Over the past few months, we have seen stories about how some users have succeeded in creating content via generative AI services that get around content guardrails. Perhaps the biggest example happened earlier this year when explicit deepfake images of pop artist Taylor Swift made their way to social networks. They were reportedly created by a group that used certain text prompts in the AI image generator Microsoft Designer.
Today, Microsoft announced a public preview of a new feature in its Azure OpenAI Service. This service gives businesses and organizations access to OpenAI"s generative AI models over Microsoft"s Azure cloud servers. The new "Risks & safety monitoring" feature is designed to give businesses a real-time look at how their AI systems are being used and detect when people are trying to abuse those systems.
In a blog post, Microsoft says this feature can show information on a variety of different end-user requests on their AI services that are considered harmful. They include:
- Total blocked request count and block rate
- Blocked requests distribution by category
- Block rate over time by category
- Severity distribution by category
- Blocklist blocked request count and rate
Based on that information, businesses can make changes to their AI content guardrails so their services are not used against their guidelines.
There"s also a feature that will help businesses detect users who are constantly making abusive or harmful requests to their AI services. Microsoft says:
If any content from a user is flagged as harmful and combining the user request behavior, the system will make a judgement on whether the user is potentially abusive or not. Then a summarized report will be available in the Azure OpenAI Studio for further action taking.
The business can then make changes to its system based on its own code of conduct to prevent its system from being compromised by these abusive end users.
The new monitoring features for Azure OpenAI Service are available for customers in the East US, Switzerland North, France Central, Sweden Central, and Canada East regions. There"s no word on when this public preview will be expanded to more countries or reach general availability.