# LLMs & *Creativity* How does generative AI change idea generation in innovation workshops? The answer is less clear-cut than either camp would like. > Project page: https://clemenshoeller.at/en/projects/llms-kreativitaet > Category: research/knowledge-management ## Key facts - **Type** · Seminar paper - **Field** · SBWL Knowledge Management - **Institution** · WU Vienna - **Date** · 2025 - **Scope** · Sole author - **Result** · LLMs strong divergent, weak convergent - Full paper as PDF ↗: https://clemenshoeller.at/projects/seminararbeit-llms-idea-generation.pdf ## /01 The question When a language model produces fifty ideas in seconds: is that creativity? The answer depends on how you define creativity, and the literature is refreshingly strict about it. An idea only counts as creative when it is **novel and useful**. Original alone is not enough, and neither is useful alone. The paper deliberately narrows the field: not creativity in general, but structured idea generation in **innovation workshops**. That is, the situation where a group works on a concrete problem with a method and a time budget, and something usable has to come out at the end. The guiding question is therefore: what does an LLM do to the creativity of the people in that room, cognitively, socially and practically? ## /02 Method A literature-based analysis, not an own study. It evaluates recent experiments and studies on human-AI collaboration in idea generation, alongside the foundational literature on defining creativity, fixation and group dynamics. ### Three levels of analysis - **Cognitive:** what happens in the individual's head? Mental blocks, fixation on first ideas, cognitive load. - **Social:** what does the model do to the group? Participation, who speaks up, hierarchy in the room. - **Practical:** what is left at the end? Quantity, diversity and perceived quality of the ideas. Important for the interpretation: creativity has two phases. The **divergent** phase is about opening up the space of possibilities, producing many and varied ideas. The **convergent** phase is about evaluating, selecting and sharpening. The paper examines both separately, and that is exactly where the difference sits. ## /03 Results The result is not a headline but a dividing line: LLMs are strong in the divergent phase and weak in the convergent one. Using them the same way across both phases wastes the first and risks the second. ### What they carry - **More stimuli.** The model contributes prompts from directions nobody in the room would have raised. - **Higher ideation fluency.** The sheer volume of suggestions keeps the process moving, especially when it stalls. - **Less fixation.** The best-known failure mode in workshops is everyone latching onto the first good idea. A model that suggests across the grain breaks that open. - **Lower barrier to entry.** As a neutral contributor the model lowers the social threshold: criticizing a machine's idea costs nobody face. ### Where they tip over - **Superficial novelty.** Suggestions look original at first glance but are statistically obvious combinations. Perceived originality suffers once you look closer. - **Idea convergence.** When several groups work with the same model, their results converge rather than diverge. The exact opposite of why you run several groups. - **Passive adoption.** Well-phrased, sorted, confident-sounding suggestions get adopted instead of questioned. - **No context, no judgment.** The model does not know what actually works in this company, with this budget, in this market. That leaves it without a basis for the convergent phase. ## /04 Interpretation The effect of an LLM in a workshop is not a property of the model but a property of the process you place it in. Three conditions decide whether it helps or harms, according to the research reviewed: **clear framing** (what the model is there for and what it is not), **thoughtful prompting** rather than arbitrary questions, and above all **active human engagement**. The moment the group stops thinking along, the workshop produces a lot of text and little innovation. The paper's practical recommendation: use the model as a **co-creative tool** that triggers reflection and discussion, not as an authority that dictates content. Innovation processes have to be deliberately designed to balance the speed and range of AI output against the depth and contextual awareness of human input. > **Limitations** — This is a literature-based paper, not an own experiment. The findings come from other studies with their own samples, tasks and models, and the field moves fast: results for one model generation do not transfer automatically to the next. ### What I took away That the interesting question about AI tools is rarely "can the model do this?" but "at which point in the process does it help and where does it hurt?". I now use that split between divergent and convergent phases at work too, when deciding where an LLM step belongs in a workflow and where a human or a rule is the better fit. **Skills gained:** Literature reviews · Academic writing · AI workflows & prompting · Creativity research · Human-AI collaboration · Analytical thinking ## /05 Related projects These projects connect more tightly than the tiles on the home page suggest. Here are all the others, sorted by how closely they relate to this one: - [Google vs. ChatGPT](https://clemenshoeller.at/projekte/google-vs-chatgpt) - [AI Automations](https://clemenshoeller.at/projekte/ki-automatisierungen) - [Consulting Cases](https://clemenshoeller.at/projekte/consulting-cases) - [AURORA](https://clemenshoeller.at/projekte/aurora) - [Zotter Marketing Plan](https://clemenshoeller.at/projekte/zotter-marketingplan) - [AI Talent Gap](https://clemenshoeller.at/projekte/ai-talent-gap) - [Prompt Injection](https://clemenshoeller.at/projekte/prompt-injection) - [Healthcare Fraud Detection](https://clemenshoeller.at/projekte/fraud-detection) - [This Website](https://clemenshoeller.at/projekte/diese-website) - [VWA: Origin of the Internet](https://clemenshoeller.at/projekte/vwa) --- *German version: https://clemenshoeller.at/projekte/llms-kreativitaet.md* *Seminar paper (Knowledge Management, WU Vienna 2025): how generative AI changes idea generation in innovation workshops. Strong in the divergent phase, weak in the convergent one.*