The Machine Is Writing the Rubric

The most consequential risk from generative AI may not be that it becomes alien. It may be that it becomes extremely familiar: competent, fluent, statistically well-mannered, and deployed so widely that its preferences start passing for everyone else’s. That is less a robot uprising than a formatting standard with delusions of grandeur.

A Nature survey of the evidence finds a recurring pattern: LLMs can improve the judged novelty of an individual output while making the overall set of outputs more alike. In creative-task experiments, model answers were more similar to one another than human answers; in brainstorming, ChatGPT users produced ideas judged more creative but less diverse. Research cited in the feature also found reduced variation in writing style after ChatGPT’s release across scientific articles, local news, arXiv preprints and Reddit posts. The product may be a better first draft. The fleet effect is a narrower search space.

That distinction matters to engineering leaders. A model that makes each employee 15% more articulate can still make the organisation collectively worse at noticing what it has not thought of. Autocomplete is an anchor, not merely a convenience; it supplies a plausible path before the user has finished locating their own. One experiment cited by Nature found that participants who had access to ChatGPT for five days later gave more similar answers on tasks performed two months without it. The researchers called that a “creative scar.” The concern is not that every prompt becomes identical tomorrow. It is that a layer sold as assistance quietly becomes the house style for thought.

Against this backdrop comes Gemini 3.8 Flash, positioned by Google DeepMind for cost-effective, production-ready agents: multimodal inputs, a context window of up to one million tokens, output up to 64K tokens, and adjustable effort levels to trade latency and cost against quality. This is real deployment machinery, not a mood board with a GPU allocation.

But a model card is not a certificate of innocence; it is the vendor’s carefully bounded description of where it believes innocence has been tested. Google lists hallucinations, continued jailbreak-resistance work, higher token use at higher effort levels, and the possibility that “There may also be occasional slowness or timeout issues.” For an agent, that final caveat is not decorative. A slow human assistant is late; a slow automated actor may be halfway through a workflow whose assumptions have already expired.

The safety accounting deserves the same production scrutiny as the context window. The card says “Safety performance across non-English languages regressed slightly relative to 3.7 Flash,” while its automated evaluation table reports a +5.4-percentage-point multilingual safety change where lower is better. It also says its frontier-safety conclusion is inferred from the prior 3.7 Flash assessment because 3.8 allegedly adds no meaningful relevant capability. Perhaps. Yet “similar enough to the previous build” is a perfectly sensible release rationale right until it becomes the postmortem’s opening sentence. Teams deploying agents should treat model cards as inputs to their own evals, observability and blast-radius design—not as outsourced judgment.

If models are converging on familiar language, retrieval systems offer a more industrial version of the same problem: manufacture the sources that will teach the machine what sounds authoritative. A Trellner Research test of Perplexity’s Sonar models submitted 380 software-buying categories and found 7,534 citations across the two tiers. Nearly 60% pointed to domains ranked below 100,000 in Tranco’s top-million list, and 23.4% to domains outside it. Popularity is not quality, obviously. Neither is a necktie, but both are useful signals when the alternative is a man in a balaclava saying he has a white paper.

The sharper finding is structural. Three apparently related sites—wifitalents.com, worldmetrics.org and gitnux.org—contained 215,128 generated software-buying pages according to the report, and accounted for 181 citations. The researchers infer shared control from common infrastructure and templates rather than establish it as fact. Still, the setup illustrates the incentive: search-shaped content has been promoted from SEO nuisance to potential input layer for AI recommendations. One site’s homepage title was reported as “Facts & Grounding Page,” which is the sort of phrase one writes when attempting to reassure a crawler that it has not wandered into a freshly painted tunnel.

Nor should agreement between model tiers be mistaken for corroboration. Sonar and Sonar Pro returned byte-identical citation lists in 289 of 380 categories, with 0.898 overlap in URL sets. They are largely two interfaces to the same epistemic plumbing. The report did not test whether these pages changed final recommendations, examined only Perplexity in one day’s runs, and released no independent validation. That restraint is useful. It still leaves a practical procurement rule: inspect recommendation provenance, test URL resolution, and ask whether supposedly independent answers are actually reading the same content farm in different hats.

Universities are discovering what enterprises will shortly rediscover: once fluent output becomes abundant, evaluating the final artifact alone is insufficient. A Nature report on assessment redesign cites a UK survey in which roughly 94% of undergraduates said they used generative AI to assist assessed work, while 12% said they directly inserted generated text. A separate US study estimated that 9% used AI on coursework despite knowing it violated rules. These are not merely cheating statistics. They describe a collapsed boundary between doing work and supervising its production.

The better educator responses sound suspiciously like sane engineering practice. Students are asked to critique model summaries, provide AI interaction logs, preserve revision histories, defend code line by line orally, or show analysis screenshots alongside a narrative. One chemistry-data course assigns most points to working Python code and the remainder to a one-to-one explanation of how the specific code works. In other words: ship it, then explain the diff. Civilization may yet survive via the code review.

This is the common thread across model deployment, retrieval and education. Traceability is becoming the unit of trust. For agent systems, retain tool calls, sources, intermediate state and human approvals. For vendor evaluation, test the actual retrieval stack rather than admiring benchmark silhouettes. For internal writing and code, distinguish permitted assistance from unaccountable generation, then instrument the boundary. The coming premium will not be on prose that merely sounds intelligent—there is now an industrial surplus of that. It will be on people and systems that can show their work, identify its sources, and remain capable of disagreeing with the machine.