Three corrections moved AI trust from product claims into evidence. Meta withdrew a likeness-based image feature after its opt-out design drew criticism, Apple alleged that former employees carried hardware secrets to OpenAI, and OpenAI found that a widely cited coding benchmark contained a large share of broken tasks.
1. Meta removes an image feature built around public Instagram accounts
Meta discontinued a Muse Image feature that let people generate pictures by mentioning public Instagram accounts. The company had offered an opt-out under Sharing and Reuse settings, but critics argued that public posts and likenesses should not become image-generation inputs by default. Meta said the feature "missed the mark" and removed it days after launch.
The broader set of Instagram AI effects remains available, making this a targeted rollback instead of a retreat from generated media. An account-level opt-out failed to resolve consent, impersonation, and endorsement risk at creation time. Product permission and a person's reasonable expectation for reuse proved to be different standards.
Sources: AP on Meta's removal of the Instagram image feature
2. Apple alleges stolen secrets accelerated OpenAI's hardware work
Apple sued OpenAI, io Products, and two former Apple employees in California federal court, alleging a coordinated theft of confidential hardware information. The complaint names OpenAI hardware chief Tang Yew Tan and engineer Chang Liu and covers product designs, manufacturing processes, supplier information, and files allegedly taken from Apple's internal systems.
The claims are allegations, and the defendants had not filed a court response when the reports appeared. The case turns talent migration into a provenance problem: experienced engineers can carry general knowledge, but protected documents and components remain separate assets. Hardware programs gain a defensible boundary from source records, clean-room procedures, and auditable material handling.
Sources: AP on Apple's trade-secret lawsuit
3. OpenAI finds roughly 30% of SWE-Bench Pro tasks are broken
OpenAI audited SWE-Bench Pro after frontier-model pass rates on its 731-task public split rose from 23.3% to 80.3% in eight months. Its agent-assisted pipeline identified 200 broken tasks, or 27.4%, while five-engineer reviews identified 249, or 34.1%. Problems included overly strict tests, underspecified or misleading prompts, and tests with insufficient coverage.
The company now estimates that roughly 30% of tasks are broken and has retracted its earlier recommendation to adopt the benchmark. OpenAI is an interested model developer, but the paired human review and published failure taxonomy make the criticism testable. Coding scores built from repository history become credible only after task-level audits expose ambiguous prompts and invalid tests.
Sources: OpenAI's SWE-Bench Pro audit