AI Adoption

The AI Failure Mode Nobody Is Watching For

August 17, 2026

A business ships a campaign it is proud of. The concept came together fast, the team used AI through most of the production, and the result looked sharp: a clean visual metaphor and a confident headline structure. Ten days later, someone on the team sees the same visual metaphor and a near-identical headline structure in a stranger’s ad. Different industry, different country, no plausible connection.

Nobody copied anyone. Both teams asked capable models for a sharp campaign, and both received the same sharpness, because they drew it from the same well. This is default tool behavior. It has a name, it has research behind it, and almost no one has put anyone in charge of watching for it.

Every leadership team using AI knows about hallucination. The confident wrong number in a proposal, the invented case law, the feature the chatbot promised that does not exist. Hallucination is loud. It embarrasses someone specific in front of someone specific, it produces an incident and a Slack thread, and it gets a review process. Fear of the loud failure is why most AI governance conversations start and end with accuracy.

Convergence is the other failure mode, and it is structured to escape exactly those defenses. Each individual output is fine. Grammatical, on-message, factually clean, defensible in any review. No error to catch means no incident and no review triggered. The failure only exists in aggregate: output by output, the company’s voice and arguments drift toward the middle of the distribution the models learned from, which is the same middle every other company’s tools draw on. One day the founder reads the last three months of the company’s content and cannot find the company in it.

Hallucination embarrasses you once. Convergence erodes the reason customers chose you, and it does so without ever producing a moment you could point to.

The exposure lands hardest on lean teams, which is a bitter piece of design. A twelve-person marketing function inside a growing company gets the most real productivity from AI, so it adopts fastest and reviews lightest. There is no brand police, no layers of approval, often no one whose job description includes reading the output in aggregate. The very structure that lets a small team move like a big one leaves nobody watching the one failure mode that compounds silently.

Convergence is the default, and the research says so

The mechanism sits in how the systems work, and the effect has been measured.

A language model is, at its core, an averaging machine. It learned from enormous amounts of text and images, and when you ask it for “a confident B2B campaign,” it produces the statistical center of every confident B2B campaign in that training data. Steer it hard, with rich context about who you are, and it can produce something at the edge of the distribution. Leave it unsteered, and it regresses to the middle, because the middle is what it is built to find. Your competitors’ tools regress to the same middle. Convergence is the machine working exactly as designed, for everyone at once.

The measured version comes from Doshi and Hauser’s experiment published in Science Advances. Writers produced short stories, some with access to AI-generated ideas, some without. Access to AI ideas made individual stories measurably better: rated more creative, better written, more enjoyable, with the biggest lift going to the less creative writers. And the AI-assisted stories were significantly more similar to one another than the stories humans wrote alone. Individual quality up, collective diversity down, in the same experiment.

Translate that to a market. Every company using these tools gets a lift, and the lift comes packaged with a pull toward sameness. Your content gets better and less yours in the same motion. The researchers called it a social dilemma: each participant benefits individually while the collective pool of novelty drains. In a market, the drained pool is differentiation, and the companies drawing from it fastest are draining it fastest.

Distinctiveness is the asset a growing business can least afford to lose

An enterprise can survive sounding generic. When the brand blurs, it buys its way back into memory: sponsorships, retargeting, category ads, a sales floor large enough to brute-force pipeline. Attention is purchasable at that scale, and plenty of forgettable enterprises remain profitable.

A $20 million business has no such cushion. It is distinguishable or it is invisible. Its entire go-to-market runs on a prospect remembering, three weeks after a first touch, that this firm sounded different from the other six tabs open that day. For a growing company, the distinct point of view carries most of the moat.

Set the two curves next to each other. AI adoption is least governed exactly where the productivity gains are largest, in lean teams at growing companies. And the cost of convergence is highest exactly where distinctiveness carries the business, in those same companies. The businesses with the most to lose from the silent failure mode are the ones running fastest toward it, with the fewest people watching.

One more use for this lens before the defenses. Voice doubles as a diagnostic. When a company’s outputs cannot hold a consistent voice even with good tools and honest effort, the cause usually sits upstream of any prompt: sales is promising one thing, product believes another, and the founder’s positioning lives in the founder’s head. A model handed contradictory context produces mush, at volume. If your AI output sounds like nobody, ask first whether your functions ever agreed on who somebody is. The tools did not create that misalignment. They put it in print.

The defenses you can build now

Convergence is structural, so the defenses are structural. None of them is a better prompt. All of them are buildable inside a quarter without an enterprise budget.

Codify who you are, in an artifact. A brand context document that states your actual point of view in concrete terms: what you believe that competitors do not, and the arguments you make versus the ones you refuse to touch. Not “confident but approachable.” Positions. This document is the steering input that pulls a model off the statistical middle, and writing it will surface the upstream disagreements while they are still cheap to settle.

Curate an example set. A model learns your voice from your proudest shipped work far better than from any description of it. We wrote up the full discipline of building one separately: roughly ten pieces, chosen by whether you would send them to a prospect unprompted, with the good-but-generic work deliberately excluded, because the model averages everything you show it.

Review shipped output by sampling, in aggregate. Convergence is invisible piece by piece, so the review cannot run piece by piece. Once a month, pull a sample of what actually shipped across functions, put it next to your example set and one competitor’s output, and ask two questions. Does this sound like us? Could this have come from anyone? Twenty minutes, calendar-anchored, and it catches in one sitting what no per-item approval flow can catch at all.

Name an owner. Someone with the standing to look at defensible, error-free work and say “this has drifted” and route it back. Without a name attached, drift-watching is everyone’s job in principle and nobody’s on Thursday. This role is the unglamorous line item missing from most AI business cases, and it is the entire difference between having guardrails and having had them.

The tools are not going to develop taste on your behalf, and the market is not going to slow down while you decide whether this matters. Model defaults improve every quarter, which means the statistical middle gets more polished every quarter, which makes generic output more comfortable to accept every quarter. The companies treating distinctiveness as infrastructure, with an artifact, an example set, a sampling cadence, and a name, will be the only ones left sounding like themselves. Everyone else will sound very good, and exactly alike.

Sources

Doshi, A. R., and Hauser, O. P., “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances 10(28), 2024. https://www.science.org/doi/10.1126/sciadv.adn5290