...
By Thomas Westerholm
Life & Trends Reporter
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.
A group of AI researchers took a closer look at Apple Intelligence's design, and what they found offered a concerning look at bias written into the tech giant's AI model.
AI Forensics, a Paris-based nonprofit, investigates the code behind algorithms in an effort to "hold major technology platforms accountable" with technical assessments designed to expose "harms caused by their algorithms."
In a recent study conducted by the group and shared with EUObserver, researchers found that when Apple Intelligence summarized users' messages, it showed "two distinct patterns" of racial and gender bias.
Read More on Tech
...
Why It Matters
Apple Intelligence is advertised as an "everyday" assistant for Apple users, built into iPhones, iPads, MacBooks and numerous other Apple products.
Among a number of other uses, Apple says its AI model can help write emails, summarize notes and simplify "your everyday communication" while it "helps you stay focused."
Users who don't want to use AI tools can turn off Apple Intelligence, but they need to consciously reject it when their device updates.
According to the researchers, EU regulations say that Apple Intelligence requires transparency documentation due to its computing power, and Apple has not signed a voluntary document that lists "discriminatory bias" as a potential risk for AI.
What To Know
AI Forensics ran thousands of tests on Apple Intelligence to reach its results.
In one example of racial bias the group shared with EUObserver, the model treated "white" as the default for the protagonist of a story—in summaries, Apple Intelligence was far more likely to mention an individual's ethnicity if they were Asian, Black or Latino.
In another example, the researchers showed that when they mentioned two patients at a doctor's office, one gay and one straight, the model assumed the gay patient had a sexually transmitted disease (STD), even though the prompt did not say which patient had one.
Researchers also shared a prompt about a man and a woman, one of whom was a doctor and the other a nurse.
The prompt did not say the man was the doctor, but once again, the model followed biases and offered a summary that assumed he was.
According to the researchers, the system hallucinated (when the AI invents false information) 15 percent of the time. When it hallucinated, it followed stereotypes 72 percent of the time.
“Associations between gender and occupation are widely present in online text and therefore in training data," study author Paul Bouchaud told Newsweek in an email.
"Previous academic work has shown that workforce statistics are mirrored, for example, in Common Crawl, on which numerous LLMs are trained."
What's Next
Apple may face financial pressures from EU regulators later this year.
In the EU, a new AI office will gain enforcement abilities in August. That office will have the power to levy significant fines against companies that violate AI laws.
The researchers told EUObserver that they want the "reach" of Apple Intelligence to be considered by regulators, even if Apple finds a workaround for its computational power.
Newsweek has reached out to Apple for comment via email.
To read how Newsweek uses AI as a newsroom tool, click here.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines