Unlock Editor’s Digest Lock for Free
Microsoft is starting to rank artificial intelligence models based on safety performance as Software Group seeks to build trust with cloud customers to sell AI products such as Openai and Elon Musk’s Xai.
Sarah Bird, Microsoft’s Head of Responsible AI, said the company will soon add the “Safety” category to its “Model Leaderboard.” This is a feature that the developers launched this month to rank iterations from a variety of providers, including China’s Deepseek and French Mistral.
A leaderboard that can be accessed by tens of thousands of clients using the Azure Foundry Developer Platform is expected to affect AI models and applications purchasing through Microsoft.
Microsoft currently ranks three metrics: quality, cost and throughput. This is how quickly the model produces output. Bird told the Financial Times that the new safety rankings ensure that “people can shop in person and understand” when deciding what the capabilities of the AI model “to buy.”
The decision to include a safety benchmark is because Microsoft customers tackle the potential risks posed by new AI models for data and privacy protection, especially when deployed as autonomous “agents” that function without human supervision.
Microsoft’s new safety metrics are based on its own toxicgen benchmark, which measures implicit hate speech, and AI Safety’s center for weapons of mass destruction proxy benchmarks. The latter evaluates whether the model can be used for malicious purposes such as building biochemical weapons.
Rankings allow users to access objective metrics when selecting from a catalog of over 1,900 AI models, allowing users to create choices based on the information available.
Cassie Kozyrkov, Google consultant and former chief decision scientist, said: “The real challenge is to understand trade-offs. What cost is higher performance? What risks and lower costs?”
Alongside Amazon and Google, the Seattle-based group is considered one of the biggest “hyperschools” dominating the cloud market.
Microsoft has also established itself as an agnostic platform for generating AI, signing a deal to sell Xai and human models, and a rival to Start-up Openai, which was backed by an investment of around $14 billion.
Last month, Microsoft said it would begin offering a model for Xai’s Grok family in the same commercial terminology as Openai.
This move happened despite the version of Grok that raised the alarm when “invalid changes” in the code repeatedly referenced “white genocide” in South Africa when responding to questions on social media site X. Xai said it has introduced a new monitoring policy to avoid future incidents.
“The model has a platform, there’s some internal review, and then it’s up to the customer to understand that using benchmarks,” Bird said.
There are no global standards for AI safety testing, but the EU AI law will be enforced later this year, enforcing businesses to conduct safety tests.
Recommended
Some model builders, including Openai, show that they are spending less time and money to identify and mitigate risks. FT reports that it has previously cited several people who are familiar with the startup safety process. The startup said it had identified efficiency without compromising safety.
Bird declined to comment on Openai’s safety testing, but said it was impossible to ship high-quality models without investing “a huge amount” in the assessment, and the process was automated.
In April, Microsoft also launched an AI Reading Team Agent that automates the processes of stress testing computer programs by launching an attack to identify vulnerabilities. “You just specify the risk, you specify the level of attack difficulty, and then you’re attacking the system,” Bird said.
There is concern that without proper supervision, AI agents can take rigged actions that do not hold the owner responsible.
“The risk is that the leaderboard can help to keep decision makers in false sense of security,” Kozyrkov said. “Safety indicators are the starting point, not green light.”
