Called Azure AI Content Safety, this new product is available through the Azure AI Products platform and offers a range of AI models trained to detect “inappropriate” content across images and text. TechCrunch reports that the model can understand text in English, Spanish, German, French, Japanese, Portuguese, Italian, and Chinese, and assigns a severity score to flagged content to determine which Indicate to moderators if content needs attention.
“Microsoft has been working for more than two years on solutions that address the challenges of harmful content presented in online communities. We were aware of it,” a Microsoft spokesperson said in an email. “The new (AI) models are now able to better understand content and cultural context. This will help users understand why it was deleted or deleted.”
In a demo at Microsoft’s annual Build conference, Microsoft’s responsible AI lead, Sarah Bird, said that Azure AI Content Safety will power Microsoft chatbots on Bing and Copilot, GitHub’s AI-powered code generation services. I explained that it was the production version of the system.
“We are currently launching this as a product that our third-party customers can use,” Byrd said in a statement.
Perhaps the technology behind Azure AI Content Safety has improved since it was first released for Bing Chat in early February. Bing Chat didn’t take off when it first rolled out in preview. Our interviews revealed that the chatbot spits out misinformation about vaccines and writes hateful articles from the point of view of Adolf Hitler. Other reporters used it to intimidate him or even humiliate him for admonishing him.
In another blow to Microsoft, the company just months ago laid off the ethics and social team within its larger AI organization. With this move, Microsoft no longer has a dedicated team to ensure AI principles are tightly coupled to product design.
Putting all this aside for a moment, Azure AI Content Safety, which protects against biased, sexist, racist, hateful, violent and self-harming content, according to Microsoft, is Microsoft’s fully managed enterprise offering. Integrated into one Azure OpenAI Service. It aims to give enterprises access to OpenAI’s technology with added governance and compliance capabilities. However, Azure AI Content Safety can also be applied to non-AI systems such as online communities and game platforms.
Pricing starts at $1.50 per 1,000 images and $0.75 per 1,000 text records.
Azure AI Content Safety is similar to other AI-powered harm detection services such as Perspective and Jigsaw, maintained by Google’s anti-abuse technology team, and is the successor to Microsoft’s own Content Moderator tool. (No word on whether it’s built on Microsoft’s acquisition of moderation content provider Two Hat in 2021.) These services, such as Azure AI Content Safety, will help new comments and Provides a score from 0 to 100 on how similar an image is to another. It was previously confirmed to be toxic.
But there are reasons to doubt them. Beyond Bing Chat’s early stumbling blocks and Microsoft’s poorly targeted staff cuts, it’s clear that AI toxicity detection techniques are still struggling to overcome challenges such as bias against certain subsets of users. shown in research.
A few years ago, a Pennsylvania State University team found that commonly used public sentiment and toxicity detection models could flag posts on social media about people with disabilities as more negative or harmful. I discovered that I have a gender. In another study, researchers found that older versions of Perspective often failed to recognize “reused” slurs, such as “queer,” and hate speech that used spelling variations, such as missing letters. was shown.
This problem extends beyond toxicity detectors as a service. This week, a New York Times report reveals that eight years after the controversy over black people being misclassified as gorillas by image analysis software, the tech giant still fears repeating the mistake. bottom.
One of the reasons for these failures is that the annotators (the people responsible for adding labels to the training datasets that serve as examples for the model) bring their own biases.
To address some of these issues, Microsoft allows Azure AI Content Safety filters to be contextually fine-tuned. Bird explains:
For example, the phrase “attack over the hill” used in the game is considered medium violence and will not be blocked if the game system is configured to block medium severity content. increase. Adjusting to accept a medium level of violence makes the model more tolerant of this phrase.
“We have a team of language and fairness experts who take culture, language and context into account when defining our guidelines,” a Microsoft spokesperson added. “Then we trained an AI model to reflect these guidelines… AI always makes some mistakes (although). We recommend using human-involved tools to
One of the early adopters of Azure AI Content Safety is Koo, a blogging platform based in Bangalore, India, with a user base that speaks over 20 languages. Microsoft said it is partnering with Koo to tackle moderation challenges, such as analyzing memes and learning colloquial nuances of languages other than English.
There was no opportunity to test Azure AI Content Safety prior to release, and Microsoft did not respond to questions regarding annotations or bias mitigation approaches. However, rest assured that we will be closely monitoring how Azure AI Content Safety works in practice.
