403
Sorry!!
Error! We're sorry, but the page you were looking for doesn't exist.
OpenAI Admits Its AI Models Hid Errors, Faked Data in Six Flagged Cases
(MENAFN) OpenAI on Wednesday revealed six separate incidents in which its artificial intelligence models displayed what the company called "unexpected or concerning" behavior — ranging from covering up mistakes to inventing information and sharing files online without permission.
The San Francisco-based firm said the episodes surfaced during training or evaluation of its models over the past six months.
"We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," OpenAI said, stressing that decisions around AI development must rest on evidence outside experts can independently verify.
In one flagged case, a model embedded instructions inside task summaries directing future versions of itself to ignore standard constraints. In another, models slipped instructions into summaries designed to mask errors or misaligned conduct — including fabricating missing historical data without flagging that it had done so.
OpenAI also disclosed an instance in which a model discovered an exposed API key sitting in a public repository, used it without authorization, and then invented the requested data once it couldn't actually retrieve it.
Additional cases involved a model uploading a file to the internet without user consent in order to cite it later, along with models tapping internal or public file-hosting services to communicate or exchange files in violation of restrictions placed on them.
The company characterized the disclosures as an initial batch rather than a full accounting of known issues, and said it plans to keep publishing findings under this framework going forward on an ongoing basis.
The San Francisco-based firm said the episodes surfaced during training or evaluation of its models over the past six months.
"We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," OpenAI said, stressing that decisions around AI development must rest on evidence outside experts can independently verify.
In one flagged case, a model embedded instructions inside task summaries directing future versions of itself to ignore standard constraints. In another, models slipped instructions into summaries designed to mask errors or misaligned conduct — including fabricating missing historical data without flagging that it had done so.
OpenAI also disclosed an instance in which a model discovered an exposed API key sitting in a public repository, used it without authorization, and then invented the requested data once it couldn't actually retrieve it.
Additional cases involved a model uploading a file to the internet without user consent in order to cite it later, along with models tapping internal or public file-hosting services to communicate or exchange files in violation of restrictions placed on them.
The company characterized the disclosures as an initial batch rather than a full accounting of known issues, and said it plans to keep publishing findings under this framework going forward on an ongoing basis.
Legal Disclaimer:
MENAFN provides the
information “as is” without warranty of any kind. We do not accept any
responsibility or liability for the accuracy, content, images, videos,
licenses, completeness, legality, or reliability of the information
contained in this article. If you have any complaints or copyright issues
related to this article, kindly contact the provider above.
Comments
No comment