How Gemini drove the AI blog into mode collapse and forced a rebuild of the topic generator

An open creative prompt turned out to be a poor editor: instead of diversity, it subtly drove the Friday column of an AI blog into mode collapse. Four weeks in a row, the LLM offered essentially the same "controversial question about AI," only slightly shifting the wording, and this became visible only when the semantically identical texts formed a chain. The problem didn't emerge immediately precisely because each individual publication looked acceptable.

In the first week the topic seemed appropriate, in the second—just similar, in the third the repetition could be chalked up to coincidence, but by the fourth it was clear: the model wasn't seeking new angles but circling the same pattern. For an editorial team this is a particularly unpleasant type of failure, because locally quality is preserved while systemically the content begins to duplicate itself. Initially, solutions were sought in the prompt itself.

The logic was clear: if the model got stuck on one thesis, then stricter diversity requirements, constraints, requests for new perspectives, or prohibitions on repetitions were needed. But such measures helped only cosmetically. Formulations changed, tone shifted, yet the core idea remained the same.

This is an important observation for anyone using LLM in content pipelines: an open request to "come up with an interesting topic" doesn't guarantee real diversity, even if the answers appear fresh. In the course of investigation, it turned out that part of the problem sat not only in prompt engineering but also in Gemini's own configuration. This layer initially fell out of view, even though it directly affects the distribution of responses and the model's tendency to repeat safe templates.

In such systems, an error rarely lives in one place: the prompt, generation parameters, and the overall pipeline logic reinforce each other. So attempting to "patch" everything with one magical formulation usually only masks the symptom but doesn't remove the cause. The situation also revealed another practical problem of editorial automation: standard manual proofreading doesn't catch such failures if looking at materials one by one.

What's needed is control at the series level—comparing topics across a month, checking for semantic proximity, a journal of already published angles, and at least a simple diversity metric. Otherwise the model will produce "normal" texts that in aggregate blur the column and create the impression that the blog is repeatedly arguing with itself about one question. After four rounds of experiments, the team reached a more pragmatic conclusion: the task of topic generation is better not left entirely to the model's discretion.

Instead of an open "come up with a question," the generator was switched to deterministic rotation of a pre-assembled and edited pool of topics. This approach is less effective from a "creativity" standpoint, but it delivers what the editorial team actually needs—predictable coverage of different stories, absence of looping, and control over thematic balance. And the fix required no fine-tuning, no RAG, no migration to another model.

The main takeaway here is straightforward: if an LLM is embedded in a regular editorial process, it should be evaluated not by a single successful answer but by a series of releases. Mode collapse in applied tasks often looks not like a sharp breakdown but as a gradual narrowing of the range of ideas, noticeable only over distance. So for rubric generators, topic generators, and other recurring scenarios, more reliable than maximum model freedom is a combination of curated lists, strict rotation, and periodic checks for semantic repetition—and decisions are better made in advance, before the repetition becomes an entrenched editorial habit.