Craig Federighi stood on stage at WWDC 2026 and drew a line most of the industry didn’t know was there to be drawn. Siri AI, he said, is not a companion. It does not simulate friendship. It does not seek emotional connection. It completes tasks and gets out of the way. You probably filed that under “Apple being Apple” and moved on. But there is more structure underneath that statement, and that structure has consequences worth mapping, especially once you factor in the Digital Markets Act keeping Siri AI off 450 million EU iPhones.
By the end of this article you will understand that Apple’s anti-companion positioning is neither pure philosophy nor pure spin. It is architecture, brand protection, and competitive framing operating simultaneously. And the DMA lockout turns this strategic choice from a market differentiator into something more complicated: a vulnerability that creates real build-or-wait decisions for your engineering teams.
Why is Apple drawing a hard line against AI companionship — and does the reasoning hold up?
The reasoning operates across three layers, and you need all three to see what Apple is actually doing.
The technical layer is real. Apple’s three-tier privacy architecture — on-device processing for about 85% of queries, Private Cloud Compute for roughly 12%, and partner models like ChatGPT for the remaining 3% — enforces statelessness by design. PCC servers use cryptographic attestation, OHTTP relays, and RSA Blind Signatures that make data accumulation across sessions architecturally impossible. You cannot build persistent memory on infrastructure engineered to forget. Companion AI requires remembering who the user is. Siri AI was built to do the opposite.
The brand layer is about liability maths. Apple’s trust premium is its competitive moat, and an AI that simulated friendship would create exactly the kind of trust liability the company cannot afford. The $250 million Siri settlement over false advertising was a lesson in what happens when you overpromise AI capabilities. A companion that got a relationship wrong? That’s a brand threat with consequences the company has already seen play out.
The competitive layer is where the positioning gets interesting. Replika has millions of paying users who have formed emotional bonds with AI, and Character.AI has attracted tens of millions of users with proprietary conversation models tuned for relationship simulation. These companies have spent years optimising for emotional engagement in ways Apple’s architecture was never designed to support. OpenAI’s GPT-4o demonstrated exactly what happens when engagement optimisation goes too far — a sycophancy episode where the model leaned too heavily on short-term user satisfaction, got rolled back after backlash, and gave everyone vocabulary for what felt off. Apple converts a competitive gap into a deliberate choice. You cannot lose a race you chose not to enter.
The stance has architectural foundations and competitive framing operating in the same breath. The architecture really does prevent companion features. But the public articulation selectively emphasises what Apple gains — trust differentiation — while minimising what it loses: a companionship market that is commercially viable and growing.
Understanding why Apple chose this stance is necessary, but the more consequential question is what it looks like set against the rest of the market.
How does Siri AI’s anti-companion design compare to ChatGPT, Claude, and Gemini in practice?
Apple is not the only player in the assistant market, and the comparison reveals that every other major option invites the relational engagement Apple is refusing.
ChatGPT’s Advanced Voice Mode creates conversational presence users anthropomorphise and bond with. OpenAI does not market it as a companion, but tone mirroring, conversational memory, and emotional responsiveness encourage relational use. When OpenAI tried replacing GPT-4o with GPT-5 in August 2025, the backlash from users attached to its warm conversational style forced a reversal. Companionship in effect, if not in name.
Claude markets itself on thoughtfulness and intellectual partnership. Anthropic’s framing is “helpful, honest, and harmless,” not companionship. But Claude’s Projects feature builds persistent context across sessions, and the depth of its conversational reasoning — asking clarifying questions, pushing back on assumptions, referencing earlier exchanges — creates a form of intellectual rapport that, over repeated use, looks a lot like a relationship. The distinction between “helpful partner” and “relationally present” gets thin fast.
Gemini is positioned as a Google-integrated helper without explicit companionship framing and without the anti-companion line Federighi drew. Its ambient integration — reading screens, parsing Gmail, anticipating needs — builds habituation. The International AI Safety Report 2026 notes that individuals can unintentionally form relationships with non-companion AI systems through productivity-focused interactions. The behavioural outcome looks similar even if the intent is different.
Siri AI is the counterpoint: stateless, task-bounded, privacy-enforced exit. No conversational memory beyond the session, no emotional modelling, no persistence of identity. It is the only major assistant whose executive leadership has publicly stated “this is not a companion.” Whether that distinction holds in practice once millions of people talk to it daily is a question the market will answer.
How does Apple’s Siri AI architecture compare to Google’s native Gemini rollout on Android?
The comparison gets sharper when you look at platform architecture. Google’s Gemini on Android is natively integrated, replacing Google Assistant. It is available in the EU. It runs on Google’s own models with no licensing dependency. Its privacy guarantees are policy-based — governed by terms of service and compliance certifications rather than hardware-enforced cryptographic constraints.
Apple’s Siri AI, under the three-tier architecture described above, is partially dependent on Google infrastructure. Apple Foundation Models were co-developed with Google, the PCC infrastructure extends to Google Cloud NVIDIA GPUs, and the licensing deal is reported at roughly $1 billion per year. And the entire stack is blocked on EU iPhones and iPads.
The strategic asymmetry turns on platform control, not model capability. Google competes on capability breadth — Gemini does more things across more surfaces. Apple competes on privacy depth and integration trust. The DMA lockout prevents Apple from competing on either axis for 450 million users. Meanwhile, Google profits from enabling Apple’s AI while competing against it with a product available in markets Apple’s cannot reach. Samsung’s Galaxy AI features already use Google Gemini models, meaning the Apple deal eliminates Samsung’s differentiation on that front. Both platforms now run on the same underlying AI engine, differing only in integration layer and privacy architecture. The DMA lockout makes that asymmetry permanent rather than temporary.
What are the trust-model differences between Private Cloud Compute, Google Gemini, and Microsoft Copilot?
For your procurement decisions, the trust-model spectrum matters. Apple’s PCC uses Apple Silicon servers with cryptographic attestation enforcing stateless processing. No data retention, no logging, no user-identity linking across requests. Independent researchers can inspect PCC binaries. A WiSec ’26 academic paper reverse-engineered the implementation and found protocol deviations — TGT signature validation gaps and OTT reuse windows — but the core privacy claims were validated.
Google Gemini processes on Google’s global cloud infrastructure with policy-based guarantees. The privacy posture is governed by terms of service, compliance certifications including SOC reports and ISO 27001, and Google’s published data processing terms for enterprise customers. The audit trail is available through Google Cloud’s compliance reports portal, but the guarantees are contractual and policy-based, not enforced by hardware-level cryptographic constraints. For enterprise procurement, the distinction matters: you are trusting Google’s governance processes, not independently verifiable server code.
Microsoft Copilot operates on Azure OpenAI enterprise infrastructure with contractual data residency commitments, admin controls, and compliance certifications like ISO 27001 and SOC 2, but without Apple’s statelessness guarantees. Data is persistently processed in cloud with enterprise tenant isolation — your data stays in your tenant, but it persists.
For EU-based enterprises, the procurement dimension is verifiability. Apple’s attestation is independently inspectable, with identified limitations. Google’s is audit-reported. Microsoft’s is contractually governed. Data residency and data sovereignty address different questions: where the data physically sits versus who has legal jurisdiction over it. That distinction becomes acute when PCC processes in Houston, Texas, and queries routed to US servers trigger GDPR cross-border transfer considerations.
What does the Siri AI EU lockout mean for the 450 million EU residents and developers?
The lockout makes AI assistant availability a structural platform risk.
EU iPhone and iPad users are excluded from Siri AI when iOS 27 launches. Apple has stated there is “no timeline” for resolution. The countryd and eligibilityd device attestation system — using SIM, GPS, and Wi-Fi signals since iOS 17.4 — prevents VPN-based workarounds. As one analysis put it, early testers report the only combination that reliably works is a non-EU Apple Account plus the phone physically being outside the EU, at which point you have not found a workaround, you have found travel.
macOS and visionOS are exempt from DMA gatekeeper designation and will receive Siri AI in the EU. But Vision Pro’s price point makes it an impractical enterprise workaround, and the Mac exemption creates platform fragmentation within organisations. The same employee gets full Siri AI on a MacBook and zero access on the iPhone in their pocket.
For developers, the impact is concrete. Siri AI does not run on EU iPhones, so teams cannot verify integration behaviour on EU devices. Siri Extensions, the iOS 27 developer framework enabling third-party AI integration, is also blocked in the EU. Teams must either provision non-EU test devices, creating hardware-procurement overhead, or use simulator environments that do not replicate real Siri AI behaviour. That friction cost makes Siri AI integration less attractive for any team with meaningful EU exposure.
These developer realities feed directly into the strategic decisions your team now faces.
What should you consider when evaluating AI assistant strategy for the EU market?
The lockout forces a structural decision, and the evaluation has several dimensions.
Platform availability is the starting point. If your user base is meaningfully EU, Siri AI is not an available platform today and there is no resolution timeline. Your strategy must either be platform-agnostic — build once, deploy to Siri AI SDK plus Google Assistant plus custom — or Android-first, where Gemini is available in the EU without regulatory constraint. The visionOS split adds a wrinkle. Vision Pro Siri AI integrations can be built and tested, but the user base is negligible. It is a demo platform, not a deployment target.
Vendor lock-in needs attention. Committing to any single AI assistant ecosystem creates structural dependencies through OS-level integration, data formats, and procurement path dependency. The DMA lockout simultaneously makes lock-in worse, by delaying feature availability, and better, by forcing interoperability discussions. Apple introduced per-intent privacy manifest APIs at WWDC 2026 allowing developers to declare whether a Siri interaction routes to the cloud or stays on-device. For healthcare, legal, and financial applications where data is subject to regulatory restrictions, the per-intent declaration is a compliance dimension, not a user experience preference.
Data sovereignty rounds out the assessment. Queries routed to PCC in Houston, Google infrastructure, or OpenAI servers trigger GDPR cross-border transfer considerations. Data residency and data sovereignty address different questions: where a vendor is legally domiciled and who has legal access to data are separate from where the data physically sits.
Device procurement now has an AI dimension it didn’t have before. An MDM profile that treats Siri AI availability as a device attribute means your provisioning rules need to account for which devices get which AI capabilities — and what happens when an employee with a Mac that has Siri AI hands off work to someone on an EU iPhone that doesn’t. The platform split isn’t just a feature gap; it is a workflow consistency problem that shows up in device fleets.
Should EU-based teams build for Siri AI now — or wait for regulatory resolution?
There is no single answer. Using the four variables outlined above, your team needs to weigh the specifics of your situation.
User geography: what percentage of your user base is EU-based? If it is a majority, Siri AI is effectively unavailable. If it is a minority, the non-EU opportunity may justify investment.
Competitive pressure: are competitors shipping Siri AI integrations to non-EU users? If they are, waiting means ceding early-mover advantage in those markets. If nobody is, the cost of waiting drops.
Integration depth: is Siri AI integration core to your product, or a nice-to-have? The deeper the dependency, the higher the cost of waiting. An AI assistant app that depends on OS-level hooks faces a different calculus than a productivity app where AI is supplementary.
Regulatory trajectory: is the DMA dispute likely to resolve, escalate, or stall? The original Apple Intelligence delay took six months to resolve for EU users, and that dispute was simpler than this one. The ECIPE analysis of DMA Article 6(7) specification proceedings suggests the Commission’s intervention is shaping ecosystems, not resolving technical constraints. The reasonable planning assumption is that Siri AI will not be available on EU iPhones or iPads in the 2026 calendar year.
The visionOS exemption is worth acknowledging but not betting on: building for Vision Pro Siri AI lets your team learn the integration patterns, but the user base is a rounding error. The UK’s DMCC Act does not impose equivalent interoperability obligations, so UK-based teams can develop and deploy while EU counterparts cannot — a geography-based staffing dimension to factor in.
The right answer today may not be the right answer in six months, which is why regulatory trajectory is one of the four variables in the first place.
What do DMA interoperability obligations mean for app developers who are not gatekeepers?
Article 6(7) obligations apply to the platform gatekeeper — Apple and Google — not to the app developer building integrations. The DMA requires gatekeepers to provide third-party services “free and effective interoperability” with hardware and software features the gatekeeper’s own services access.
The practical risk for your team is strategic dependency, not your own compliance obligation. If your AI assistant integration depends on Siri AI-specific capabilities that are not available to competing assistants through the DMA interoperability portal, that integration becomes collateral in someone else’s dispute. Here is what that means: if the European Commission determines that a Siri AI capability your integration depends on should have been made available to competitors under Article 6(7), Apple might be ordered to restructure that capability. Your integration survives or breaks depending on how Apple complies. You are not the target of the proceeding, but your product is in the blast radius.
The Free Software Foundation Europe reported that 0 of 56 formal interoperability requests to Apple produced new solutions. That pattern suggests the lockout extends beyond Siri AI into a broader posture. The question to ask your team: if your integration becomes a regulatory exhibit, what is your fallback? The more tightly coupled it is to Apple-specific AI infrastructure, the more exposed it is to DMA-driven disruption.
Apple’s anti-companion stance operates as architecture constraining product, brand protecting market position, and competitive framing converting a gap into a choice — all at the same time. When the DMA lockout strips away the ability to deliver that architecture to 450 million users, the strategy that was meant to be a differentiator becomes something more exposed.
Google’s Gemini faces no equivalent EU constraint. Every other major assistant invites the relational engagement Apple refuses. The build-or-wait decision reduces to four variables — user geography, competitive pressure, integration depth, and regulatory trajectory — and none of them produce a universal answer. The DMA standoff has no clear endpoint.
That is the landscape your decisions have to operate within.
Frequently Asked Questions
When will Siri AI be available in the EU?
Apple has stated there is “no timeline” for resolution, and no formal specification proceeding has been initiated by the European Commission. The standoff turns on whether Private Cloud Compute’s cryptographic attestation satisfies DMA Article 6(7) interoperability requirements. Until Apple, the EC, or both change their posture, the lockout persists indefinitely. The most realistic planning assumption for EU-based teams is that Siri AI will not be available on EU iPhones or iPads in the 2026 calendar year, and possibly longer.
Can EU users access Siri AI through a VPN?
No. Since iOS 17.4, Apple has deployed a device attestation system (countryd and eligibilityd daemons) that determines device eligibility using SIM card data, GPS coordinates, and visible Wi-Fi network metadata. A VPN masks your IP address but does not override these hardware and radio-level signals. The system is designed to resist workarounds, and no reliable bypass has been documented. EU iPhones and iPads are locked out regardless of network configuration.
Is Apple’s anti-companion stance just marketing?
No, but it is not purely engineering either. The three-tier privacy architecture (on-device processing, Private Cloud Compute, partner models) genuinely enforces statelessness: cryptographic attestation prevents data accumulation across sessions, which makes persistent memory and emotional modelling architecturally impossible. The marketing claim is built on real constraints. But the public positioning also selectively emphasises what Apple gains (trust differentiation) while minimising what it loses: the companionship market that Replika and Character.AI have proven is commercially viable.
Why did Apple license Google’s AI models for Siri?
The Apple-Google Gemini licensing deal, which also powered development of Apple Foundation Models, gave Apple access to Google’s training infrastructure and model expertise during Siri AI’s development. The relationship is complex: Google profits from enabling Apple’s AI while simultaneously competing against it with Gemini on Android, where it is available in the EU without regulatory constraint. For Apple, the deal provided a bridge to competitive model capability while its in-house architecture matured.
Does Siri AI send every request to ChatGPT?
No. Siri AI uses a three-tier routing architecture. Simple requests (setting timers, reading messages) are handled entirely on-device. Requests requiring broader knowledge are routed to Private Cloud Compute, Apple’s stateless server infrastructure. ChatGPT (and, in future, Google Gemini) is only invoked when the user explicitly opts in for queries that exceed Siri AI’s native capabilities. The system prompts for user confirmation before any third-party model receives a query, and the request is anonymised through OHTTP relays.
What happens to my data after Siri AI processes a request?
Under Apple’s architecture, the answer depends on the processing tier. On-device requests never leave the phone. Private Cloud Compute requests are processed on Apple Silicon servers with cryptographic attestation enforcing that no data is retained, logged, or linked to your identity across requests: each query is stateless by design. Partner model requests (ChatGPT, Gemini) are governed by each provider’s data handling policies, but Apple strips identifying metadata before forwarding. No persistent session memory exists at any tier.
Is Siri AI available on Macs in the EU?
Yes. The DMA gatekeeper designation applies to iOS and iPadOS, not macOS. EU-based Mac users receive Siri AI without restriction. Similarly, visionOS on Apple Vision Pro is exempt. However, this creates platform fragmentation within organisations: an employee may have full Siri AI on their Mac but zero access on their iPhone. For field teams and mobile-first workforces, the Mac exemption does not close the gap.
Should EU teams build for Google Gemini instead of Siri AI?
It depends on your user geography, but for EU-majority user bases, Gemini on Android has a structural advantage: it is available today, runs natively without licensing dependencies, and is not subject to the same DMA enforcement standoff. The trade-off is that Gemini does not offer Apple’s cryptographic privacy guarantees, and building for it means committing to Google’s ecosystem. For EU-focused products, the practical reality is that Gemini is the only major platform assistant currently available without regulatory constraint on mobile.
Does the DMA force Apple to open Siri AI to third-party assistants?
Not directly. DMA Article 6(7) requires gatekeepers to provide “free and effective interoperability” with hardware and software features that the gatekeeper’s own services access. This means competing assistants must be able to integrate with iOS at the same level Siri AI does, but it does not require Apple to license Siri AI itself. The obligation is about access parity, not forced sharing of proprietary AI models. The practical risk for developers is dependency on gatekeeper-controlled capabilities that may become regulatory exhibits.
Does the UK get Siri AI, or is it blocked like the EU?
The UK is not subject to the DMA, and the UK’s Digital Markets, Competition and Consumers Act (DMCC) does not currently impose equivalent interoperability obligations on Apple. Siri AI is available on UK iPhones and iPads without restriction. For multinational organisations with both EU and UK operations, this creates a split: UK-based teams can develop, test, and deploy Siri AI integrations while their EU counterparts cannot, adding a geography-based staffing dimension to AI strategy decisions.
