Apple uses AI to build its foldable iPhone
Its 100-part hinge is calibrated in the factory, but long-term durability remains unproven.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:003:26
Read transcript
Apple’s first foldable iPhone makes factory AI part of its durability pitch. The iPhone Duo’s hinge contains more than 100 components, controlling the opening and closing motion, supporting the center of the display, and helping the screen stay flat when open. Apple says algorithms match each hinge with the housing that fits it best. Then a confocal laser scans every unit’s surface topology, and a 3D printer adds as many as 25 micro-layers of custom photopolymer to reduce residual waviness. That is a notable use of AI: not a feature the customer operates, but a production system managing variation in a highly demanding physical component. Apple pairs the hinge process with high-strength glass, custom adhesives meant to relieve bending stress, multi-layer lamination, and a nano-textured polymer cover it says is up to 40 percent stiffer than other industry materials. Those interventions support the company’s argument that the Duo is engineered for repeated folding. They do not, however, prove long-term reliability. Foldables face stresses conventional phones avoid, and that test can only happen through extended use. Preorders begin October 16, with availability October 23. The larger point is that AI is moving deeper into the product stack—from software interfaces into manufacturing decisions and quality control. That same shift toward embedded intelligence appears on the wrist. Apple Watch Series 12 and Ultra 4 add Live Rewind, which shows up to 15 seconds of a previous conversation as text after a double press of the Digital Crown. Siri Recap takes a longer view, using ambient listening to generate a title and key points after a conversation. Apple says these Audio Intelligence features do not create or store audio recordings. The software arrives with watchOS 27 and will also reach compatible older Watches, so the change is broader than the new hardware. The practical question is whether convenience outweighs the privacy and accuracy questions that come with ambient capture. The economic version of that question comes from Anthropic’s interactive scenario explorer. Users can vary assumptions about AI capability, adoption, autonomy, and productivity, then see modeled U.S. outcomes for 2030. In its substantial scenario, GDP is 8.3 percent above a no-AI baseline, while labor receives 56.1 percent of output and knowledge-worker wages are essentially flat. Its extreme scenario lifts GDP 32.4 percent, but labor’s share falls to 45.2 percent and knowledge-worker wages drop more than 10 percent. Anthropic presents this as a conditional map, not a forecast. The useful distinction is between AI that augments workers and AI that performs knowledge work autonomously at speed: more output does not guarantee a broadly shared gain. And in security, OpenAI is applying the same automation logic to the repair loop. Its Defense Factory uses Codex, isolated environments, and engineering tools to investigate vulnerabilities, prepare tested patches, and verify that fixes reach production. OpenAI says an internal sprint involved more than 250 people across more than 100 service areas, with a 0.81 percent false-positive rate after runtime validation. But merged patches were not always deployed, which is precisely why people still review consequential changes and independently check fixes in production. Across these stories, the thing to watch is not simply whether AI can produce an answer or a patch. It is whether the surrounding system—manufacturing controls, privacy boundaries, labor institutions, or deployment checks—can reliably absorb what the models do.
