Apple Intelligence beta flagged a phishing email as "Priority"
Apple’s new “Apple Intelligence” features in the iOS Mail beta flagged a clear phishing email as “Priority” and surfaced it to the top of the inbox, raising concerns about using AI to curate security-sensitive content. Commenters note that spam filters have always been imperfect, but argue this kind of prioritization amplifies the impact of failures by lending scams extra credibility and hiding visual cues users rely on. The incident fuels broader skepticism about deploying LLM-style systems in adversarial contexts, the push to inject “AI” into every product, and whether users can or should fully disable such features.
Beta quality vs expectations
- Many note this is pre-release software; bugs, even serious ones, are expected.
- Others argue “it’s a beta” shouldn’t mute criticism: reporting and public discussion are how issues get fixed.
- Some feel the incident is overblown clickbait; others say it’s a valid example of a dangerous failure mode.
Spam filtering vs. Apple Intelligence
- Unclear exactly how Apple Intelligence interfaces with spam filters.
- Likely: traditional spam/Junk classification happens first; the new “Priority” feature then ranks inbox mail.
- The problematic behavior is not just missing the phish, but explicitly elevating it as “Priority” and hiding key warning cues.
- Several commenters stress that this is fundamentally a spam-filter failure; the AI priority layer is a secondary contributor.
LLM limitations and adversarial use
- Many emphasize that LLMs are probabilistic, error-prone, and “gullible,” especially with adversarial input.
- Attackers can iteratively test emails against the same models and tune them until they pass as important/legitimate.
- Some argue that, in adversarial domains like phishing, there is currently no safe way to rely on LLMs.
User experience and safety risk
- Concern that an AI-prioritized inbox can actively increase harm by pushing scams to the top and making them look official.
- Comparison to a human assistant saying “this is important, act on it,” thereby overriding users’ natural suspicion.
- Debate over whether AI systems should ever second‑guess spam filters or prioritize untrusted content.
Comparisons to Gmail and others
- Mixed experiences: some say Gmail spam/“important” filters are excellent; others report lots of spam and false positives.
- Several note that Gmail and others already mislabel phishing as “important,” so Apple is not unique.
Control, opt‑out, and product direction
- Apple Intelligence features can be disabled globally and per app; some are relieved, others expect this to be user‑hostile in practice.
- Broader criticism that companies are forcing AI into working systems (like email) for marketing/Wall Street reasons, not user need.
- Worry that email will be turned into an algorithmic feed, increasing opacity and potential for manipulation.