Q&A with cognitive scientist Gary Marcus on technical limitations and moral inadequacies of LLMs, constructing a regulatory framework for AI, and more
The U.S. should create an artificial intelligence agency led by a cabinet-level position, Marcus argues Bluesky: @privacyguru.bsky.social . X: @aratofrat and @amy_brand . LinkedIn: Steve Rosenbush Bluesky: Debra J Farber / @privacyguru.bsky.social : Gary Marcus has been right all along to critique LLMs capabilities (for which he was ridiculed) and his latest post lays it out perfectly. He calls out the limitations of LLMs while pushing for a future where AI systems are truly reliable, robust, & aligned with human needs. [embedded post] X: @aratofrat : @WSJ Gary Marcus has it wrong. The last thing we need is regulation (control) by the government. Encourage collaboration between private industry on best practices, and adopt protections present in laws like the patient safety quality improvement act. Non-governmental frameworks @amy_brand : Cognitive scientist Gary Marcus says AI regulation should include a cabinet-level government position and a federal department modeled on the FDA. That's just the beginning. https://www.wsj.com/... via @WSJ LinkedIn: Steve Rosenbush : My column this week: An in-depth discussion with cognitive scientist Gary Marcus, who was an early voice expounding on the limitations of large language models …
Context & Ripple Effects
Marcus's case for a dedicated AI regulator extends a long-running argument that scaling language models is not a sufficient route to dependable systems. That critique was reinforced in related coverage of research finding no formal reasoning in language models and earlier debate over whether a paradigm shift is needed beyond deep learning.
The governance proposal also lands against a record of limited U.S. legislative action on AI, while legal uncertainty around model outputs has made questions of accountability more concrete. It frames technical reliability and institutional oversight as linked rather than separate issues.
First-order effects
- Marcus's proposal gives policymakers and AI-governance advocates a specific institutional model—a cabinet-level agency, modeled on the FDA—to debate instead of relying solely on broad calls for AI rules.
- AI developers face a more explicit critique of LLM deployment: demonstrating usefulness is distinct from demonstrating reliability, robustness, and alignment with human needs.
Second-order effects
- The intervention sharpens a divide between proponents of formal public oversight and those, such as commentators examining liability for generative-AI outputs, who must account for legal exposure even without a new agency.
- If reliability becomes a central regulatory test, model builders and customers may place more weight on evaluation, safeguards, and documented limits—not just capability claims.
Third-order effects
- The broader contest is over whether AI governance should be organized around product-style preconditions for deployment or industry-led standards; the eventual answer will shape how much compliance capacity becomes a competitive advantage.
- Persistent evidence that LLM behavior is pattern-based rather than reasoned could shift attention from scaling alone toward systems designed to meet auditable reliability requirements, though that technical transition remains unsettled.
The trend: AI policy is increasingly converging on the idea that claims of advanced capability must be matched by demonstrable reliability and accountable deployment institutions.