OpenAI stated that they might revise their safety measures if another firm unveils a high-risk artificial intelligence system lacking safeguards.
OpenAI wrote in its
Preparedness Framework
report that if another company releases a model that poses a threat, it could do the same after “rigorously” confirming that the “risk landscape” has changed.
This document outlines the process through which the firm monitors, assesses, predicts, and safeguards itself from potential catastrophes caused by artificial intelligence models.
“If another front-runner AI developer unveils a high-risk system lacking similar safety measures, we might revise our criteria,” OpenAI stated in a blog post released on Tuesday.
Nonetheless, we would initially thoroughly verify that the risk environment has genuinely shifted, openly admit that we are implementing changes, evaluate whether these adjustments do not substantially elevate the total risk of significant damage, and simultaneously maintain protections at a higher safeguarding threshold.
Prior to making a model available to the broader audience, OpenAI assesses potential significant harm by pinpointing likely, quantifiable, novel, serious, and irreversible risks. They also develop protective measures for such risks. These dangers are subsequently categorized into four levels: low, medium, high, or critical.
Several of the risks that the firm already monitors include their models’ abilities in areas such as biology, chemistry, cybersecurity, and self-enhancement.
The firm mentioned it’s also assessing fresh risks, including the capability of their AI model to function continuously without human intervention, its potential for self-replication, and the threats this might entail within the realms of nuclear and radiological sectors.
“Risks associated with persuasion,” like the use of ChatGPT for political campaigns or lobbying, will be managed separately from this framework and examined differently.
Model Spec
, the document that dictates ChatGPT’s behavior.
‘Quietly reducing safety commitments’
Steven Adler, a former OpenAI researcher, said on
X
that the revisions in the firm’s readiness document indicate it is “stealthily scaling back its safety pledges.”
In his posting, he highlighted a December 2023 pledge from the firm to test “customized versions” of their artificial intelligence models; however, he also mentioned that OpenAI plans to focus solely on testing models where the training parameters or “weights” will be made public.
“People might completely differ on whether fine-tuned model testing is necessary and preferable. It’s better for OpenAI to withdraw a commitment altogether rather than keeping it without following through,” he stated.
But in both scenarios, I would prefer OpenAI to be more transparent about retracting their earlier promise.
This announcement follows the release of a new series of AI models from OpenAI this week, known as GPT-4.1, apparently unveiled without
system card
Or safety report. Owing to this, Euronews Next reached out to OpenAI for comment regarding the safety report; however, they had not received a response by the time of publishing.
This follows the action of 12 ex-OpenAI staffers who have lodged a complaint.
brief
last week in Elon Musk’s case brought against OpenAI, which alleges that a shift to a for-profit company could lead to corners being cut on safety.
