Apple's integration of ChatGPT was "dramatically underperforming" and "persistently underperforming," according to OpenAI statements in court filings. The statements were made in the company's ongoing dispute with Elon Musk's xAI, now known as SpaceXAI. Apple added the ChatGPT option to Siri in December 2024, yet OpenAI now describes its competitive effect as minimal. For companies deploying AI assistants, the episode shows that placement inside a major operating system does not guarantee adoption.
ChatGPT opt-in inside Apple Intelligence
Apple introduced the ChatGPT link in December 2024 as an extension to Apple Intelligence, not as a default engine. Users had to go through a multi-step opt-in process before Siri could call on ChatGPT for more demanding requests. One visible use was Visual Intelligence, where ChatGPT could add context to images shot with an iPhone camera. The filings discussing the performance are heavily redacted, so usage volumes and internal targets remain undisclosed. Meanwhile, Google's Gemini now powers much of what Siri AI can do, which reframes the ChatGPT tie-in as one of several external models rather than the core.
The design kept Apple Intelligence as the first layer and offered ChatGPT as an escalation path when a request needed more power. That structure preserved user choice but added friction at the exact moment of use, since the option had to be found and enabled in advance. OpenAI's filing does not detail click-through or retention, yet its language points to low activation rather than technical failure. The company, through expert Dr. Catherine Tucker, calculated the share of generative AI consumers who reached ChatGPT through Apple Intelligence and found it consistent with minimal usage. In other words, the bottleneck was entry, not capability.
The disclosure appeared in litigation where xAI accused OpenAI of harming competition, and OpenAI responded that the Apple channel had an "indisputably de minimis" effect on rivals. That framing matters because the integration was once seen as privileged distribution on a very large base of devices. Early Apple Intelligence was viewed as incomplete and behind competing assistants, which likely shaped first impressions and reduced willingness to try add-ons. If users distrusted the base system, an optional extension had little chance to prove itself. The result leaves open whether low ChatGPT use reflected Siri limits, setup friction, or preference for the standalone ChatGPT app.
What poor OS integration means for business
For businesses, the lesson concerns deployment of AI copilots and agents inside existing workflows. An assistant embedded in a phone, CRM, or service desk only creates value if employees encounter it without extra steps and permissions. A small firm that relies on standard iPhones or off-the-shelf tools feels this directly: one additional opt-in can cut regular use to near zero. A larger organization with managed devices and central configuration can pre-enable integrations, set defaults, and train staff on when escalation to a stronger model helps. The difference between available and actually used becomes the metric to track in pilots.
The filing also shows what buyers should verify before counting on platform distribution. Much of the evidence is redacted, and no public activation rate, retention curve, or task success data was disclosed. That means the news alone does not prove that users rejected ChatGPT, only that the Apple Intelligence route delivered little traffic. When evaluating vendors, ask what share of requests flows through each model, how opt-in and consent screens affect completion, and where data goes when a query escalates. Test the standalone app against the embedded path on the same tasks, from drafting and summarization to image explanation.
The marker to watch is how Apple positions model choice in the rebuilt Siri AI and whether ChatGPT remains a visible, low-friction option alongside Gemini. If future disclosures or court records show rising use through Siri or a changed default, the earlier failure will look like an interface problem. If traffic stays minimal, it will point to a lasting preference for separate apps. That split will tell buyers whether to invest in OS-embedded paths or in standalone deployments.
