Testing the Apple Intelligence beta shows the prerelease software still has a long way to go, with hallucinations, marking scam emails as priority, and more
Geoffrey A. Fowler / Washington Post :
Context & Ripple Effects
Apple introduced Apple Intelligence in developer betas with text-generation and Siri changes while leaving some capabilities absent. Its own earlier prompt instructions explicitly emphasized avoiding fabricated information, making the reported hallucinations a direct test of whether those safeguards work in practice the initial developer-beta rollout the system's anti-hallucination instructions.
The report arrives as Apple Intelligence moves through successive beta releases. Later hands-on coverage also characterized iOS 18.1's experience as uneven, suggesting the issue is not confined to a single isolated output type the first public-beta release.
First-order effects
- Beta users can receive unreliable generated information and misleading mail prioritization, including scam messages elevated as important.
- Apple must improve model behavior and the priority-classification experience before those features can be treated as dependable parts of its software.
Second-order effects
- Incorrect prioritization makes users more likely to verify AI-ranked inbox items manually, reducing the immediate convenience of an assistant meant to surface what matters.
- The failures raise the quality bar for Apple’s staged beta rollout: features can be shipped separately, but user-facing trust depends on their reliability working together.
Third-order effects
- If comparable failures persist across consumer AI assistants, product competition will shift from feature breadth toward dependable task routing, provenance, and controls for high-consequence recommendations.
- Email triage illustrates a broader constraint on AI embedded in operating systems: mistakes are more damaging when the assistant chooses what the user sees first, rather than merely drafting optional text.
The trend: Consumer AI is moving from standalone generation into operating-system workflows, where trust hinges on reliable prioritization as much as fluent output.