The term comes from the process. Pasteurization makes milk safe and shelf-stable by removing the organisms — good and bad — that give it character. AI writing tools do the same thing to human communication: they remove the friction, the idiosyncrasy, the rough edge that carries meaning. What remains is safe, smooth, and uniform. Technically correct. Informationally thin.
This is not a complaint about AI tools. It is a complaint about how organizations are deploying them — as a substitute for thinking rather than a tool for doing more of it.
The Research Says It Out Loud
This is not a vibe. It is documented.
A controlled study found that when multiple users co-wrote argumentative essays with an instruction-tuned LLM (InstructGPT), the result was significantly reduced lexical and conceptual diversity — and different authors’ essays became measurably more alike. The model produced average phrasing and smoothed out each author’s unique stylistic markers. USC researchers analyzed more than 130 studies and arrived at a conclusion worth reading slowly: “The concern is not just that LLMs shape how people write or speak, but that they subtly redefine what counts as credible speech, correct perspective, or even good reasoning.”
That last part is the one to sit with. It is not that everyone writes the same way. It is that everyone starts to think the same things sound reasonable. The tool trained on the statistical average produces the statistical average and then grades everything else against it.
In a separate study on AI writing assistance and cultural expression, Indian participants who used AI writing tools produced text that was measurably more similar to American writing — not just in style, but in content choices and cultural framing. Effect size: Cohen’s d = 0.91. Large. The AI was not asked to homogenize. It just did, because the training data did, because that is what optimization toward a mean produces.
The student essay research put it in terms that apply directly to the enterprise: “Just as monoculture in agriculture maximizes short-term yield while introducing long-term fragility, the homogenization of thought may undermine the very preconditions for knowledge creation, critical inquiry, and innovation.” Quality improves. Diversity collapses. Both can be true simultaneously, and the conventional evaluation framework only measures the first one.
The Enterprise Version
In the corporate setting, AI Milk looks like this:
Meeting summaries that capture the agenda and miss the argument. The dissent that was raised, not resolved. The assumption that everyone agreed to but nobody said out loud. The thing the most senior person in the room believed that shaped the direction without being stated as a decision. Gemini summarized the meeting. The summary is accurate. The meeting is not in it.
Emails that are professionally indistinguishable. PR Daily surveyed 300 communications professionals at the start of 2026 and found broad prediction that “messaging will sound identical” as AI writing is normalized. One professional noted a breaking point with “trite social posts and emoji-heavy AI writing.” Another: “We’ll start to prefer imperfection in comms because it will feel more human.” When your external audience can no longer tell your brand’s communication from your competitor’s, you have not gained efficiency. You have lost differentiation.
Strategic documents that do not move the business. The deliverable is produced. The box is checked. The executive team gets a summary with headers and bullet points. The insight that would have changed the decision did not survive the smoothing process.
Three AI agent employees with nobody qualified to review them. High-velocity output. No domain expertise applied to evaluation. The agents are productive in the same way a machine running without gauges is productive — until it is not, at which point the failure is large and expensive and difficult to trace because nobody was watching the instruments.
The Adoption Mandate That Produces It
The AI Milk problem has a supplier: the organizational pressure to deploy tools without an evaluative framework for whether the output is actually better.
When enterprise AI adoption is framed as “get on board or be left behind,” the implicit argument is that the tool is so obviously beneficial that resistance constitutes willful ignorance. This is not analysis. This is missionary posture, and it produces the worst kind of adoption: compliance without judgment. People use the tool because they were told to. They send the AI-generated email because it is faster. They circulate the Gemini meeting summary because it exists. Nobody asks whether the output serves the purpose it was supposed to serve, because the organization is measuring adoption, not outcomes.
The evals data from this board’s Emerging Careers research is the enterprise version of the same failure: 89% of AI teams have observability (they can see what the system is doing) but only 52% have evals (they have a way to know whether the output is correct). Companies deployed AI and then measured whether it was running, not whether it was working. AI Milk is what you produce when you measure running and ignore working.
The Bill That Has Not Arrived Yet
There is a secondary problem forming behind the homogenization problem: the enterprise token bill.
When AI tools are deployed at scale — every email, every meeting summary, every document, every agent running in the background — the token costs accumulate at a rate that the initial business case did not price. The ROI analysis for enterprise AI adoption was typically performed at the proof-of-concept scale, where the tool is impressive and the cost is marginal. At production scale, where every knowledge worker is generating AI-assisted output throughout the workday, the compute cost looks different. That reckoning has not arrived at most organizations yet. When it does, the question “what did we get for this?” will require an answer that the output quality measurement never captured, because nobody built the instrument to measure it.
What Actually Wins
Deep expertise wins. Not because AI tools are useless — they are not — but because the value of the tool scales with the quality of the judgment applied to it. The domain expert who uses AI to do more of what they are already excellent at produces something different from the non-expert who uses AI to do something they could not otherwise do. The first person is amplified. The second person is replaced by a plausible-sounding average.
The person who asks “who reviews the output?” and means it — who actually has the expertise to evaluate whether the agents are producing something correct and useful — is not behind the adoption curve. They are running a better quality control system than most organizations that are busy counting their AI seat utilization.
The differentiated voice, the unexpected framing, the specific insight that only comes from a career of doing a thing — none of that is in the statistical average the model was trained to produce. It is in the human who uses the model as a tool rather than a replacement. And as every team around them produces AI Milk, the person with something distinctive to say becomes increasingly easy to find in the noise.
I have been in enough process improvement programs to recognize the pattern: the organization adopts the tool, measures the adoption, and calls that success. The harder question — what changed in the output that matters — goes unmeasured because it is harder to count. AI adoption at the enterprise is doing the same thing faster than most previous technology cycles, because the tool is impressive enough that questioning whether it is working feels like questioning whether electricity is a good idea.
The man with three AI agent employees is not wrong to want leverage. He is wrong to want it without the domain expertise to evaluate what the leverage produces. An agent that generates 1,000 wrong decisions per day is not a productivity tool. It is a liability that runs faster than you can catch it.
The researchers are right that monoculture maximizes short-term yield while introducing long-term fragility. That is also the definition of every enterprise AI deployment story that ends in a case study about what went wrong. The short-term yield is the slide deck number: adoption rate, seats activated, hours saved. The long-term fragility is the organization that cannot remember how to think distinctively because it outsourced the thinking and then forgot to keep the expertise that was supposed to check the output.
Enterprise AI is producing AI Milk. It is technically safe, nutritionally homogenized, and shelf-stable. It will not spoil. It will also not nourish anyone in a way they could not have been nourished by the generic version from the competitor who ran the same prompt. Deep expertise, specific insight, and the willingness to ask who reviews the output — that is the raw milk. Harder to produce. Harder to scale. Considerably more valuable to the person who can tell the difference.
Sources
Zhivar Sourati, Alireza S. Ziabari, and Morteza Dehghani, "The homogenizing effect of large language models on human expression and thought," Trends in Cognitive Sciences, 2026 (USC analysis of 130+ studies; credible speech redefinition) · personaMail arxiv.org/pdf/2602.17340 (InstructGPT co-writing study: reduced lexical and conceptual diversity; authors’ essays become more alike) · "Does AI Homogenize Student Thinking? A Multi-Dimensional Analysis," arxiv.org/pdf/2603.21228 (quality improves / diversity collapses simultaneously; monoculture-fragility analogy) · Angie D., "The AI Homogenization Problem," Medium/ILLUMINATION, June 2026 (cultural expression study: Indian participants’ writing shifted to American framing, Cohen’s d = 0.91) · PR Daily, "Your predictions: How AI in comms will evolve in 2026," January 2026 (300 communications professionals surveyed; messaging will sound identical; trite social posts breaking point; preference for imperfection) · Knowledge Hub Media, "How to Avoid AI Content Homogenization," July 2026 (sameness carries commercial cost; consumers engage less with AI-perceived content; trust reduction) · TheMoneyZoo Emerging Careers Board, AI Evaluation Engineer Blueprint (89% observability / 52% evals adoption; $1.9B/yr undetected losses; the enterprise measurement gap).
