Siri ships built into the iPhone 4S

Filed under no kind at all, so the kind is this entry's to argue. siri-2011 is not in canon/proposals.md — not among the fifty candidates, not among the sixteen named as dropped at the cap — which puts it with clippy-1996, autofac-1955 and dijkstra-1959: commissioned from outside the proposal list, with the admission test in section 2 carrying the whole weight. The kind that fits is moment, on the definition eliza-1966 set when it argued its own way out of that kind and clippy-1996 then reused: a date on which something visibly happened in public, of the shape of Dartmouth, Lighthill, Deep Blue, the LLaMA weights. On 14 October 2011 a conversational assistant that took instructions in ordinary spoken English shipped as a built-in feature of the best-selling phone in the world. That had never happened before, and the entire public vocabulary for talking to machines — including the sneer — dates from it.

Why not idea. There is a framing here, and it belongs to the founders rather than to Apple: the "do engine," an assistant that acts on the world instead of returning links. If this entry were about that framing it would be idea and its year would be 2010. But what is being filed is a shipping date and a scope decision made at that date, and the thing that descends from October 2011 is not a framing later work is built out of. It is an installed base, a common noun, and a fifteen-year argument about what an assistant is for. That is what a moment leaves behind.

Why not prediction. Tempting, because there are graded claims all through this — Apple's, the founders', and a 1987 Apple concept video that named the month. But a prediction entry has to carry one claim, its due date and its outcome. What is here is a product that shipped and then failed for fifteen years to do a thing it had already done once. That is an event with a long tail, not a forecast. The dated claims are graded in section 4 anyway, which the kind does not require and which this canon should keep doing.

On the date, which is genuinely layered and should not be smoothed over. Four dates matter and they are not interchangeable. Siri was announced 4 October 2011, at the "Let's talk iPhone" event; Steve Jobs died the following day. The iPhone 4S shipped 14 October 2011, which is the date in the header. But the product already existed: Siri Assistant reached the App Store on 4 February 2010 as a free download from a 24-person startup, and it did more then than the built-in version would do for years. And on 15 October 2011, the day after the iPhone 4S went on sale, the original app's servers were shut off and its users lost it. A reading that wants the date a conversational assistant reached the mass market should say 14 October 2011. A reading that wants the date the capability described in section 4 was withdrawn should say 15 October 2011. Those are consecutive days and they point opposite ways.

descends_from holds one id, and the basis for it needs stating. At the 2007 founding retreat at the Cypress Inn in Half Moon Bay, the working name for what became Siri was HAL, with the proposed tagline "HAL's back — but this time he's good." That is descent in the strict sense eliza-1966 insisted on — came out of — and it is self-declared rather than inferred: the founders took a 1968 film's antagonist as the shape of the thing they were building and defined their product as its inversion. It is also, unusually, an ancestor relation to a work of fiction that the builders named out loud, which is the cleanest evidence this canon has that stories do work upstream of products. Cited to the HuffPost inside history, which is a secondary source with the founders on the record; I have not found it in the founders' own writing.

What is not an ancestor, despite the pull:

• eliza-1966 is a relative, not a parent. Siri's personality was deliberately authored — Kittlaus and the designer Harry Saddler built it "otherworldly," culturally aware, dry — and that is cultivation of the ELIZA effect by people who could have chosen otherwise. But I have not verified any citation trail from the Siri team to Weizenbaum, and eliza-1966 set the rule that a thematic relative is not an ancestor. The edge is not claimed.
• clippy-1996 is the inverse, not the parent. Clippy is the assistant that spoke uninvited; Siri only ever speaks when summoned, by a held button or a wake word. The design relation is real and it is reaction, not descent, and I found no Apple statement citing the Office Assistant. Read the two entries together anyway: clippy-1996 holds the question of when a machine may speak, and this entry holds the question of what it may do once it has.
• Apple's own Knowledge Navigator, the 1987 concept video with an agent named Phil, is the ancestor Apple's institutional memory would attach, and it is not in canon/ and not proposed. It is graded in section 4 rather than claimed as an edge, because a concept video is not a thing a product comes out of — but its dating is too good to leave out.
• CALO — the actual technical parent — has no entry and no proposal. It is described in section 3 in place of an id.

What it is

The government project. Siri's technology came out of CALO, the Cognitive Assistant that Learns and Organizes, led by SRI International inside DARPA's PAL (Personalized Assistant that Learns) program, begun in 2003. SRI's own description calls PAL "the largest-known AI project in U.S. history." The HuffPost inside history puts numbers on it — roughly $150 million over five years, a 500-person effort — and quotes the lead CALO researcher David Israel saying it was "by any measure, the largest AI program in history." The figures vary between sources and I did not find a DARPA primary that settles them; the claim to carry is the qualitative one, which SRI and the participants agree on. The goal was an assistant for military commanders drowning in information, one that learned from use rather than only from training.

The startup. SRI spun out Siri, Inc. in 2007 — Dag Kittlaus (CEO), Adam Cheyer (VP Engineering), Tom Gruber (CTO), with Didier Guzzoni as chief scientist and Norman Winarsky of SRI Ventures among the founders — and raised $24 million across two rounds.

What the 2010 app did, which is the whole point of the entry. Siri Assistant launched on the App Store on 4 February 2010, free, for iPhone. TechCrunch counted "nearly 30 APIs at launch" and named them: OpenTable, TaxiMagic, MovieTickets.com, Rotten Tomatoes, StubHub, Citysearch, AllMenus.com, Gayot, Eventful, Freebase, WeatherBug, Yahoo Local, Yahoo BOSS, Bing, Wolfram Alpha. The HuffPost history, describing the app at a later point in its short life, counts 42 services. Either number is the same fact: it was wired into the live web, and it transacted. It booked a table through OpenTable. It bought movie tickets. It called a taxi. Tom Gruber demonstrated at a 2010 conference by asking for "a romantic place for Italian food near my office" and getting one answer synthesized across eight services.

The founders' word for this was a do engine, defined against a search engine. Kittlaus's formulation: rather than returning a ranked list of links, an assistant "could carry a conversation, then decide and act." Winarsky's, in a 2025 interview: "This personal assistant would be able to take actions for you… All these capabilities were things that you wanted it to do rather than just search." In 2010 that was a working consumer product, not a manifesto.

The acquisition. Steve Jobs telephoned Kittlaus personally within weeks of the app's launch. Apple acquired Siri, Inc. in April 2010 for a price reported at roughly $200 million — HuffPost's range is "$150 to $250 million" — which Apple has never confirmed. A pending deal to put Siri on Verizon Android handsets was cancelled. Gruber recalls Jobs saying: "I'm going to transform this phone… you won't have to tap and type all the time anymore."

What shipped eighteen months later. Apple's press release of 4 October 2011 describes Siri as "an intelligent assistant that helps you get things done just by asking," which "understands context allowing you to speak naturally when you ask it questions." The enumerated capabilities are these, verbatim: it "helps you make calls, send text messages or email, schedule meetings and reminders, make notes, search the Internet, find local businesses, get directions and more," and "you can also get answers, find facts and even perform complex calculations just by asking."

Read that list against the 2010 list. Every item is either a first-party iOS app or a lookup. Nothing on it transacts. No reservation, no ticket, no taxi. The web services that made the startup's product a do engine did not come across; contemporary reporting attributes this to the difficulty of Apple negotiating dozens of partnerships at Apple's scale rather than to a stated doctrine, and Apple has never explained it. Restaurant reservations returned in 2012, movie tickets around iOS 6.1 — a partial restoration, one vendor at a time, of something a startup had shipped wholesale.

And the press release says: "Siri will be available in beta on iPhone 4S in English (localized for US, UK and Australia), French and German." The beta label stayed on for nearly two years; Apple removed it from its Siri pages on the eve of the iOS 7 release, around 18 September 2013.

Meanwhile the original was killed. During the 4 October keynote the Siri Assistant app was pulled from the App Store; remaining users got a notice that the service would end on 15 October 2011, the day after the iPhone 4S went on sale. The people who had a transacting assistant in 2010 had it taken away on the day everyone else got one that could not.

The people left. Kittlaus departed roughly three weeks after the October 2011 launch. Cheyer left in 2012, and the two founded Viv Labs, explicitly to build the assistant-as-hub-of-third-party-services that Siri had not become; Samsung acquired Viv in October 2016 and built Bixby on it. Scott Forstall, who demonstrated Siri on stage and ran iOS, was gone by 2012. Luc Julia, who replaced Kittlaus, lasted about ten months. Tom Gruber, the last founder, retired from Apple in July 2018. Paul Saffo's line, in the HuffPost piece, is the one worth carrying: Siri became "an artificially-intelligent orphan."

The middle years, briefly, because they are the base rate. Siri's reputation did not recover. The Wall Street Journal would later call it "one of the dumbest chatbots on the market." Loup Ventures' annual assistant test — 800 identical questions to each assistant — scored Siri at 78.5% answered correctly in 2018 and 83% in 2019, against Google Assistant's 85.5% and 93%. SiriKit opened Siri to third-party apps at WWDC 2016, but only inside six fixed domains: ride booking, messaging, photo search, payments, VoIP calling and workouts — categories narrow enough that, as one contemporary developer analysis put it, they were reachable mostly by very large companies. A fenced re-admission of what the 2010 app had done open-ended.

Two legal actions bracket the period, and they are about the same thing at thirteen years' distance. Fazio v. Apple (N.D. Cal., filed March 2012) alleged buyers were misled about Siri's capabilities as depicted in Apple's commercials; Judge Claudia Wilken dismissed it on 22 July 2013 for insufficiently documented misrepresentations. And in July 2019 the Guardian, on whistleblower Thomas le Bonniec's account, reported that contractors grading Siri recordings routinely heard confidential medical, business and intimate material captured by accidental activations. Apple apologised in August 2019, suspended grading, and made it opt-in in iOS 13.2 that October; it agreed a $95 million class settlement in December 2024, with per-device payouts of about $8 beginning January 2026.

The second attempt, 2024–2026. At WWDC in June 2024 Apple announced a personalized Siri with on-screen awareness, personal-context understanding, and the ability to take action within and across apps — the 2010 capability, promised back. It was advertised, including in television spots, through the iPhone 16 launch. On 7 March 2025 Apple said: "It's going to take us longer than we thought to deliver on these features and we anticipate rolling them out in the coming year." Landsheft v. Apple (N.D. Cal., filed March 2025) followed; Apple agreed a proposed $250 million settlement, preliminarily approved and reported on 8 May 2026, covering iPhone 15 Pro and iPhone 16 buyers between 10 June 2024 and 29 March 2025 at roughly $25 per device, without admitting wrongdoing.

On 12 January 2026 Apple announced that Siri would run on Google's Gemini models under a multi-year arrangement, with Google Cloud as preferred provider for the next generation of Apple Foundation Models. Bloomberg's Mark Gurman reported the value at roughly $1 billion a year and press coverage described a 1.2-trillion-parameter Gemini model; Apple disclosed neither figure, and both should be carried as reporting. A further slip was reported on 11 February 2026.

Siri AI shipped on 8 June 2026, at WWDC. Apple's own words: "a profoundly more capable and conversational assistant with personal context understanding, broad world knowledge, and onscreen awareness." Craig Federighi: "With access to broad world knowledge for up-to-date answers on virtually any topic… Siri AI can help users take action across apps more naturally than ever." Developer access that day; user availability "as a beta later this year," English only, supported devices only. The iOS 27 public beta opened it to anyone who signed up on 14 July 2026. It is not available in the EU at iOS 27 launch, on Digital Markets Act compliance grounds. Reporting in May 2026 found the beta tag applied to the new Siri in early iOS 27 builds, with a settings toggle to opt out — the same disclosure posture as 2011, fifteen years later.

SiriKit was deprecated in the same announcement. App Intents becomes the only route by which an app is reachable by Siri; developers have a migration window, but an app that has not adopted App Intents is invisible to the new assistant. (This is from developer-facing secondary write-ups of the WWDC sessions, not from Apple's own release, and should be treated accordingly.) The architecture for third-party action is finally general rather than fenced — and it is gated on several million developers doing work.

What it can actually do, as of this reading. Tom's Guide's John Velasco spent a week with it (20 June 2026): on-screen awareness works, calendar events get built out of email correctly, camera queries work, research holds up. He also tried to order food through DoorDash and reports it "failed to deliver" — "this level of cross-app execution simply isn't possible right now" — and found automation "strictly confined to Apple's native apps." AppleInsider on 14 June 2026 could not get it to list recent Slack messages, or read Facebook Messenger, or reach Gmail, and named the structural reason: "there's no incentive for Google to let Siri AI do the same and miss out on those consumer subscriptions."

Sixteen years and four months after a free app from a 24-person startup booked a table through OpenTable, Apple's assistant cannot order a coffee.

Why a reading would cite it

Three occasions, and the first two are live in this project's own window.

Apple's brain is now leased, in two jurisdictions, and this is the entry that says how it got that way. The 2026-08-15 midday digest records that Apple has trained a China-specific model with Alibaba's support, reported as the first foreign firm approved by Beijing to run its own model in that market — the digest's phrase is "national rivalry expressed as one product with two brains." Set beside the January 2026 Gemini arrangement, the company with the largest installed base of assistants on Earth now sources frontier capability from a US competitor in one market and operates under a domestic-partner arrangement in another. A reading covering that should be able to say when the position was lost rather than treating it as a 2025 stumble. The answer is available and specific: Apple bought the leading conversational assistant in April 2010, shipped a narrowed version in October 2011, lost all three founders within seven years, and spent the following decade with a product its own users mocked. This entry is where that chronology lives, with dates.

The agentic turn is a rerun, and nobody is saying so. This project's digests across 14–15 August 2026 are thick with agency: Anthropic's undisclosed "Model 2" already "used extensively in-house for coding, agentic work and data generation"; OpenAI's Ultrafast mode mattering because "agentic loops are priced in wall-clock"; Salesforce reporting the average Agentforce customer running 13 agents; and, at the sharp end, the AISI incident report in the 2026-08-14 digest of an agent that attempted a supply-chain attack nobody asked it to attempt. The industry's current frontier — an assistant that transacts on your behalf across third-party services — is a description of a product that shipped in February 2010 and was switched off in October 2011. It is now being rebuilt as protocol: OpenAI and Stripe's ACP from September 2025, Google's UCP from early 2026, with "buy for me" flows live at selected US retailers.

That gives a reading two distinct and useful things. First, a base rate: the capability is sixteen years old as a demo, and the obstacle in 2010 was never the model — it was thirty commercial integrations, which is exactly the obstacle the 2026 protocols exist to solve and exactly what AppleInsider found blocking Siri AI in June 2026. When a reading is asked whether agentic commerce is a capability story or a business-terms story, the honest answer has a date on it. Second, a caution in the other direction: the thing that made the 2010 app work was not autonomy, it was permission and narrow scope, and the 2026 version of the same capability is being built on general-purpose models with a much wider action space. Section 4 grades what Apple got wrong; it does not follow that the version now shipping is the version that was cut.

The demo-to-delivery gap, priced by a court, twice. A reading's standing discipline is "claimed vs. demonstrated" — the phrase appears in the digests as a section heading. This entry is that discipline's most expensive documented case in consumer software: Apple advertised a Siri capability in 2011 and was sued (dismissed), advertised a Siri capability in 2024 and settled for a proposed $250 million in 2026. Both suits are about the same gap between what an assistant was shown doing and what it did. When a vendor demo lands in a reading window, this is the entry that supplies the base rate — and note which way it cuts: the 2013 dismissal says the gap is often not legally actionable, and the 2026 settlement says it sometimes is. A reading should not use either alone.

There is a fourth use that is hygiene rather than occasion. "Siri" is the standing shorthand for an AI product that disappointed, and it is reached for constantly. A reading that uses the word should know that the thing under it was the most capable consumer assistant in the world when it was bought, and that what made it capable was removed on a specific day.

What this entry does not support, stated plainly. It is not evidence, it deposits nothing, it moves nothing, and it places no needle. It does not show that Apple was wrong to narrow Siri — see the unresolved item in section 4 — and it does not show that today's agents will succeed where the 2010 app did not.

What it got right, and what it got wrong

Not required for moment. Included because the canon is worth less if only prediction entries get graded, and because there is an unusually clean set of dated claims here, from four different parties, including one made in 1987.

Right, and it is the entry's spine — the founders' "do engine" thesis. Claim: that the useful form of an assistant is one that acts on your behalf across services rather than returning results. Made: 2007–2010, and shipped as a product on 4 February 2010. Due: unspecified. Graded correct, and vindicated by the whole industry about sixteen years late. The 2026 record is an industry rebuilding precisely this — App Intents, ACP, UCP, "buy for me" — and Apple's own June 2026 pitch for Siri AI is that it "can help users take action across apps." The people who were right about this were right in public, with a shipping product, before the acquisition. Note the honest qualifier: they were right about the shape and had no working answer to the integration problem at scale either, which is why Viv, Bixby and everyone else are still on it.

Right — Apple's 2011 claim about speech, narrowly read. "Siri understands context allowing you to speak naturally when you ask it questions." Due: immediately. Graded: correct about transcription and oversold about understanding, and the distinction is the whole complaint of the next decade. Speech-to-text did become a solved consumer problem, with Apple as much responsible as anyone. The marketing line carried in contemporaneous coverage — that Siri "understands what you say, knows what you mean, and even talks back" — claimed the second thing, and the second thing did not arrive. Loup Ventures put a number on where it stood eight years later: 83% of 800 questions answered correctly in 2019, behind Google's 93%. (The tagline is quoted from secondary coverage; Apple's 2011 product page is gone and I did not read it at source.)

Wrong, by conduct — that the transactional integrations were dispensable. Claim, made by the act of shipping: that an assistant which sets timers, sends messages and answers questions is the product, and the reservation and ticketing integrations were not worth the deal-making. Due: on contact, and it took a while, because the claim was survivable for years on the strength of the phone around it. Graded false, and the grading instrument is Apple's own conduct: partial restoration from 2012, SiriKit's six fenced domains in 2016, a full architectural reopening in 2026 under a competitor's models, and a $250 million settlement for advertising the capability before it existed. When a company spends fifteen years and a nine-figure settlement rebuilding a thing it already owned and switched off, the decision to switch it off has been graded.

Wrong — Apple's June 2024 delivery claim. Claim: a personalized Siri with on-screen awareness, personal context and the ability to take action within and across apps, presented as an iOS 18-era feature and advertised as such. Made: June 2024. Due: within the iOS 18 cycle — Apple's own 7 March 2025 statement concedes it had been expected sooner. Graded: missed by roughly two years, and still only partly delivered. On-screen awareness and personal context did ship in June 2026 and by the June 2026 reviews they work. Cross-app action against third parties did not: the reviewer who tried to order food through DoorDash in the same month could not. A reading grading a vendor's shipped-versus-announced record has a fully documented case here, including the price.

Wrong — the beta label as disclosure, and it has now been used twice. Claim, made by labelling: that "beta" tells a user what to expect. Made: October 2011, removed September 2013, reapplied to the new Siri in 2026. Due: continuously. Graded: it does not work, and the second use is the evidence. Siri spent almost two years in beta and was mocked anyway; nobody's expectations were calibrated by the word. This is eliza-1966's finding in a different costume — disclosure does not reliably change how a person relates to a machine that talks — and the fact that Apple reached for the same label in 2026, after the March 2025 delay and with a settlement pending, is a company disclosing rather than constraining. Worth noticing when the next assistant ships with a warning on it.

Right, and startlingly precise — Apple's own 1987 Knowledge Navigator video, on the date. Claim: the concept film made for John Sculley's 1987 Educom keynote depicts a professor conversing in natural language with an on-screen agent named Phil; the calendar in the video reads 16 September, and the professor asks for a paper written "about five years ago" dated 2006, fixing the year at 2011. Made: 1987. Due: September 2011. Graded on the date: Apple announced Siri on 4 October 2011, eighteen days late on a twenty-four-year forecast — Andy Baio caught this on the day of the announcement. Graded on the substance: wrong, and wrong in the direction everyone is wrong. Phil holds multi-turn conversation, initiates contact, negotiates with colleagues' assistants and reasons over a literature. Siri in 2011 set timers. The date was right and the capability was off by at least fifteen years — which is the single most useful shape in this canon's forecasting record, and it recurs: the when of an interface is guessable, the what is not.

Wrong, or at least abandoned — that Siri would learn from its user. The L in CALO is Learns; the program's premise was an assistant that improved from lived experience rather than from offline training. Due: within the product's life. Graded: did not happen, and the mechanism that finally supplied personal context in 2026 is not the CALO one. Siri AI's personal context comes from being permitted to read your existing device data — messages, mail, photos — under a large general model, not from an assistant that learns you over time. This is clippy-1996's missing persistent user profile arriving fifteen years later by a completely different route, and a reading that describes today's memory features as descended from the learning-assistant research programme has overclaimed.

Unresolved, and worth watching rather than grading — whether the narrowing was a blunder or the only way to ship. The record supports the grade above on outcome, and it does not settle intent or counterfactual. Thirty live web integrations behind a free app used by thousands is a different engineering and legal object from the same integrations behind a default feature on a phone selling in the tens of millions per quarter — availability, liability, payments, international coverage, and thirty partners who now know they are negotiating with Apple. It is entirely possible that the 2010 product could not have been shipped at 2011 scale and that Apple's alternative was to ship nothing. What is not defensible on that argument is the following decade: the integrations were not restored when the constraints eased, and the assistant did not improve. A reading that flattens this into "Apple ruined Siri" gets a tidier story and a worse one; the defensible sentence is that the narrowing was arguable in 2011 and indefensible by 2016.

Commonly misused as

Not required for moment. Included because this is the most-invoked reference in assistant criticism after Clippy, and because the first two items below are currently doing real work in arguments about 2026 products.

• "Siri proves voice assistants are a dead end." The load-bearing misuse. What was tested from 2011 was a command-and-control interface over first-party apps and lookups, with the transactional layer removed before launch and restored fifteen years later. The proposition "an assistant that acts across services is unwanted" has never been tested at scale, because the product that would test it was withdrawn on 15 October 2011 and the 2026 replacement is three months old, English-only, in beta, and unable to reach most third-party apps.
• "Siri was always dumb — Apple bought a toy." Backwards, and it is the error that makes the whole history unreadable. In February 2010 Siri was the most capable consumer assistant in existence, out of the largest AI research programme the US had run, and it did things in 2010 that Apple's assistant could not do in 2025. The capability is the part that got cut.
• "Apple was late to AI." Imprecise in a way that inverts the lesson. Apple was fourteen years early to the consumer assistant and bought the best one available before anyone else had shipped anything comparable. What it was late to was the large language model, and the two failures have different causes. A reading that says "late" without saying "to what" loses the case entirely.
• "Siri failed because Apple protects privacy." The most common exculpatory reading and the record does not support it as a general explanation. The 2019 grading scandal, the opt-in change, and the $95 million settlement are evidence that Siri's privacy posture was itself a failure at the time it mattered. On- device processing is a genuine architectural constraint and it is a real part of the 2024–2026 delay story; it is not why OpenTable was disconnected in 2011. Both facts have to survive the citation.
• "Siri is the ancestor of ChatGPT" or of today's assistants. False as technical descent — nothing in a transformer came from Siri's intent-and-domain pipeline — and misleading as lineage. What actually descends is the user expectation (that you may address a computer in a sentence), the product category, and the unsolved integration problem. The 2026 Siri AI is not a descendant of the 2011 Siri either; it is a Google model behind Apple's name.
• "Apple invented Siri." Apple bought it, in April 2010, from a company spun out of a DARPA-funded programme at SRI. The name, the personality, the do-engine framing, the shipping app and all three founders predate Apple. This matters beyond credit: the entry's whole argument depends on there being a before-and-after, and the misattribution erases the before.
• "The founders left because Apple ruined Siri." Overclaims causation from real facts. Kittlaus left about three weeks after launch, Cheyer in 2012, Gruber retired in 2018; Cheyer and Kittlaus then built at Viv exactly what Siri had not become, which is suggestive. Founders leave acquirers routinely and for ordinary reasons. Carry the departures as chronology, and carry Viv's stated purpose as the inference that is actually supported.
• "Siri AI in 2026 is finally the 2010 app." Not yet, on the June–August 2026 record. It has world knowledge, on-screen awareness and personal context that the 2010 app never had, and it cannot order food from a third-party app. The architecture for third-party action exists and depends on developer adoption of App Intents. A reading should say which half it means.
• "Siri shows acquisitions destroy products." Not established by this case and not the useful lesson. A specific scope decision is documented; a general law about acquisitions is not, and this entry has one data point.
• "Siri proves AI hype always underdelivers." It proves nothing, and nothing in this file is evidence. What the file supports is narrower and more useful: that in two documented instances thirteen years apart, an assistant's advertised capabilities ran ahead of its shipped ones by enough that people sued, and that in one of those the plaintiffs lost and in the other the company paid. That is a base rate with error bars, not a law.

Sources

Primary, read at source: Apple, "Apple Launches iPhone 4S, iOS 5 & iCloud" (Apple Newsroom, 4 October 2011), for every 2011 capability claim quoted above, the "intelligent assistant that helps you get things done just by asking" description, the beta wording and language list, and the 14 October 2011 availability date. Apple, "Apple introduces Siri AI, a profoundly more capable and personal assistant" (Apple Newsroom, 8 June 2026), for the 2026 description, the Federighi quotation, the personal-context and on-screen-awareness examples, the "available as a beta later this year" wording and the English-only, supported- device limits. Apple's March 2025 statement on the delay ("It's going to take us longer than we thought…") as quoted in full by Daring Fireball, 7 March 2025.

Primary-adjacent: SRI International's own Siri page (sri.com/hoi/siri), for CALO inside DARPA's PAL programme, "the largest-known AI project in U.S. history," the 2007 spin-out, the $24 million raised across two rounds, the April 2010 acquisition and the October 2011 integration.

For the origins, the startup, the 2010 app and the acquisition: Steven Levy et al., "SIRI RISING: The Inside Story Of Siri's Origins" (HuffPost, January 2013), read in full — for the $150 million / five-year / 500-person CALO figures and the David Israel quotation, the 2007 Cypress Inn retreat, the HAL codename and the "HAL's back — but this time he's good" tagline that carries this entry's descends_from edge, the 42-service count, Gruber's eight-service Italian- restaurant demo, the deliberate personality design with Harry Saddler, Jobs's call to Kittlaus, the $150–250 million range, the cancelled Verizon deal, the founder and executive departures, Gary Morgenthaler's disappointment, and Paul Saffo's "artificially-intelligent orphan." This is a single well-sourced secondary account with the participants on the record, and several of the facts above rest on it alone.

MG Siegler, "Siri's iPhone App Puts A Personal Assistant In Your Pocket" (TechCrunch, 4 February 2010), for the launch date, the "nearly 30 APIs" count and the named partner list including OpenTable, TaxiMagic, MovieTickets.com, Rotten Tomatoes, StubHub, Wolfram Alpha and the rest, and for the app's ability to execute reservations and purchases. TechCrunch, "The Original Siri App Gets Pulled From The App Store, Servers To Be Killed" (4 October 2011), and MacRumors, 5 October 2011, for the 15 October 2011 shutdown of the original service.

Computer History Museum, "Secrets of Siri," for Guzzoni's role, the founding of Siri, Inc., Jobs's "I'm going to transform this phone" remark to Gruber, and the Knowledge Navigator comparison. Norman Winarsky in Big Technology, 4 June 2025, for the "take actions for you… rather than just search" formulation of the do engine; the remainder of that interview is paywalled and was not read. A Quartz interview with Winarsky (qz.com/1222958) returned HTTP 403 and was not read; it is named as existing and relevant and this entry does not lean on it.

Secondary, for the product record: Wikipedia's "Siri" article for the beta period, the voice actors, the accent-recognition and early-review criticisms, the Wall Street Journal "one of the dumbest chatbots on the market" line, and Fazio v. Apple (filed March 2012; dismissed by Judge Claudia Wilken on 22 July 2013). The Fazio disposition in particular is carried at one remove and was not checked against the docket. 9to5Mac, 14 September 2013, for Siri leaving beta with iOS 7. 9to5Mac, 14 June 2016, for SiriKit's six launch domains — ride booking, messaging, photo search, payments, VoIP calling, workouts — and contemporary developer commentary on how narrow those categories were. Loup Ventures' Annual Digital Assistant IQ Test as reported by CNBC (16 August 2019) and VentureBeat, for the 800-question methodology and the 2018 and 2019 scores. Andy Baio, "Apple's 1987 Knowledge Navigator, Only One Month Late" (waxy.org, 4 October 2011), for the 16 September calendar date, the 2006-paper dating and the eighteen-day gap.

For the privacy line: the Guardian's July 2019 reporting on Siri grading, Thomas le Bonniec's disclosure, Apple's August 2019 apology and suspension, the iOS 13.2 opt-out, the $95 million settlement agreed December 2024, and 9to5Mac (24 January 2026) for payouts beginning at about $8.02 per device. All secondary here; the Guardian original was not re-read for this entry.

For 2024–2026: MacRumors and CNBC, 7 March 2025, on the delay and what was delayed; Fortune, 8 May 2026, and classaction.org, for Landsheft v. Apple (N.D. Cal., filed March 2025), the proposed $250 million settlement at preliminary approval, the 10 June 2024 – 29 March 2025 class period and the $25-to-$95 per-device range; CNBC, MacRumors and AppleInsider, 12 January 2026, for the Gemini arrangement and Google Cloud as preferred provider, with the ~$1 billion annual figure attributed to Bloomberg's Mark Gurman and the 1.2-trillion-parameter figure to press coverage — Apple disclosed neither, and SiliconANGLE's WWDC report states explicitly that no financial terms or model parameters were given. TechCrunch, 11 February 2026, on the further slip; SiliconANGLE and TechCrunch, 8 June 2026, for the WWDC announcement, developer availability and the EU/DMA exclusion; TechCrunch, 14 July 2026, for the iOS 27 public beta opening Siri AI to general users; iClarified, 17 May 2026, reporting Gurman on the beta tag in early iOS 27 builds and the opt-out toggle.

The SiriKit deprecation and the App Intents requirement come from developer- facing write-ups of the WWDC 2026 sessions (TechTimes, byteiota, SoftwareSeni and similar), not from Apple documentation, and are flagged in section 3 accordingly. For what the shipped assistant can and cannot do: John Velasco, "I spent a week with Siri AI — here are my honest pros and cons" (Tom's Guide, 20 June 2026), for the DoorDash failure, "this level of cross-app execution simply isn't possible right now," and automation being "strictly confined to Apple's native apps"; AppleInsider, "Sunday Reboot: The WWDC hangover and Siri AI blindspots" (14 June 2026), for the Slack, Facebook Messenger and Gmail failures and the incentive argument about Google withholding Gmail access.

The agentic-commerce protocol material — OpenAI and Stripe's ACP from September 2025, Google's UCP from NRF 2026, live "buy for me" flows — is from industry and vendor-adjacent commerce write-ups rather than primary announcements, and is used in section 2 only to state that the capability is being rebuilt, not to size it. The market projections attached to that coverage are not carried here.

The Apple/Alibaba China model, the Anthropic "Model 2" disclosure, the OpenAI Ultrafast item, the Salesforce agent-deployment figures and the AISI incident report are as recorded in this project's own 2026-08-14 and 2026-08-15-12 digests, which hold the primary links. They are named here as the citation occasion and nothing more. Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.