How Gemini Pushed an AI Blog into Mode Collapse and Forced a Rebuild of Its Topic Generator

An open-ended creative prompt turned out to be a poor editor: instead of delivering variety, it quietly drove the AI blog's Friday column into mode collapse. For four weeks in a row, the LLM proposed essentially the same "contested question about AI," changing only the wording slightly, and this became visible only once the texts, identical in meaning, formed a chain.

The problem did not show up right away precisely because each individual post looked acceptable. In the first week the topic seemed fitting, in the second merely similar, and in the third the repeat could still be written off as coincidence. By the fourth, it was clear that the model was not looking for new angles but circling the same pattern. For an editorial team this is an especially unpleasant kind of failure: quality holds locally, while systemically the content starts duplicating itself.

At first the team looked for a fix in the prompt itself. The logic was clear: if the model is stuck on one thesis, the diversity requirements need to be spelled out more strongly, with added constraints, requests for new angles or a ban on repeats. But these measures helped only cosmetically. The wording changed, and so did the tone, but the core of the idea stayed the same. This is an important observation for anyone using LLMs in content pipelines: an open request to "come up with an interesting topic" does not guarantee real variety, even if the answers look fresh.

During the debugging, it turned out that part of the problem lay not only in prompt engineering but also in Gemini's own configuration. That layer was initially overlooked, 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, the generation parameters and the overall pipeline logic reinforce one another. That is why trying to "finish the cure" with a single magic phrasing usually just masks the symptom without removing the cause.

The situation also exposed another practical problem of editorial automation: ordinary manual proofreading does not catch such failures if you look at pieces one by one. You need control at the series level: comparing a month of topics, checking for semantic similarity, keeping a log of angles already published and at least a simple diversity metric. Otherwise the model will keep producing "normal" texts that, taken together, blur the column and create the impression that the blog is arguing with itself over and over about a single question.

After four rounds of experiments, the team reached a more down-to-earth conclusion: 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 through a pre-assembled and edited pool of topics. This approach is less impressive in terms of "creativity," but it delivers what an editorial team really needs: predictable coverage of different stories, no looping and control over thematic balance. Fixing it required neither fine-tuning, nor RAG, nor migration to another model.

The main takeaway is simple: if an LLM is built into a regular editorial process, it should be judged not by one good answer but by a series of issues. In applied tasks, mode collapse often looks not like a sudden breakdown but like a slow narrowing of the range of ideas, noticeable only over distance. So for section planners, topic generators and other recurring workflows, what works more reliably is not maximum freedom for the model but a combination of a curated list, strict rotation and periodic checks for semantic repeats. Decisions are best made in advance, before the repetition turns into a settled editorial habit.