Microsoft releases inaugural AI transparency report

In its Accountable AI Transparency Report, which primarily covers 2023, Microsoft touts its achievements round safely deploying AI merchandise. The annual AI transparency report is without doubt one of the commitments the corporate made after signing a voluntary settlement with the White Home in July final yr. Microsoft and different corporations promised to determine accountable AI programs and decide to security.

Microsoft says within the report that it created 30 accountable AI instruments previously yr, grew its accountable AI workforce, and required groups making generative AI functions to measure and map dangers all through the event cycle. The corporate notes that it added Content material Credentials to its picture technology platforms, which places a watermark on a photograph, tagging it as made by an AI mannequin. 

The corporate says it’s given Azure AI clients entry to instruments that detect problematic content material like hate speech, sexual content material, and self-harm, in addition to instruments to judge safety dangers. This consists of new jailbreak detection strategies, which had been expanded in March this yr to incorporate oblique immediate injections the place the malicious directions are a part of information ingested by the AI mannequin.

It’s additionally increasing its red-teaming efforts, together with each in-house pink groups that intentionally attempt to bypass security options in its AI fashions in addition to red-teaming functions to permit third-party testing earlier than releasing new fashions.

Nonetheless, its red-teaming models have their work reduce out for them. The corporate’s AI rollouts haven’t been proof against controversies.

Natasha Crampton, chief accountable AI officer at Microsoft, says in an electronic mail despatched to The Verge that the corporate understands AI continues to be a piece in progress and so is accountable AI. 

“Accountable AI has no end line, so we’ll by no means think about our work below the Voluntary AI commitments carried out. However we now have made robust progress since signing them and look ahead to constructing on our momentum this yr,” Crampton says.