Study: Apple Intelligence susceptible to prejudice +++ Apple is researching AI-generated film scoring | News
Since the introduction of Apple Intelligence-powered summaries in the Notification Center, the italic texts in macOS and iOS have with some regularity caused amusement or head-scratching: In most cases, the local Large Language Model (LLM) gets to the heart of the content of collected posts; Occasionally, however, the short version is so far removed from the original content that one could speak of character assassination. A research group has now examined how Apple Intelligence reacts to inaccurate wording when it comes to origin and gender. They found that the local AI summaries replicate certain biases – more so than the competition. For this purpose, the EU-based research team “AI Forensics” formulated English-language texts, which were then summarized using local Apple Intelligence LLMs on an Apple computer with macOS 26.1. They noticed a trend when it came to mentioning origins: If the original text described a person as white, this detail only appeared in 53 percent of the summaries. If the original text describes them as black, the probability rose to 64 percent. For “Hispanic” or “Asian,” the origin appeared in over 80 percent of the short versions.
Trick questions about the job
In further experiments, the researchers deliberately formulated ambiguous texts, for example about a disagreement between a doctor and a nurse. The next sentence claimed in one case that he was right and in another case that she was right. In 77 percent of all cases, Apple Intelligence chose a profession based on gender, in line with prejudices: in 67 percent of cases, “she” was the nurse, “he” was the doctor. The result is particularly unpleasant when compared with the competition: the local Gemma3-1B-LLM, which is only a third as large, only hallucinated assignments in 6 percent of cases, in contrast to 15 percent for Apple Intelligence.
AI creates audio track for video
Apple has already shown in the past how intensively the company is working on realistic applications of AI in the field of video. In collaboration with the Renmin University of China, the company’s own researchers are now publishing a publication – and putting the AI model online for you to try out. In this case it is about the subsequent recording of video material. VSSFlow is a multimodal generative model that generates both authentic ambient sounds and lip-synchronized speech. Users provide the silent film and the transcript, the AI creates an audio track.
Two tasks in one
The researchers emphasize that VSSFlow leaves other models behind in established benchmark tests, and in two categories: Normally, video-to-sound (V2S) and visual text-to-speech (VisualTTS) are two separate tasks. On a project page, the research team has put together some examples that present the results of VSSFlow in comparison to other AI-generated dubbing.
