Overview

The hearing did not turn AI risk into a number that is easy to answer. It moved the discussion toward more practical questions: when risk is hard to predict, who verifies it, who can stop a system, and who is responsible when harm occurs?

What the hearing put on the public record

The New York City Council held an AI oversight hearing on October 5. Jacob Coxon, a former Anthropic employee, issued a sharp warning about the industry’s safety culture and pace of development. That was a witness’s judgment drawn from personal experience, not an established risk finding.

Speaker Julie Menin asked companies whether they could quantify the risks. A Meta representative said he did not have the relevant framework in front of him; OpenAI representative Morgan Dwyer gave no probability, saying that any 1%, 10%, or 20% risk would be unacceptable. In a separate exchange about legal responsibility if a model caused serious harm, the representative said the company was responsible for developing and evaluating its systems for safety, while Menin pressed for a clear yes-or-no answer.

A number needs a shared scale

The issue, as I see it, is not simply whether someone can offer a percentage. It is how risk is defined: which capability is being assessed, over what period, on what evidence, and with whose independent review. Without shared terms, a figure can look precise without helping the public decide what to do.

Risk estimates can guide testing and contingency plans, but they cannot by themselves answer who may pause a system, who keeps incident records, or how affected people can seek help. Uncertainty in an estimate should not become a reason to postpone preparation indefinitely.

Make responsibility answerable step by step

Reports said the New York City Council was considering ten proposals touching on independent third-party validation, human shutdown capability, and chatbot privacy. They remained proposals under consideration, not rules enacted at the hearing. The next question is whether broad principles can become clear triggers, independent checks, and enforceable remedies.

Write the process before something goes wrong

Maintaining a website has made one distinction increasingly clear to me: releasing a feature is not the same as establishing accountability for it. Shutdown conditions, who keeps change records, and where readers can turn with a problem should be clear before launch, not improvised after an incident.

A hearing is not a verdict

The hearing did not settle whether catastrophic risks can be quantified reliably, nor did it define the precise legal duties companies would face across different kinds of harm. A former employee’s warning, company testimony, and lawmakers’ proposals occupy different evidentiary positions; they should not be collapsed into a single settled account.

Its practical value was to put “how large is the risk?” and “who is responsible?” on the public record as questions requiring answers. What matters next is not who sounds most certain, but whether proposed rules and independent testing produce mechanisms that can be reviewed and enforced.

Sources consulted

Associated Press, “AI industry insiders voice alarms to NYC council as local governments take up safety concerns,” published October 5, 2026 at 18:45:54 UTC (October 6 at 2:45 a.m. Beijing time). Read the report

CBS News New York, “New York City Council holds landmark AI oversight hearing,” page marked updated October 5, 2026 at 8:34 p.m. EDT (October 6 at 8:34 a.m. Beijing time). Read the report