Buried in a discovery filing for OpenAI v. SpaceXAI is a phrase that should make every enterprise architect who bet on “Apple Intelligence extensibility” sit up: the ChatGPT integration in Siri, OpenAI’s own internal metrics show, has “persistently underperformed” since it launched in December 2024. No percentages. No user counts. Just that word — persistently — which is corporate-speak for “we kept waiting for the hockey stick and it never came.”
The easy read is a trust story. Users don’t want their assistant quietly shipping requests to a third party, the argument goes, so they opted out. It’s a tidy narrative. It’s also mostly wrong. The more honest explanation is duller and more useful: Apple designed the handoff to be forgettable, and it succeeded. Underperformance wasn’t a verdict on ChatGPT. It was the predictable output of the plumbing.
What did OpenAI actually say — and what didn’t it say?
The load-bearing claim here is one word doing a lot of work. “Persistently underperformed” appears in a filing about a commercial dispute, not a product retrospective, which matters for how much weight it can hold. OpenAI has an incentive in that proceeding to characterise the Apple partnership a particular way. The filing, as reported, gives no adoption figures, no monthly-active anything, no baseline for what “expected” performance even was. So resist anyone — including me — who tries to hang a number on it.
Here’s the factual spine, and I’ll be precise about where each piece comes from. Per Apple’s own release notes for those updates, the integration shipped in iOS 18.2, iPadOS 18.2 and macOS 15.2 in December 2024. On my own test devices running the current September 2026 releases, the handoff is still present. And per the reporting on the filing, OpenAI’s telemetry since launch disappointed the people who built it. Everything past that spine is inference, and I’ll flag mine as I go.
Was this ever going to be a hit?
Think about how you actually reach ChatGPT through Siri. You ask Siri something. Siri decides — or you tell it — that this is a “world knowledge” question better handed off. Then, by default, a card appears asking whether you’d like to send the request to ChatGPT. You tap yes. Then the answer comes back.
That confirmation step isn’t a bug. On a test device running the current iOS, it’s governed by the Confirm ChatGPT Requests toggle, on by default, under Settings → Apple Intelligence & Siri → ChatGPT. (Names and paths here are what I see on-device as of September 2026; Apple has been known to reshuffle these labels between point releases, so treat them as current rather than eternal.) Apple built a doorbell into the handoff on purpose.
Here’s the analogy I keep coming back to. Imagine hiring a translator, but every single time you want them to speak, you first have to summon them by name and then confirm — out loud, per sentence — that yes, you really do want them to translate this one. Most people would just… learn the other language a bit, or give up. The friction isn’t hostile. It’s just enough to keep the feature from becoming a habit, and features that don’t become habits don’t compound.
Layer on the alternatives. Anyone who genuinely wants ChatGPT has the actual ChatGPT app one tap away, with memory, history, voice mode and the model they chose. The Siri bridge, by contrast, is stateless, capped, and interrupts you to ask permission. It was never the destination. It was a courtesy ramp. Courtesy ramps don’t post big adoption numbers, and I’d argue nobody sober ever expected them to.
What does ChatGPT actually see when Siri calls it?
This is the question that decides whether the feature is deployable in a regulated fleet, and it’s where the design is genuinely clever. What follows comes from Apple’s published privacy documentation for the ChatGPT integration — its Apple Intelligence and privacy material and the on-device disclosure Apple shows when you first enable ChatGPT — not from my own packet capture, so read the retention claims as documented promises rather than something I’ve independently proven on the wire.
When you use the handoff without connecting a ChatGPT account — the default state — Apple positions itself as a privacy broker between you and OpenAI. Apple states that requests are relayed so your IP address is obscured from OpenAI, and that under the partnership terms OpenAI does not store those requests or use them for training. Critically for identity people, Apple’s disclosure says OpenAI does not receive your Apple ID, your Managed Apple ID, the device serial or UDID, or any MDM-supplied enrolment identity. From OpenAI’s side the request arrives effectively anonymous. That’s a real architectural boundary described in Apple’s documentation, not just a policy line — but the no-retention half of it is a contractual commitment you’re trusting, not something an admin can verify from outside.
The moment a user signs in with their personal ChatGPT account, Apple’s own documentation says that boundary changes character. Now the request is associated with their OpenAI identity, subject to OpenAI’s normal data and retention policies, and Apple notes that its IP-obfuscation protection no longer applies in the same way. Still no Apple ID or device identifier crosses over — the link is the user’s own ChatGPT credentials, which they supplied deliberately. But you’ve traded Apple’s anonymising relay for a named session under a consumer contract your organisation never signed.
For a corporate device that distinction is the whole ballgame. Anonymous relay: annoying, low-risk. Personal-account sign-in on a managed device: an unsanctioned data path where employees route who-knows-what through a personal OpenAI account tied to corporate context. That second scenario is the one you actually need to govern.
Can you turn it off across a fleet?
Yes — and you’ve been able to since the day it shipped, which is the part the “will enterprises trust this?” framing tends to miss. Apple gave admins the kill switch in the same release as the feature.
On supervised devices, the restriction payload com.apple.applicationaccess carries two keys. Both are documented in Apple’s device management restrictions reference, and I’ve confirmed the behaviour on supervised hardware in my own MDM testing:
allowExternalIntelligenceIntegrations— set tofalseto remove the ChatGPT (and any future third-party model) extension entirely. No handoff, no card, no doorbell.allowExternalIntelligenceIntegrationsSignIn— set tofalseto keep the anonymous relay available but block users from attaching a personal account.
That second key is the surgical one. You can let staff use the anonymised, non-retained relay for genuinely useful world-knowledge questions while forbidding the personal-account path that creates the real data-governance headache. A profile that does exactly that:
One correction to a common assumption: this is enforced through MDM and supervision, not through a toggle in the Apple Business Manager console. ABM manages your Managed Apple IDs and device assignments; there’s no “disable third-party AI” checkbox in it. The control lives in the restrictions payload your MDM pushes, and — again, from my testing — the keys only take effect on supervised hardware. BYOD devices enrolled via User Enrolment with a Managed Apple ID sit in a data-separated world where you don’t get this lever at all — another reason the personal-account path deserves your attention.
To sanity-check that the restriction actually landed on a Mac in the fleet:
So what actually failed?
Not user trust — or at least, the filing gives you no evidence for that. What “persistently underperformed” documents is a feature that was optional, gated behind a confirmation tap, stateless, and permanently outclassed by the standalone app it was pointing at. Given those constraints, low adoption isn’t a surprise. It’s the spec working as written.
For anyone building a deployment strategy on Apple Intelligence extensibility, the lesson isn’t “third-party AI in Siri is doomed.” It’s that Apple ships these bridges as low-commitment, high-friction, easy-to-govern add-ons — and treats them as removable from day one. That’s genuinely good news for the security team and a cold shower for anyone who assumed the ChatGPT partnership was a strategic on-ramp rather than a demo that happened to ship.
The handoff still works on my test devices in September 2026. Almost nobody reaches for it. Apple built it that way, and the numbers just caught up to the design.