Is it already too late to reduce catastrophic AI risk? — Harvard Gazette

The Harvard Gazette published an interview in October 2026 under the headline 'Is it already too late to reduce catastrophic AI risk?' The interview considers which actors hold the most power to mitigate significant harms from AI systems and identifies AI model developers as that group, on the grounds that they control the systems that get misused.

The interview states that a long list of safeguards exists but is not consistently implemented. It does not set out which safeguards appear on that list, who compiled it, or how compliance with it is assessed. It also notes a consistent tendency among some users to download open models, train away their harmful-task-refusal behaviours, and re-upload the resulting models, describing this as a recurring practice rather than an isolated case.

The framing places responsibility for mitigation with the developers of the systems rather than with users, regulators or other outside parties. The gap the interview describes is between the existence of a body of safeguards and their consistent application in practice, with the re-uploading of modified open models cited as one route by which safeguards are undone after the fact. The question posed in the headline — whether reducing catastrophic risk is already beyond reach — is the question the interview addresses, with the discussion organised around where the power to act sits.