Turns out AI is even less creative than we thought

A study published in Patterns found that when Stable Diffusion XL and LLaVA iteratively generated images for 100 rounds, the outputs rapidly drifted toward a small set of 12 recurrent visual motifs (e.g., pastoral landscapes, rainy nightscapes, beaches, generic interiors) across 1,000 iterations. The research highlights a pronounced lack of visual diversity in current generative-image workflows, underscoring limits to creative originality and potential challenges for firms and advertisers that rely on AI to produce distinctive visual content.

Analysis

Market structure: The research increases differentiation between commoditized generative outputs and high-value bespoke creative content. Winners include GPU/cloud providers (NVDA, AMZN, MSFT, GOOGL) and software that enables fine‑tuning/asset management (ADBE) because clients will pay for models trained on proprietary datasets; losers are low-margin stock/image mills and template ad aggregators (e.g., pure-play image brokers) that compete on volume, risking ~5–15% price compression over 12–24 months. Expect pricing power to concentrate in owners of unique IP and curated datasets, not generic model hosts.

Risk assessment: Tail risks include adverse copyright rulings or compulsory licensing regimes (10–25% probability over 12–24 months) and reputational/advertiser flight if platforms keep serving “elevator music” creatives (could cut ad ROI by >5% and trigger product pivots within months). Hidden dependencies: model performance depends on diverse, labeled datasets and content-moderation tooling; failure to invest in those drives customer churn. Catalysts to watch: major model releases, high‑profile legal judgments, or ad-platform A/B tests showing CTR drops >3% that accelerate migration to human-created ads.

Trade implications: Favor infrastructure and premium creative software exposure: NVDA (compute demand) and ADBE (creators ecosystem) over commoditized image providers. Use defined-risk option structures to express exposure: buy-call spreads on NVDA 3–6 month expiries 15–25% OTM to capture continued fine-tuning demand while limiting capital. Consider a small short/hedged exposure to pure-play stock-image providers or ad-template platforms sized at 1% of NAV to harvest mean reversion if they fail to pivot.

Contrarian angle: The consensus that “AI destroys creative value” misses the bifurcation — human originality will command price premiums, so companies that facilitate premium marketplaces or licensing (legacy media, boutique marketplaces, IP licensors) can outperform unexpectedly. Historical parallel: commoditization of stock photography in the 2000s led to resurgence in premium commissioned work; if courts force licensing fees, licensors (e.g., major publishers/IP holders) could see a 5–10% revenue upside within 2 years. Watch legal outcomes and ad-CTR signals to identify entry points.

Key Decisions for Investors

• Establish a 2–3% long position in NVDA (or equivalent GPU-focused ETF exposure) using a 3–6 month call spread 15–25% OTM to capture continued fine-tuning and compute demand while capping downside.
• Allocate 2% long to ADBE (creative software + asset management) on the view premium creators retain pricing power; add on any pullback >8% within 3 months.
• Initiate a 1% short position in a pure-play stock-image/templated-ad provider (e.g., SSTK) funded by selling 3-month OTM calls or buying 3-month 10% OTM puts if available, anticipating margin pressure and client churn.
• Monitor and be ready to rotate 1–2% into legacy IP licensors/media (e.g., large publishers or Getty-equivalent) if a court ruling within 12–24 months mandates training-data licensing or sets damages >$50M; enter after ruling with a 6–12 month horizon.