With artificial intelligence becoming ever more deeply embedded in daily life, Microsoft has opened the doors to its AI engine room, offering a glimpse into the arduous process behind creating AI technologies that people can trust.

In its latest Responsible AI Transparency Report, the tech giant charts its progress over the past year in pursuit of ‘responsible innovation’, highlighting new model development, stronger safety frameworks, and hands-on guidance for customers navigating high-stakes AI.

The company claims that ‘North Star’ principles like fairness, safety, privacy, transparency, and accountability are the bedrock of its AI creation, while generative AI systems are built using a ‘govern, map, measure, and manage’ approach designed to mitigate risks across development lifecycles.

In practice, that means Microsoft has, for instance, established an Office of Responsible AI to guide policy and product development, shaped by evolving global regulations, while at the same time, a dedicated engineering team builds tools and technical guidance to help users meet responsible AI standards across Microsoft’s platforms.

Meanwhile, the tech trend setter has also upped its threat mapping process, essentially hunting out the bugs in AI models that might give rise to any new or previously unforeseen risks in its development.

According to the report, in 2024, Microsoft’s AI Red Team (AIRT) carried out 67 operations across its Copilot tools and other AI models, including every major model added to Azure OpenAI Service and each new Phi release, with work extending into research in emerging areas like image-to-image, video-to-text, text-to-video, and text-to-audio AI.

Once Microsoft assesses the capabilities of a new AI model, it applies a ‘defence in depth’ strategy to manage potential risks, including using content classifiers to filter out harmful inputs and AI-generated outputs, while at the application level, models are ‘grounded’ with input data and supported by safety messages to ensure outputs align with responsible AI standards and expectations.

While other AI developers continue to grapple with content safety, Microsoft claims it made significant progress last year in detecting and blocking unsafe AI-generated material, including sexual, violent, and hateful content, as well as copyrighted material like song lyrics, news articles, and code.

However, it is the Sensitive Uses and Emerging Technologies programme, which Microsoft seems most proud of, which provides hands-on counselling for high-impact and higher-risk uses of AI.

In 2024, 77% of cases that received consultations from Microsoft’s Sensitive Uses and Emerging Technologies team were related to genAI development, such as the use of voice capabilities in Copilot, with a growing number involving the use of agentic AI applications. 

Recommended reading

• KPMG UK Interview: Why Trust is AI’s Most Underrated Asset
• New Report Reveals AI Trust Divide in Cybersecurity
• Trust Divide Emerges Over AI in Customer Communication

“The past year has seen a wave of AI adoption by organisations of all sizes, prompting a renewed focus on effective AI governance in practice,” wrote Teresa Hutson, Microsoft’s corporate vice president, and Natasha Crampton, the firm’s chief responsible AI officer, in the report’s foreword.

“In the year ahead, we will focus on developing more flexible and agile risk management techniques, advancing a vibrant ecosystem through shared norms and effective tools, and supporting effective governance across the AI supply chain. 

“Together, these efforts will enhance the abilities of not only AI developers and deployers like Microsoft, but also our customers and partners, to implement governance efficiently and effectively, fostering the trust we need at a pace that matches AI innovation.”