Nevertheless, it is a serious problem when summaries misrepresent headlines, and edge cases where this happens are unfortunately unavoidable. Apple can't just fix these abstracts with a software update. The only answers are to help users understand the downsides of the technology so they can make more informed decisions, or to remove or disable the feature entirely. Apple apparently goes for the former.
We're oversimplifying it a bit here, but in general LLMs, like those used for Apple's notification summaries, work by predicting parts of words based on what came before them and are unable to predict the content they contain. really understand summarizing.
Furthermore, these predictions are known to not always be accurate, with incorrect results occurring a few times per 100 or 1,000 outputs. As the models are trained and improvements are made, the error rate can be reduced, but it will never reach zero if numerous summaries are created every day.
Deploying this technology at scale without users (or even the BBC, it seems) really understanding how it works is risky at best, whether it's the iPhone's summaries of news headlines in notifications or the AI Google summaries at the top of search engine results pages. Even if the vast majority of summaries are completely accurate, there will always be users who see inaccurate information.
These summaries are read by so many millions of people that the magnitude of errors will always remain a problem, no matter how relatively accurate the models are.
We wrote extensively a few weeks ago about how the rollout of Apple Intelligence seemed rushed, at odds with Apple's usual focus on quality and user experience. However, with current technology, there is no refinement to this feature that Apple could have made to achieve a zero percent error rate with these notification summaries.
We'll see how well Apple manages to get its users to understand that the summaries may not be accurate, but really letting all iPhone users know how and why the feature works this way would be a tall order.
