TLDR: A study found that Large Language Models (LLMs) struggle with genuine creativity, often defaulting to generic phrasing, a phenomenon termed “Galton’s Law of Mediocrity.” Through a two-phase experiment involving ad concept simplification and regeneration, researchers observed that creative elements like metaphors and emotions disappear first during compression. When LLMs tried to recreate ads from simplified inputs, they produced lexically varied but superficially novel text, failing to restore true originality. However, providing explicit creative cues (markers) significantly improved the models’ ability to align with original creative ideas, suggesting that targeted guidance can help steer LLMs away from predictable outputs in creative tasks.
Large Language Models (LLMs) have transformed how we interact with technology, generating fluent and coherent text across countless applications. However, a new research paper titled Galton’s Law of Mediocrity: Why Large Language Models Regress to the Mean and Fail at Creativity in Advertising delves into a critical limitation: their struggle with genuine creativity. Authored by Matt Keon, Aabid Karim, Bhoomika Lohana, Abdul Karim, Thai Nguyen, Tara Hamilton, and Ali Abbas, this study formalizes the LLM tendency to default to safe, generic phrasing as a ‘Galton-style regression to the mean’ in language.
The core idea, inspired by Francis Galton’s 1886 principle of regression towards the mean, suggests that just as extreme traits in a population tend to drift back to the average over generations, LLMs, by maximizing next-token likelihood, privilege high-frequency, statistically probable patterns. This leads them to suppress rare or unconventional expressions, resulting in outputs that are polished but often lack true originality. This phenomenon is particularly evident in creative domains like advertising, where uniqueness and emotional resonance are paramount.
The Two-Phase Creativity Stress Test
To investigate this, the researchers designed a unique two-phase empirical study using a database of over 1000 advertising ideas. The study aimed to answer three key questions:
- When ads are gradually simplified, do generic product details or unusual creative highlights last longer?
- If given only bare facts, can models regenerate creative flourishes or do they resort to clichés?
- Can explicit creative cues, or ‘markers,’ improve the recovery of originality during regeneration?
Phase 1: Forgetting involved prompting state-of-the-art LLMs (including GPT-4o, GPT-5 variants, Gemini-2.5-Pro, and Claude Sonnet 4) to compress original ad concepts through mild, moderate, and extreme reductions. The goal was to observe what creative elements disappeared first. For instance, an ad about car sensors with a metaphor like ‘tree people’ would see the metaphor, humor, and sensory details vanish early, leaving only the core factual message: ‘Car sensors detect what you can’t see.’
Phase 2: Expansion reversed this process. Starting from the extremely compressed versions, models were asked to expand them back into full ads. This phase had two conditions:
- Plain Expansion: Models expanded without any additional guidance.
- Marker-Driven Expansion: Models were given 3-4 ‘creative markers’ (short, ad-specific phrases like emotional cues, metaphors, or visual details) extracted from the original ad to guide their regeneration.
Evaluation involved a hybrid framework combining quantitative metrics (cosine similarity for semantic fidelity, METEOR for surface overlap, entropy for lexical diversity, and 4-gram uniqueness for structural novelty) with qualitative human and LLM judgment.
Key Findings: Creativity’s Fragility and the Power of Guidance
The results provided clear evidence for creativity regression. In Phase 1 (Forgetting), both cosine similarity and METEOR scores dropped sharply as ads were simplified, indicating a loss of original meaning and wording. Lexical diversity (entropy) also declined, while 4-gram uniqueness rose, meaning the compressed outputs used new phrasing but were semantically thinned. Qualitatively, emotional and metaphorical markers were the most fragile, declining by over 60-70% between mild and extreme forgetting, while brand-specific cues and slogans were more resilient. This confirmed that the imaginative components of ads are the first to vanish.
In Phase 2 (Expansion), models showed a consistent pattern. During Plain Expansion, they generated lexically varied and novel-sounding outputs (high entropy and 4-gram uniqueness), but these expansions recovered only a fraction of the original meaning and surface overlap (low cosine similarity and METEOR). The ‘novelty’ was often superficial, relying on a narrow pool of high-frequency advertising clichés, such as ‘whisper secrets’ or ‘endless possibilities.’
However, Marker-Driven Expansion significantly improved recovery. Cosine similarity and METEOR scores were noticeably higher, indicating better semantic alignment and recognizable overlap with the originals. Entropy also increased, suggesting richer lexical variety. This demonstrated that providing structured creative cues helped models restore more of the original creative content and diversify their stylistic output, moving closer to the mix of emotional, visual, and slogan-like content found in real advertising.
Also Read:
- Unveiling Self-Preference: How Large Language Models Develop Human-Like Bias
- Navigating the Promise and Pitfalls of AI as Synthetic Social Agents
Implications for AI Creativity
The study concludes that LLMs’ tendency to regress towards generic phrasing is a consistent property, driven by their next-token prediction objective, training data biases, and decoding methods. Surface novelty, while easily generated, is not equivalent to true creativity. While external structure, such as creative markers, can guide models towards better alignment and richer expression, fully restoring genuine originality remains a challenge.
These findings suggest that future work should focus on designing training objectives that reward diversity and penalize clichés, adopting adaptive sampling strategies, and incorporating richer evaluation measures to foster truly creativity-sensitive LLMs. The research highlights that while LLMs are powerful tools, human guidance and structured signals are crucial to steer them away from mediocrity and unlock their potential for genuine creative generation.
