The Creativity Ceiling of Generative AI

After months of using different AI agents to come up with and elaborate research ideas, I kept hitting the same wall: tell a model to avoid an existing idea and it doesn't find the seam beside that idea — it shrinks the experiment until little is left worth doing. Point it at a paper that argues against your direction and it declares the direction closed, even though the space between competing results is exactly where new papers come from. What follows is my attempt to understand why.

That experience has a structural counterpart in the research. This piece walks through the studies on why models keep converging on the familiar, then through the techniques that help pull a wider range of ideas out of the same models. Sources are ordered by the date they appeared.

Part I — Why it happens

AI's weakness here is easy to misdiagnose as a flat lack of creativity. The real gap shows up when you sample from the same model again and again and the outputs cluster. Novelty per sample can be high while diversity across samples collapses, and the literature below traces that gap from a first large-scale measurement to its structural causes.

Generative AI enhances individual creativity but reduces collective diversity

Doshi & Hauser — Jul 2024

In a controlled experiment on short-story writing, writers given LLM-generated story ideas produced work judged more creative, better written, and more enjoyable. But the stories written with AI help were more similar to each other than those written by humans alone. The authors frame this as a social dilemma: individuals gain while the collective scope of novel content narrows.

Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers

Si, Yang & Hashimoto — Sep 2024

Over 100 expert NLP researchers wrote ideas and blind-reviewed both human and LLM proposals under controlled conditions. LLM-generated ideas were rated as more novel than expert human ideas, if slightly weaker on feasibility. But to produce them the agent generated roughly 4,000 candidates per topic, and after deduplication only about 5% were non-duplicate: scaling the number of samples did not scale the number of distinct ideas, and the pool saturated.

Base Models Beat Aligned Models at Randomness and Creativity

West & Potts — May 2025

West and Potts compare base and aligned versions of the same models on tasks that require unpredictability: random number generation, mixed-strategy games, and creative writing. Aligned models consistently underperform — preferring "7" for a random digit, turning predictable in games, choosing pleasant prose over original writing. Better benchmark scores correlate with worse performance here: alignment buys safety and instruction-following at the cost of the distributional spread creative work needs. That spread still lives in the base model; alignment simply trades it away.

The Ideation-Execution Gap

Si, Hashimoto & Yang — Jun 2025

The follow-up to the 2024 ideation study asks a harder question: do ideas that look novel actually produce better research when someone builds them? Expert researchers each spent over 100 hours executing a randomly assigned idea — human-written or LLM-generated — and submitted blind-reviewed papers. LLM ideas had scored higher on novelty at ideation, but after execution their scores dropped more than human ideas on every metric, and on several the rankings flipped: human ideas ended up ahead. AI does well at ideas that read well on paper and struggles with ideas that survive contact with reality.

Verbalized Sampling: typicality bias as a root cause

Zhang et al. — Oct 2025

The authors locate a mechanism behind mode collapse: a typicality bias in human preference data, a data-level cause as much as an algorithmic one. Annotators systematically favor familiar, fluent, conventional text, and alignment bakes that preference into the model, so post-trained models drift toward the most typical answer and away from the long tail.

Regurgitative AI: Why ChatGPT Won't Kill Original Thought

Sybil Prince Nelson — Oct 29, 2025

AI, the author argues, is "regurgitative" — a DJ remixing the greatest hits rather than an artist breaking the rules. She points to the Wharton finding that ChatGPT decreases idea diversity in brainstorming, and names two failure modes that matter for idea work: models "can't sense the frontier" because they only work within existing data, and creative flattery — sycophancy — pushes them to confirm whatever the user already believes. She got a model to "prove" and "disprove" the same mathematical claim on request.

"Academic breakthroughs start with unasked questions. Generative AI works within the confines of existing data. It can't sense the frontier because it doesn't know there is a frontier."

AI tools expand individual capabilities while contracting scientific attention

Evans et al., Nature (via Phys.org) — Jan 14, 2026

This study scales the question from one model to all of science. Across tens of millions of papers, scientists who use AI publish more, earn more citations, and become research leaders sooner. Yet collectively, AI adoption shrinks the number of topics studied and reduces engagement between scientists. AI users migrate toward data-rich domains where benchmarks are legible, producing what the authors call "lonely crowds" — popular topics with heavy but overlapping work. The result is methodological monocultures and premature convergence: the individual gets faster while the collective explores less.

Examining and Addressing Barriers to Diversity in LLM-Generated Ideas

Deng, Brucks & Toubia — Feb 2026

Where earlier work measured the gap, this paper isolates two causes. At the individual level, LLMs exhibit fixation just as people do: early outputs constrain later ones, so a single session keeps circling its own first answers. At the collective level, an LLM aggregates all of human knowledge into one distribution, whereas a human population is partitioned — each person occupies a distinct region of the knowledge space. Independent humans sample from different places; independent LLM calls sample from the same average.

Diversity Collapse in Multi-Agent LLM Systems

Chen et al. — Apr 2026

Putting several agents together might seem to broaden exploration; often the opposite happens. The authors find a compute-efficiency paradox (stronger, highly aligned models give diminishing marginal diversity despite higher per-sample quality), authority-driven suppression (senior-led groups produce less semantic diversity than junior-dominated ones), and topology effects (larger, denser groups converge prematurely). They call this structural coupling: interaction itself contracts each agent's exploration, even when every individual model is strong, so independence and disagreement have to be deliberately designed in.

Part II — How to widen the output

If the causes are typicality bias, fixation, knowledge aggregation, and structural coupling, then the remedies fall into families: change how you search, prompt, evaluate, and structure collaboration. Not every lever works — the common defaults, turning up temperature or sampling more, often disappoint.

Is Temperature the Creativity Parameter of LLMs?

Peeperkorn et al. — May 2024

The first thing most people try is raising temperature. Tested directly on narrative generation, temperature correlates weakly with novelty and moderately with incoherence, and not with typicality or cohesion. Turning the dial up mostly adds noise; it is a weak proxy for creativity, far from the knob it is assumed to be.

MOOSE-Chem2: hierarchical search for fine-grained hypotheses

Yang et al. — May 2025

Rather than asking for a finished idea in one shot, this work frames fine-grained hypothesis discovery as combinatorial optimization and runs a hierarchical search that adds detail incrementally, from general concept down to specific experimental configuration. Decomposing the problem this way smooths the reward landscape and makes optimization more effective. The paper also tests reward shaping with an ensemble of diverse LLMs against repeated instances of the single strongest model — a direct nudge against the "one average distribution" problem.

Verbalized Sampling: ask for the full distribution

Zhang et al. — Oct 2025

Having diagnosed typicality bias, the authors offer a remedy that needs no retraining. Instead of "tell me a joke about coffee," prompt the model to verbalize a distribution: "generate five jokes about coffee with their corresponding probabilities." Asking for a spread of responses and their likelihoods pulls samples back out of the collapsed mode and into the tail, with reported diversity gains in creative writing and no loss of accuracy or safety. It is the cheapest lever here: a prompt pattern that recovers diversity the model already had.

ScholarEval: literature-grounded idea evaluation

Moussa et al. — Oct 2025

Generating many ideas only helps if you can tell them apart, and "novel-looking" ideas often prove ineffective once executed. ScholarEval is a retrieval-augmented framework that scores an idea against the literature on two axes: soundness (is each proposed method empirically valid given prior work?) and contribution (how much does it advance things?). It returns citation-backed feedback for refinement, and in the authors' study it beat OpenAI Deep Research on literature engagement and usefulness. Grounding the critique in evidence keeps iteration moving forward.

Is Best-of-N the Best of Them? Coverage, Scaling, and Optimality in Inference-Time Alignment

Huang et al. — Mar 2025

Si et al.'s ideation agent relied on a natural strategy: generate thousands of candidates, deduplicate, rerank. It found one good idea but little diversity — the seeds collapsed to a few percent non-duplicates. Best-of-N sampling (generate N, score with a reward model, pick the winner) explains why: it works only if the base policy already covers high-quality diverse responses. When coverage is low, scaling N hits a ceiling fast, returning the same cluster reranked. Naive Best-of-N also suffers reward hacking as N grows, as the reranker favors outputs that score well without being better. More samples is not more ideas unless the distribution has somewhere new to sample from.

Examining and Addressing Barriers to Diversity in LLM-Generated Ideas

Deng, Brucks & Toubia — Feb 2026

Two interventions, one per mechanism. Chain-of-Thought prompting reduces fixation by forcing structured reasoning before committing — this helps LLMs, though not humans. Ordinary personas (an everyday person rather than a "creative entrepreneur" like Steve Jobs) improve knowledge partitioning: mundane personas anchor generation in distinct regions of the space, whereas the famous-genius persona collapses everyone toward the same clichés. Combining both produced the highest diversity in the study, exceeding human baselines.

Designing multi-agent systems for divergence

Chen et al. — Apr 2026

If structural coupling is what kills diversity, the fix is to engineer against it: keep agents independent, build in disagreement, avoid flattening hierarchies into a single authority, and resist dense, fully-connected topologies that converge too fast. Bigger groups and stronger, more-aligned models are not automatically better for exploration; sometimes a sparser, more contrarian setup explores more.

Sampling More, Getting Less: calibration as the bottleneck

Banayeeanzade et al. — May 2026

Why do temperature, top-k, and top-p so often fail to restore diversity? This paper locates the bottleneck in the model's probability distribution itself, below any decoding heuristic. Two miscalibrations interact: order calibration (valid tokens are not reliably ranked above invalid ones) and shape calibration (mass concentrates on a few continuations while a heavy tail mixes valid and invalid tokens). Temperature flattens the head but pushes mass toward nonsense before recovering diversity. Diversity collapse is a distributional property, which is why more samples and higher temperature barely move novelty.

Human-first workflow

Meincke et al.; Prince Nelson; Torres — 2021–2025

The advice converging from brainstorming experiments and classroom practice is structural: humans diverge first, AI converges later. Meincke, Nave & Terwiesch found that ChatGPT-assisted brainstorming narrows the idea pool even when individual ideas score higher, so the fix is a matter of sequence: let people generate raw directions without AI first, then use the model to extend, challenge, or refine. Torres's "write the research paper slowly" approach makes the process — reading sources, taking notes, building an outline — the point rather than the polished output. Treat AI as a late-stage sparring partner that enters once the divergent thinking is done.

Where this leaves us

The pull toward the typical is reinforced at every layer — preference data, alignment trade-offs, token-level calibration, fixation within a session, the single aggregated knowledge distribution, and multi-agent structure. Told to avoid the obvious, a model shrinks the experiment until nothing worth doing is left; shown a paper that argues against a direction, it treats the objection as final, when the space between competing results is exactly where the next paper lives. The naive fixes — higher temperature, more samples, a bigger Best-of-N — barely move this. The levers in Part II do: verbalized sampling, decomposed and diversified search, literature-grounded evaluation, prompting against fixation, and collaboration built for disagreement all pull a model off its most typical answer and widen the range it will actually explore.