Silicon Valley promised AI would democratize creativity. New research tracking 442 participants found the opposite: people who were more creative without AI produced better work with it. The gap didn't close. It may have widened.
Enterprises are spending billions on AI pilots, but MIT's research shows most deliver no return. It's not the technology failing. The gap between impressive demos and working systems comes down to data quality, technical debt, and organizational readiness.
Good Morning from San Francisco, Anthropic crowned its new model the coding champion. The White House invoked the Manhattan Project
Silicon Valley promised AI would democratize creativity. New research tracking 442 participants found the opposite: people who were more creative without AI produced better work with it. The gap didn't close. It may have widened.
Silicon Valley keeps selling generative AI as the great equalizer. Hand everyone the same tools, watch the playing field flatten. That's the pitch, anyway. Struggling writers get unstuck. Untrained designers produce professional work. The cognitive elite watches its monopoly crumble.
A pair of studies out of Penn State and UConn complicate this story considerably.
Researchers tracked 442 participants through creative tasks performed both with and without AI assistance. The pattern that emerged won't comfort democratization optimists: people who scored higher on creativity and intelligence measures before touching an AI tool also produced superior work after gaining access to one. The gap between high and low performers? It stayed open. Some of the data hints it may have widened.
Key Takeaways
• Baseline creativity predicted AI-assisted performance at β = .42 in Study 1; intelligence and creativity together explained 40% of variance in Study 2
• Researchers isolated AI-assisted creativity as a distinct cognitive construct, separable from both general intelligence and baseline creative ability
• Findings contradict earlier democratization research, suggesting cognitive advantages persist or amplify when everyone gets access to the same AI tools
Simone Luchini's team at Penn State built their experiments around a specific question: does AI actually level creative ability, or does that claim collapse under scrutiny?
The first study recruited 263 undergraduates and had them write short stories. Each participant wrote twice, once with ChatGPT available as an assistant and once flying solo. Order was randomized. To score the stories, the team deployed a custom transformer model called MAoSS that predicts human originality ratings. They also measured how semantically diverse each story was and how much detail writers packed in. On a separate track, participants sat for cognitive tests: pattern completion for fluid reasoning, synonym matching for vocabulary, and a timed animal-naming task for verbal fluency.
What showed up in the data: how well someone wrote without AI predicted how well they'd write with it. The standardized coefficient hit β = .42. That's not a subtle effect. Strong writers stayed strong. Weak writers stayed weak. The AI didn't close the distance.
For the second study, 184 participants came through Prolific. The design grew more ambitious here. Baseline creativity got measured through tasks entirely separate from the AI-assisted ones, including scientific hypothesis generation, freehand drawing, engineering design challenges, and metaphor creation. Then participants tackled AI-assisted tasks with GPT-4o: social media posts, product reviews, business pitches, job interview responses.
Same pattern. Creativity predicted AI performance (β = .39). So did general intelligence (β = .35). Combined, these baseline abilities explained 40% of the variance in how well people created with AI assistance.
The researchers also managed something methodologically interesting. They demonstrated that AI-assisted creativity functions as its own distinct construct. Performance on one AI task predicted performance on different AI tasks, even across substantially different creative challenges. Writing a sharp social media post correlated with generating a persuasive business pitch.
These findings run against some earlier work. Anil Doshi and Oliver Hauser published a 2024 study showing GPT-4 access helped weaker performers close the gap with stronger ones. Their less creative participants improved enough to match the output quality of more capable writers.
The contradiction matters less than it might seem. Doshi and Hauser measured baseline creativity with a single word association task. Luchini's team used validated multi-measure assessments across diverse domains. Methodological choices shape conclusions.
Other recent research aligns with the Penn State findings. Ethan Zhou and John Lee tracked visual artists and found that those producing more novel work before generative AI kept producing more novel work after adopting it. William Orwig's team observed similar persistence of advantage in AI-assisted visual art. Paul DiStefano's group found the same examining human-AI co-creativity more broadly.
The democratization narrative rested on thinner evidence than advertised.
Forty percent explained. Sixty percent unexplained. That's a lot of variance left on the table.
What else drives AI collaboration success? The researchers offer possibilities without settling on answers. Personality might matter. Attitudes toward AI. Skill at crafting prompts or knowing when to reject suggestions.
Related work from the same research group points toward interaction quality. How people engaged with the AI during tasks, the questions they asked, how they iterated on suggestions, all of it predicted creative outcomes. Those who treated the AI as a genuine collaborative partner rather than a magic output machine got better results.
This hints at something learnable. High baseline creativity doesn't automatically translate to effective AI use. And people with modest creative ability might compensate through superior collaboration instincts.
The studies captured full interaction logs between participants and GPT-4o. Mining those logs could reveal what distinguishes productive AI collaboration from wheel-spinning. What questions do high performers ask? When do they push back on suggestions? Future analysis might tell us.
Study 2 buried a revealing finding in its secondary analysis. When researchers isolated which baseline creativity measures independently predicted AI performance, only one survived: metaphor generation.
Scientific creativity dropped out. Drawing ability dropped out. Engineering design problem-solving dropped out. Verbal creativity alone, specifically the capacity to generate novel metaphors, retained predictive power after controlling for other measures.
The explanation isn't mysterious. Every AI-assisted task in Study 2 involved writing. Domain matching mattered. If participants had used AI for visual creation or scientific hypothesis generation, different baseline abilities would likely have predicted success.
Practically, this suggests AI collaboration skills may fragment by domain. Using AI effectively for design work might require different capacities than using it effectively for business writing. Training programs may need to specialize accordingly.
The researchers state the implication directly: cognitive ability assessments remain relevant for AI-augmented roles. Employers screening candidates on creativity and intelligence can expect those measures to predict performance even when employees have access to powerful generative tools.
This cuts against some recent thinking. The World Economic Forum and similar institutions have suggested AI might erode the premium on traditional cognitive skills, shifting value toward emotional intelligence or ethical judgment. The Penn State data suggests otherwise. Old-fashioned creativity and intelligence keep predicting outcomes.
For education, the picture splits. On one hand, schools should keep developing students' creative and intellectual capabilities. These transfer to AI contexts. On the other hand, AI-assisted creativity emerged as its own distinct ability in these studies. Knowing how to think creatively without AI doesn't fully prepare someone to create effectively with it. That's a separate skill requiring separate development.
The 40% explained variance is substantial. But 60% lies elsewhere. Education focused only on baseline cognitive development misses something important about AI-era performance.
Some caveats matter here. The first study pulled exclusively from one university's intro psych course, and the sample ran heavily female. Prolific gave the second study more demographic variety, though women still outnumbered men. Neither study looked at people with dyslexia, ADHD, or other cognitive differences. Those populations might interact with AI tools in ways this research couldn't capture.
Both studies also locked themselves to particular AI models: ChatGPT-3.5-Turbo and GPT-4o. Snapshots of a moving target. Models with different architectures, multimodal capabilities, or agentic behaviors might alter the relationship between human abilities and AI-assisted output. What held for GPT-4o in mid-2025 may not hold for whatever ships in 2026.
One methodological wrinkle deserves attention. Study 2 used three LLMs as creativity judges: GPT-4.1, Gemini 2.5 Flash, and Claude Sonnet 4. AI evaluating human creativity in work produced with AI assistance. The judges agreed with each other quite well (ICC = .88). Whether their consensus matches what humans would call creative is a different question entirely.
Strip away the methodology and this research asks something fundamental about technological equity. When powerful tools become universally available, do human differences shrink or grow?
Silicon Valley has mostly bet on shrinkage. The internet democratized information. Social media democratized broadcasting. Generative AI democratizes creative production. So the logic runs.
But access and outcomes aren't the same thing. Word processors are universal. Writing quality isn't. Internet access is nearly universal. Information literacy varies wildly. The Penn State research suggests generative AI follows this pattern. Everyone gets the tool. Results diverge anyway.
The studies measured relative performance rather than absolute improvement. Someone with modest creative ability probably produces better work with AI than without. But if stronger creators improve even more with AI assistance, the gap between them grows rather than shrinks.
Scale this beyond individual performance. If AI amplifies cognitive differences rather than compressing them, the inequality concerns aren't speculative anymore. They have data behind them. People with stronger baseline abilities may capture outsized benefits from AI augmentation, while others gain less from the same tools.
Q: What does β = .42 actually mean in plain terms?
A: A standardized coefficient of .42 means that for every one standard deviation increase in baseline creative writing ability, AI-assisted performance increased by .42 standard deviations. In practical terms, this is a moderate-to-strong effect. Someone in the top 15% of writers without AI would typically land around the top 25% when using AI.
Q: How did researchers prevent participants from just copying AI output?
A: The studies disabled copy-paste functionality entirely. Participants could see AI suggestions in a chat panel but had to type their stories manually in a separate writing area. System prompts also constrained ChatGPT to provide only bullet-point suggestions rather than full passages, preventing the AI from writing complete stories even if asked.
Q: How did they score creativity without human raters?
A: Study 1 used MAoSS, a transformer model trained to predict human originality ratings on short stories. Study 2 used an ensemble of three LLMs (GPT-4.1, Gemini 2.5 Flash, Claude Sonnet 4) as independent judges. The three-model ensemble achieved ICC = .88 inter-rater reliability, meaning the AI judges agreed with each other at rates comparable to human raters.
Q: Can people learn to collaborate with AI more effectively?
A: Likely yes. The studies found that 60% of variance in AI-assisted performance remained unexplained by baseline creativity and intelligence. Related research from the same team showed that interaction quality, meaning how people questioned the AI and iterated on suggestions, predicted outcomes. This suggests a learnable skill component beyond raw cognitive ability.
Q: Is this research peer-reviewed?
A: The paper is currently a preprint available on Open Science Framework (osf.io/b4ztn), meaning it hasn't yet completed peer review. However, the research team has published extensively in peer-reviewed journals on creativity assessment, and all data and analysis code are publicly available for scrutiny. The studies use established psychometric methods and validated assessment tools.
Marcus Schuler
Maria Garcia
Marcus Schuler
Facebook claims 52% daily usage while TikTok hits 24%, suggesting clear dominance. But Pew's survey measures visits, not time spent. That distinction reshapes everything about platform power, ad economics, and which apps actually own user attention.
Deezer receives 50,000 AI tracks daily—34% of all uploads. Yet they generate just 0.5% of streams, with 70% of plays flagged as fraud. The flood isn't about whether AI sounds convincing. It's about zero-cost content enabling industrial-scale royalty theft.
DeepMind's AlphaEvolve can search millions of mathematical constructions in hours, not weeks. Fields Medalist Terence Tao already builds on its outputs. But the system finds candidates, not proofs. The real shift: math discovery at industrial scale.
Enterprises report 74% positive AI returns while cutting training budgets 8%. The Wharton study reveals companies extracting productivity gains today by depleting tomorrow's capabilities—a business model that works until skills erode.
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
