The news
Gary Marcus appeared on Bloomberg’s “The Close” with Romaine Bostick on October 2. He stated that developers, including OpenAI, do not have sufficient safeguards in place for systems they cannot reliably control. Marcus is identified in the segment as co-founder of Robust AI Inc. and a former CEO of the company.
The exchange lasted several minutes and stayed on the single point that capability gains have not been matched by mechanisms for verification or intervention. Marcus did not present new data or internal documents from any lab. He spoke from his dual roles at New York University and Robust AI Inc.
Context
Marcus has held a faculty position at New York University for years and has repeatedly questioned whether current neural-network approaches can ever be made verifiable. The Bloomberg segment returned to that theme at a moment when multiple labs continue to ship larger models on short release cycles. The prior state, as Marcus described it, is one in which operators deploy systems whose internal states remain inaccessible even to the teams that trained them.
No technical specifications, safety reports, or counter-statements from OpenAI or other labs were aired during the interview. The discussion therefore rested entirely on Marcus’s assessment that the absence of control methods is structural rather than temporary.
Detail
Marcus framed the issue as one of controllability rather than scale. He said developers continue to ship systems whose internal behavior remains opaque even to their creators. The interview did not include specific metrics, incident reports, or responses from OpenAI or any other lab named in the discussion. Marcus linked the lack of safeguards directly to the inability to exert reliable control once models are released.
He argued that post-training inspection cannot substitute for built-in mechanisms that would let operators detect when a model has left its intended operating envelope. The segment offered no timeline for when such mechanisms might appear in public releases.
Why it matters
The interview restates a standing critique: progress in capability has outpaced progress in verification. When a system cannot be audited in any systematic way, claims about its safety rest on post-hoc observation rather than engineering guarantees. Marcus’s position implies that further scaling alone will not close this gap, because the underlying architectures were never designed for verifiable behavior.
For teams that must integrate these models into products, the practical consequence is continued reliance on testing regimes that cannot enumerate all failure modes. Regulators and enterprise buyers therefore face the same information asymmetry that Marcus described on air. They receive performance numbers on selected benchmarks but receive no independent means to confirm that the model will stay within declared bounds once deployed at scale.
The segment offers no timeline for when or whether labs will publish methods that would allow external verification of control. Until such methods appear, the default stance for anyone deploying the technology remains one of limited visibility into the systems they operate. That stance affects procurement decisions, insurance underwriting, and regulatory filings in equal measure. Each of those activities requires evidence that a deployed system can be monitored and, if necessary, halted—an evidence base Marcus says is still missing.
---
Sources:
No comments yet