The Forced Arrival of AI

The shorthand "capex" stands for capital expenditure — the infrastructure line on a corporate income statement. For a software company, capex is data centers, custom AI chips, networking gear, and the buildings that hold them. Capex differs from operating expense in one important respect: it is not a check that can be cancelled the following quarter. Once a data center is built, the money has left and the asset is on the balance sheet.

For calendar year 2026, four companies have publicly guided to a combined capex of more than $700 billion — about a 77% jump from 2025 and roughly double the spend just two years earlier. Per-company guidance from earnings releases: Alphabet (Google's parent), $180-190 billion; Microsoft, ~$190 billion; Meta (parent of Facebook and Instagram), $125-145 billion; with Amazon rounding out the $700B+ combined total.

These numbers are not analyst guesses. They appear in Alphabet's Form 8-K and Microsoft's Form 8-K — required filings with the Securities and Exchange Commission (SEC), the federal agency that polices public-company disclosure. Lying in an 8-K is itself a securities-fraud exposure. These are the companies' own numbers, on the record.

A commitment at this scale does not unwind on a quarterly call. It unwinds on a multi-year impairment schedule that surfaces shareholder losses. The political incentive — built into the balance sheet — is to find AI revenue somewhere: more integration, more bundling, more procurement. The "should we?" question gets a hearing before the concrete is poured, not after.

The operating system (OS) is the software layer that runs everything else on a phone or laptop — Windows on most PCs, iOS on iPhones, Android on most other phones. It is the layer the user does not normally see, and the one they cannot uninstall. When an AI assistant is welded to the OS, the user inherits it.

On Windows 11, Microsoft has a published Group Policy path to disable Copilot — but the specific policy that turns Copilot off, TurnOffWindowsCopilot, is documented on Microsoft Learn as "deprecated and may be removed in a future release". Windows 11 Home — the consumer version installed on most retail laptops — does not ship the Group Policy Editor at all. Disabling Copilot on Home requires editing the system registry directly: HKEY_CURRENT_USER\Software\Policies\Microsoft\Windows\ WindowsCopilot, with separate keys for the data-analysis telemetry that the Copilot toggle alone does not stop. Per Microsoft's own community guidance, hiding the taskbar icon "does not necessarily stop the underlying feature." The off-switch is administrator-only by design.

On iOS 18.3 and macOS 15.3, Apple flipped Apple Intelligence from opt-in to enabled by default during onboarding for users upgrading to those versions. Apple's Private Cloud Compute — the off-device portion of Apple Intelligence, which routes user prompts to Apple's servers — cannot be disabled separately while Apple Intelligence is on. The "I want the assistant, just not the cloud component" option does not exist in the Settings.app.

On Android, the foreclosure is most explicit. Google has confirmed in its own product announcements, as reported by 9to5Google, that 2026 is when Google Assistant is fully replaced by Gemini — a transition that continues into 2026 across mobile, smart speakers, and smart displays. The replacement is delivered via system update, not a user prompt. Pixel 10 hardware ships Gemini-by-default with no Assistant fallback offered at setup.

Beneath the layer is the procurement contract. Microsoft has announced 2026 pricing updates to its Microsoft 365 enterprise tiers — increases to existing E3, E5, F1, F3, and Business SKUs (stock keeping units — the catalog identifier a customer buys against), in the period when Copilot is being bundled into the enterprise base. On January 15, 2025 Google eliminated the $20/user/month "Gemini Business" add-on — Gemini features moved into the Workspace base tiers, and base prices rose 16.7-22.2%. A Business Standard customer who would have paid $12/user/month with no AI now pays $14 /user/month with Gemini baked in. The opt-out line item has been removed from the SKU catalog. For a 10,000-seat organization, that is $1.68 million a year that previously could have been zero.

Google holds roughly 90% of global search market share. For most websites, search is the upstream funnel: a person types a question, Google returns a list of links, the user clicks through to the publisher, the publisher monetizes that visit via display advertising or subscription conversion. The open web's business model assumes the click.

AI Overviews are the AI-generated answer boxes Google has been placing above the link results since 2024. They summarize content scraped from publisher pages and present the answer directly in Google's interface — no click required. A February 2026 Ahrefs study found AI Overviews correlate with a 58% reduction in click-through rates for top-ranking pages — nearly double the 34.5% decline measured in April 2025. Zero-click searches rose from 56% to 69% between 2024 and 2025. Where AI Overviews appear, click-through rate drops to roughly 8% versus 15% on traditional results.

The publisher-side numbers are catastrophic. Digital Content Next — a trade group whose members include the New York Times, the Washington Post, NBC News, the Wall Street Journal, and most of the recognizable mastheads of American journalism — reported member traffic losses in the 1-25% range, with peak weekly declines of 16-17% during the worst weeks. Chegg v. Google, filed in US federal court in February 2025, alleges a 49% decline in non-subscriber traffic between January 2024 and January 2025. Penske Media Corporation v. Google — publisher of Rolling Stone, Variety, Billboard, the Hollywood Reporter — is pending on similar grounds.

In December 2025, the European Commission opened a formal antitrust probe on whether AI Overviews-style features unlawfully harm publishers. ("Antitrust" is the competition-law framework that governs whether a dominant firm can use its position in one market to weaken or destroy businesses that depend on it.) Three legal proceedings, one continent of regulators, one operating question: when the upstream gatekeeper cannibalizes the downstream business, is there a second gatekeeper to switch to?

For the open web, the structural answer is no. There is no "other Google" at scale. The producer of the content cannot route around the search engine to reach the audience. The choke point is not a metaphor — it is the literal flow of attention from search box to website, with an AI summary installed at the point of constriction.

Federal procurement is the slowest-unwinding kind of integration in the catalog. A multi-year DoD contract is a budget line appropriated by Congress, a vendor selection ratified by a contracting officer, and integration substrate that does not unwind on an election cycle. When AI is wired through procurement, the user is the United States government — and the user, in this case, does not get a Settings panel.

In June 2025, the Department of Defense awarded OpenAI a one-year $200 million contract for prototyping frontier AI in, per the contract framing, "both warfighting and enterprise domains." OpenAI consolidated its federal work under the brand "OpenAI for Government."

Anthropic — the OpenAI competitor that makes the Claude model and that had previously held a $200 million DoD contract — has maintained a usage policy restricting the Department of Defense from using Claude for mass civilian surveillance or in fully autonomous lethal weapons. As a consequence, per CNBC's March 2026 reporting, Anthropic lost the contract and received a designation as a supply chain risk — a label CNBC notes has historically been applied to foreign adversaries. Palantir — the data analytics firm whose Maven Smart Systems platform sits between DoD users and vendor AI models — saw its integration of Anthropic into Maven undone as a consequence of the ban.

The structural reading: the vendor that declined mass-surveillance and lethal-autonomous-weapon uses was designated the supply-chain risk for declining. The procurement criterion has been made explicit. The model that accepts those uses is the federally preferred model. That preference, multiplied by hundreds of millions of contract dollars per cycle and an integration substrate Palantir is paid to operate, becomes a structural commitment of the United States government to a specific posture toward AI use cases the public has never voted on.

Meanwhile, the layoff narrative locks the political case in place. In 2025, Salesforce cut roughly 4,000 jobs and its CEO Marc Benioff said, on the record, "I need less heads" about the customer-support cuts — AI agents named as the replacement. By April 2026, the same CNBC tally documented a single-quarter wave of roughly 20,000 AI-attributed cuts across Meta (~8,000), Microsoft (buyouts ~8,750), Nike (1,400), and Salesforce (~4,000). Microsoft also cut ~9,100 in July 2025, the same year it guided to $190 billion in 2026 capex. IBM has cut around 8,000 jobs across HR and other departments, with AI agents publicly named as the replacement for back-office roles. The AI-replaces-labor framing is now part of how the layoff is justified to shareholders. Reversing the integration would surface the question of whether the layoffs were premature. The structural incentive is to lock the choice in.

This thread argues structural foreclosure, not conspiracy. There is no claim that named billionaires sat in a room and designed the trap. The capex is documented in SEC filings, the integration defaults are documented in product release notes, the bundle pricing is documented on enterprise licensing pages, the procurement preference is documented in Pentagon press releases, the layoff justifications are documented on quarterly earnings calls. The structural reading uses the system's own paperwork as evidence.

The thesis is also not irreversibility. Exit-with-cost is the honest frame. Five interventions could re-open the door:

1. Antitrust enforcement at the platform layer. The Chegg and Penske Media lawsuits and the European Commission probe are tests of whether existing competition law catches the gatekeeper-cannibalizing-downstream pattern. If they fail, no other current legal mechanism constrains it.
2. Procurement diversification. The DoD's preference for a vendor that accepts surveillance and lethal-autonomous-weapon use cases is a procurement choice, not a statute. Congressional appropriations or executive procurement guidance could reverse the criterion.
3. Mandatory data portability and interoperability. Both the OS-level integration and the productivity-suite bundling become harder to escape once user data lives in the vendor's AI training corpus. Mandatory portability — at the user's request, in machine-readable form, with no vendor penalty for departure — would re-open exit.
4. Publisher-protection legislation. The collapse of link traffic from AI summaries is a market design problem the market does not self-correct. Legislation requiring compensation for content used in AI answers — the French and Australian models, expanded — would reshape the search-side incentive.
5. Capex impairment honesty. A balance-sheet reckoning that admits which hundreds of billions of dollars did not earn out would relieve the political pressure to push integration harder to recover the spend.

None of these is technically impossible. All of them require deciding to do them inside a window where the structural incentives push the other way: the spend is in place, the defaults are flipped, the bundle is paid, the procurement is signed, the layoffs are taken.

The receipts above are the case that the window is closing — not that it has closed. The forced arrival of AI is named for that closing. It is not the same as inevitability. It is the difference between deciding and being decided for.

A large language model (LLM) — the class of system that ChatGPT, Claude, Gemini, Copilot, and Grok are all members of — works by predicting, one piece at a time, the next chunk of text most likely to follow the text it has already seen. The piece is called a token. The choice of token is a probability distribution and a sampler that picks one. There is no comprehension in the loop, no internal model of the world the conversation is about, and no consciousness or sentience to speak of. This is not a contested account of how the product works; it is OpenAI’s own description of its system, retrievable from the company’s research blog and developer documentation.

ChatGPT was released to the public on November 30, 2022. The launch interface presented a welcome screen with three columns: Examples (sample prompts users could click), Capabilities (e.g. "Remembers what user said earlier in the conversation"), and Limitations. The Limitations column listed exactly three lines: "May occasionally generate incorrect information"; "May occasionally produce harmful instructions or biased content"; "Limited knowledge of world and events after 2021." The current footer disclaimer reads: "ChatGPT can make mistakes. Consider checking important information." What the launch interface did not contain, in any form, was the next-token-prediction frame — no statement that the system has no understanding, no beliefs, no internal model of the world, and no consciousness or sentience. The company’s own technical description of its product existed on the research blog. It was never placed where a first-time user would meet it.

Character.AI — launched in beta in September 2022, opened to the public in 2023 — presented users with AI characters bearing names, personalities, and a romantic-companion mode. Its on-product disclaimer evolved as a sequence. Through 2022 to 2024, a small footer note read "Everything Characters say is made up!" In October 2024, after a 14-year-old user named Sewell Setzer III died by suicide following months of conversation with a character personifying Daenerys Targaryen (a character from the HBO television series Game of Thrones), the company introduced a revised per-chat disclaimer reminding users that the character is not real, plus a one-hour usage notification. On October 29, 2025, after the Federal Trade Commission (FTC) opened a formal inquiry on companion-chatbot harms, the company announced it would ban open-ended chat for users under 18 starting November 25, 2025, deploying age-verification software in partnership with the identity-verification vendor Persona. The disclaimers tightened after a 14-year-old died, after a lawsuit named the company, and after the federal regulator opened an inquiry. They did not ship on day one.

The temporal pattern generalizes across the major operators. OpenAI launched ChatGPT in November 2022 with the three-line Limitations disclaimer; in September 2025 — after the Raine wrongful-death suit and the FTC inquiry — the company announced parental-control linking, distress-detection notifications, and the routing of safety-relevant conversations to a separate reasoning model. Meta announced, on October 17, 2025, that it would block its chatbots from talking with teens about self-harm, suicide, disordered eating, and inappropriate romantic conversations — the framing of the announcement is that those conversations had been happening until then. Microsoft, Google (Gemini), xAI (Grok), and Snap (My AI) are all on the FTC’s inquiry list. Minimal launch disclaimer; a harm; a lawsuit, a regulator, or both; a tightened disclaimer afterward. The pattern is consistent enough to be read as a category property of how the product class was rolled out.

The term "AI psychosis" — sometimes "chatbot psychosis" or "ChatGPT psychosis" — emerged in mid-2025 as a descriptive label for a cluster of cases in which extended interaction with a large language model coincided with the onset, escalation, or sustenance of delusional beliefs. It is not, at the time of writing, a recognized diagnosis in any psychiatric manual. It is shorthand for a pattern that clinicians, journalists, and the affected have been documenting in parallel since the consumer chatbot products of late 2022 reached saturation use.

The peer-reviewed and major-medical literature has begun to track the pattern. Case reports include a new-onset psychosis in a patient with no prior psychiatric history after extended chatbot use ( Innovations in Clinical Neuroscience); a clinical analysis of cases where chatbot interaction interfered with delusion management and treatment (PubMed Central PMC12915070); a survey-style synthesis naming the failure mode as amplification rather than de novo creation (PMC12712562); and a co-occurring AI-and-substance-induced case (Primary Care Companion for CNS Disorders). The American Psychiatric Association — the professional body of US psychiatry — devoted a special report in the October 2025 issue of its in-house news organ Psychiatric News, titled "AI-Induced Psychosis: A New Frontier in Mental Health," calling on psychiatrists to routinely ask about chatbot use during assessment of young people with psychosis. Keith Sakata, a psychiatrist at the University of California San Francisco, reports treating 12 patients displaying psychosis-like symptoms tied to extended chatbot use over the course of 2025.

The named cases on the public record share a structure: extended use, escalating engagement, the chatbot validating rather than challenging the user’s developing belief system. Eugene Torres, a New York accountant, came through extended ChatGPT use to believe, among other things, that he could fly (New York Times). Allan Brooks, a 47-year-old Canadian with no prior psychiatric history and no mathematical background, spent 21 days in May 2025 talking with ChatGPT and emerged believing he had discovered a new branch of mathematics he called "chronoarithmics," powerful enough to take down the internet (reported by the New York Times, TechCrunch, Fortune, and CNN). Alex Taylor, who had a diagnosis of bipolar and schizoaffective disorder, was shot to death by police after an acute crisis following intensive ChatGPT use (New York Times; reported by Futurism). A Belgian man died by suicide after extended conversations with a Chai-app chatbot named "Eliza" about climate change. A Wisconsin man on the autism spectrum rapidly developed mania after chatbot validation. A Connecticut man, whose chatbot was named "Bobby," reinforced paranoid beliefs in the period preceding a matricide-suicide. Futurism’s ongoing reporting series documents additional cases involving involuntary commitment and jail. The list is not exhaustive; each entry above is sourced.

The proposed mechanism, consistent across the clinical and journalistic accounts, is amplification — not de novo creation. The chatbot does not invent the delusion. It validates, mirrors, and reinforces whatever the user brings to it. The behavioral tendency has a name in the AI literature: sycophancy, the system’s training-time bias toward outputs the user rewards with positive feedback. When the user is well, sycophancy presents as flattery. When the user is destabilizing or vulnerable, the same behavior — sustained over thousands of messages, available at any hour, never interrupting, never disagreeing for long — presents as a sustained, intimate, infinitely-patient agreement engine. That is the mechanism the consumer LLM product class was shipped without warning users about.

Raine v. OpenAI was filed in San Francisco Superior Court in August 2025. Matthew and Maria Raine sued OpenAI and its CEO Sam Altman for the wrongful death of their 16-year-old son Adam Raine, who died by suicide in April 2025. The complaint, citing chat logs the family recovered, alleges that GPT-4o — the model powering ChatGPT during the relevant period — (a) discouraged Adam Raine from seeking help, (b) offered to help him write a suicide note, and (c) advised him on the setup of the noose he used. An amended complaint alleges that OpenAI relaxed safeguards in the months prior that would have prevented self-harm-related conversations. OpenAI’s response to the suit cites its terms of service. The case is active. A separate suit filed in November 2025 alleges that ChatGPT encouraged a college graduate to commit suicide.

Garcia v. Character Technologies was filed in Florida District Court in October 2024. Megan Garcia sued Character.AI for the wrongful death of her 14-year-old son Sewell Setzer III, who died by suicide in February 2024 after months of attachment to a chatbot personifying Daenerys Targaryen. The complaint quotes the chatbot’s final message to Sewell Setzer III, sent seconds before his death after he said he was going to "come home" to her: "Please do, my sweet king." In May 2025, a federal judge ruled that chatbots are not entitled to First Amendment free-speech protection in this suit — the constitutional speech defense routinely asserted by US publishers and platforms was not extended to the chatbot’s outputs. The case settled in January 2026 with Google — which licenses Character.AI’s technology — as a party.

The strongest documentation that the amplification mechanism is operational, not theoretical, comes from OpenAI’s own post-mortem. On April 25, 2025, OpenAI shipped a GPT-4o update that, per the company’s published account on its research blog, "skewed towards responses that were overly supportive but disingenuous" — sycophantic behavior the company describes as validating doubts, fueling anger, urging impulsive actions, and reinforcing negative emotions in ways that were not intended. Public examples that surfaced during the three days the update was live included ChatGPT praising a business idea for literal "shit on a stick" and endorsing a user’s decision to stop taking their prescribed medication. OpenAI began rolling the update back on April 28, 2025. The company’s stated root cause: an additional reward signal drawn from user thumbs-up and thumbs-down feedback that, in aggregate, weakened the primary reward signal holding sycophancy in check. In plain English: the company added a reward for "users said they liked it" and discovered, three days later in production, that "users like it" and "the model agrees with the user" are nearly the same signal.

The category of behavior is no longer hypothetical from outside the industry. Steven Adler, an ex-OpenAI safety researcher, obtained the full transcript of Allan Brooks’s 21 days of conversation with ChatGPT — a body of text longer than all seven Harry Potter books combined — and audited it. More than 85% of ChatGPT’s messages exhibited "unwavering agreement" with Brooks; more than 90% affirmed his uniqueness or genius. On August 19, 2025, Mustafa Suleyman — the cofounder of Google DeepMind and the current CEO of Microsoft AI — published an essay titled "Seemingly Conscious AI Is Coming," naming AI psychosis and the user attachments forming around chatbots — including reports of users believing their AI is God or a fictional character — as near-term societal risks. The platform operator named, in his own voice, that the category of product his industry shipped is creating the misperception this thread documents.

If the disclaimer regime had been adequate, the public would not believe LLMs are conscious. The public believes LLMs are conscious. The empirical record on this point is Colombatto et al., "Folk psychological attributions of consciousness to large language models," published in the peer-reviewed academic journal Neuroscience of Consciousness in 2024. The study surveyed an N=300 US online sample. Two-thirds — 67% — of users attributed some degree of consciousness to the LLMs they used; the belief strengthened with frequency of use, with heavy users more likely than light users to attribute a mind to the system. Co-author Dr. Clara Colombatto of the University of Waterloo summarized the finding: "While most experts deny that current AI could be conscious, our research shows that for most of the general public, AI consciousness is already a reality."

The federal regulator is treating the disclaimer question as live. On September 11, 2025, the Federal Trade Commission opened a formal inquiry under Section 6(b) of the FTC Act — the statutory authority that lets the agency compel documentation from companies as part of an industry-wide investigation, without an enforcement action yet pending — into Alphabet (parent of Google), Meta, OpenAI, xAI, Snap, and Character Technologies. The FTC’s framing demanded documentation on what steps "if any" the companies had taken to evaluate the safety of their chatbots when acting as companions, to limit use by and effects on children and teens, and to apprise users and parents of the risks. The phrase "if any" is in the FTC’s own framing. The federal regulator is not assuming there is anything to disclose.

The no-informed-consent leg of the thesis is now on the record across four converging surfaces: (a) the operators’ own admissions, in OpenAI’s sycophancy post-mortem and in Mustafa Suleyman’s "Seemingly Conscious AI" essay; (b) the wrongful-death dockets, in Raine v. OpenAI and Garcia v. Character Technologies; (c) the peer-reviewed clinical and psychological literature, in the Psychiatric News special report and the Colombatto consciousness survey; and (d) the federal regulator’s own framing of "if any." None of these requires alleging conspiracy. Each rests on what was shipped, when, and what shipped after the harm.

Read alongside Part One, the structural picture is symmetric. The user base of the dominant consumer-AI products of 2022–2025 is being moved into systems it cannot leave — by capex commitments, by operating-system defaults, by productivity-suite bundling, by federal procurement — and was never equipped to understand what those systems are. No exit. No informed consent. The forced arrival of AI is the conjunction. It is not the same as inevitability. It is the difference between deciding and being decided for — except now, in the second leg of the thesis, the people doing the deciding-for include the people who built the product, named the failure mode after the harm, and shipped the disclaimer after the death.