How this editorial can be challenged
Who carries the cost between the moment an employee reports a risk and the moment someone with authority acts?
The engineer who reports a flaw pays in time, uncertainty and sometimes professional standing. If the decision disappears into another team, only the reporter has a personal reason to keep chasing it.
Employees can be wrong, and disclosing how to exploit a vulnerability could make it much harder to fix. Leaders cannot hand every person who files a concern veto power or access to sensitive files.
A private response can say that an owner reproduced the problem, declined it with reasons, or scheduled a retest. None requires giving the reporter a veto or making attack instructions public.
Interviews describe what some employees experienced. The public record is not a register of every complaint, the answer each received, or how many people chose not to speak.
An independent sample of both accepted and rejected reports showing timely, reasoned responses, successful retests where needed and confidential follow-up available to departing workers would weaken the case that reporters must carry these risks alone.
The job nobody assigned
Imagine you find a door that should lock and doesn't. You can tell the facilities team in thirty seconds. The expensive part comes three weeks later, when the door still looks the same. Was it repaired? Did someone decide it wasn't a problem? Do you test it again and risk being the colleague who won't let things go?
That's a second job, and the person who first noticed the fault is the one most likely to inherit it. It's unpaid in the ordinary sense: their deadlines haven't moved. Much of the argument over whether AI workers are free to speak misses what it costs them to keep speaking.
An open channel and an unanswered question
Reuters reports that staff at OpenAI brought security concerns to leaders and the board, some before ChatGPT was launched. The company says access to its most sensitive work is restricted and it invests heavily in security. The two accounts don't establish which individual warning received which response. They don't have to contradict each other to leave an employee with a reasonable question: did anyone test the thing I reported?
Credit for the unflattering detail
OpenAI's own technical account contains the most encouraging detail I read this morning. Staff checking fixes found that merging a patch didn't mean it was running everywhere it belonged. They widened the checks. A team willing to look twice learned something that its first completion count concealed.
That example doesn't vindicate or disprove the earlier employee complaints. It does show what a useful reply to a colleague can look like. Not ‘we take this seriously,’ but ‘we checked where the change actually runs; here is what is still outstanding.’ You can send that answer privately.
Who can afford to find out?
At Google DeepMind, The Washington Post reports that some employees in Britain want existing non-disparagement language changed. Google says it removed the post-employment restriction from terms for new hires and does not enforce it against former workers. An assurance has value. So does wording that remains clear after the manager who gave the assurance leaves.
Compare the position of a junior researcher with that of Geoffrey Hinton, who has a Nobel Prize and can air doubts in a Guardian interview. He also described a good future as possible. His standing doesn't prove his predictions. It does mean speaking in public costs different people different things. The absence of a second famous dissenter is a poor measure of how many colleagues have doubts.
The right to hear no
The strongest objection is that security teams would drown in appeals if everyone who filed a concern could stop a release. Some warnings will be mistaken; some contain details that really should stay secret. And UK whistleblowing protections have conditions: not every disagreement about technology qualifies.
Give a reporter a way to hear no. Name someone who can explain the decision in private. Distinguish ‘not reproduced’ from ‘not important enough to fix now.’ Offer a route to another reviewer when there is new evidence. This gives the lab discretion without making persistence a test of personal courage.
A quieter office can be a better one, or a worse one
Managers sometimes want a number that will tell them whether a culture is getting healthier. Reports going down is an especially bad one. It might mean the recurring fault was repaired. It might mean the last colleague who raised it spent months chasing an answer and everyone else noticed.
I'd start instead by asking a few people whose reports were turned down what happened. Could they find the person who made the call? Did they understand the reason, even if they disagreed? A lab that answers those questions well might hear more complaints next quarter. That could be the first sign that something is working.
Read the reporting
Opinion is ours. The factual record is linked below.
Reuters — staff accounts and OpenAI's response OpenAI — field notes on retesting deployed fixes The Washington Post — contract, staff interviews and Google's response The Guardian — Geoffrey Hinton interview UK government — what counts as whistleblowing