AI Is Not Safe Yet, Says UCLA Professor

Anthropic and OpenAI are preparing for IPOs while emphasizing AI safety, but UCLA professor Safiya Noble argues current AI systems remain unsafe because stereotypes and biases are embedded in training data. The piece is primarily a qualitative commentary on AI risk and governance rather than a new commercial or financial disclosure.

Analysis

The immediate market implication is not about the headline IPOs themselves, but about pricing power in the AI stack. If the largest model vendors have to spend the next 12-24 months proving governance, auditability, and safety, that shifts commercial leverage toward picks-and-shovels names that sell monitoring, model validation, identity, data provenance, and workflow controls rather than raw inference capacity. That is a subtle but important change: the “trust layer” can become a durable margin pool if enterprise buyers demand indemnities and compliance hooks before scaling deployment.

Second-order, this is mildly negative for the pure-play frontier model companies because safety scrutiny raises the probability of slower enterprise conversion, more disclosure burden, and potentially narrower product design freedom post-IPO. The bigger winner may be incumbents with distribution, existing compliance infrastructure, and the ability to bundle AI into broader software contracts, since they can amortize governance costs across large installed bases. A governance-heavy IPO process also tends to reveal unit economics earlier, which can compress premium private-market multiples if revenue quality depends on aggressive usage growth.

The contrarian angle is that safety criticism is not necessarily bearish for the category; it can be a moat if regulation codifies best practices and raises the barrier to entry. The real risk is a timing mismatch: public-market investors may reward AI growth today while underestimating that regulatory friction often shows up with a 6-18 month lag through slower procurement cycles, litigation, and model policy changes. If one or both IPOs market aggressively on “responsible AI,” expect the next catalyst to be whether enterprise demand proves it is a monetizable feature rather than a cost center.

Key Decisions for Investors

• Long MSFT / short a basket of pre-IPO frontier-model proxies on any post-IPO strength: the trade favors companies that can absorb governance costs into existing enterprise distribution over standalone model vendors with higher scrutiny risk; target a 3-6 month horizon and trim if enterprise adoption metrics accelerate faster than expected.
• Overweight AI infrastructure beneficiaries with compliance adjacency, such as NOW or SNOW, on weakness: these firms can sell governance, observability, and data controls as part of the AI rollout, with a favorable risk/reward if enterprise procurement shifts toward trusted platforms over model novelty.
• Buy 6-12 month call spreads on CRWD or PANW: if AI regulation tightens, security and access-control budgets should benefit from incremental demand for model/data protection; structure for limited premium outlay and upside from a compliance-driven spend cycle.
• Avoid chasing valuation in unprofitable AI IPOs until after the first post-listing earnings print: the risk is multiple compression if disclosed retention, gross margin, or inference-cost trends disappoint once public reporting forces transparency.