OpenAI hopes to win back the trust of businesses in handling their highly confidential data. On August 20th, the company reiterated its Zero Data Retention (ZDR) policy with respect to its eligible API users and presented a novel concept of Private Safety Processing, which aims to identify the wrongdoers without anyone from OpenAI having a look at your inputs. Sounds contradictory? Well, it kind of is. Here’s why.
Also read: AI Pro vs AI Plus for students: Google’s free offer isn’t the same everywhere
This is nothing new. This service has been around for some time now as an alternative for enterprise API users who do not want OpenAI to store their prompts or outputs after the request is done. No one from OpenAI is supposed to see that content, and it is not used for training future iterations of the model unless a user chooses to opt-in. This was always the starting point for industries dealing with health, financial information, or confidential business plans when considering the use of OpenAI’s API over the free ChatGPT.
There is, however, one exception that has not been altered. Content marked as possible child sexual abuse material is always stored as per US law requirements.
The problem is this. The current safety systems which are compliant with ZDR are point-in-time and evaluate each interaction separately. However, some of the more dangerous risks of AI, such as an organized jailbreak that spans several interactions or the continued activity of an autonomous agent even after it is ordered to cease doing so, can be detected only in the context of multiple interactions.
Also read: Rewriting the rules of workplace learning: Inside Workday’s bet on the future of L&D
Rival companies have addressed this problem by encouraging their clients to allow the AI company to keep sensitive material in order to conduct the checks. OpenAI believes that such an approach is unacceptable for most of its enterprise clients.
Here’s the new system. It expands safety monitoring to cover relevant interactions without granting OpenAI’s personnel access to the raw material itself. For ZDR instances, all of the data remains on infrastructure controlled by the customer. OpenAI is also developing a system that will store the content on OpenAI’s servers but in an encrypted form that will allow only the customer, not OpenAI, to decrypt it.
The automated systems will monitor the use for signs of misuse. In cases where there appear to be problems, OpenAI gets a “narrowly defined signal,” essentially meaning “this user has triggered signal X.” The enforcement actions are based on that signal alone. Should a customer wish to challenge such a finding, they can submit the necessary material themselves rather than allowing OpenAI to retrieve it.
This is not to say that OpenAI cannot take action against your data. This means that OpenAI employees will not look at it casually, and the automated flagging process will be done without viewing the content, unless you provide it to OpenAI voluntarily during the appeals process. This system is not operational yet. OpenAI is currently testing this with some early adopters and will be launching it in September.
But this is one example in a much larger trend happening everywhere there are significant AI research laboratories at the moment where companies are attempting to repair their enterprise reputations after several privacy missteps while also convincing regulators and the public at large that artificial intelligence can be trusted to act more independently. It will be interesting to see what happens once this goes live.
Also read: Why India’s smartphone protection problem isn’t about awareness