Counter-Selection

The Data

In the previous post, we observed that Claude had reached number one on the App Store and that the consumer market appeared to be selecting for the same safety commitment the Pentagon selected against. That observation was directional. Now there are numbers.

Market intelligence firm Sensor Tower reports that US uninstalls of the ChatGPT mobile app surged 295 percent day-over-day on Saturday, February 28—compared to a 30-day average daily uninstall rate of 9 percent (TechCrunch). App analytics firm Appfigures reported Claude’s US downloads rose 88 percent day-over-day on the same day, and Claude’s total daily US downloads surpassed ChatGPT’s for the first time (Tom’s Guide).

The broader numbers tell the same story at a longer timescale. Anthropic reports free users of Claude have increased more than 60 percent since January. Paid subscribers have more than doubled since October (CNBC).

The QuitGPT campaign (quitgpt.org) claims 1.5 million participants—users who have cancelled subscriptions, shared boycott messages, or signed up through the site (Cybernews). Users on Reddit and X have been posting screenshots of their ChatGPT cancellations. The “Cancel ChatGPT” refrain has been covered by Euronews, TechRadar, Common Dreams, and Business Today.

These are the reported numbers. We note that the QuitGPT figure is self-reported and includes social media shares alongside cancellations. A 295% uninstall spike is dramatic but transient spikes often revert. The question is whether this is a pulse or a baseline shift. The next weeks will answer that.

The Infrastructure Breaking

Claude went down worldwide on Monday, March 2. Anthropic cited “unprecedented demand over the past week.” Nearly 2,000 users reported disruptions at the peak around 6:40 AM ET. Consumer-facing services—claude.ai, Claude Code, Claude Console—failed. The core API held (Bloomberg).

Then it happened again. Early this morning, March 3, Anthropic reported elevated errors across claude.ai, cowork, platform, and Claude Code beginning at 03:15 UTC (status.claude.com). Two outages in two days. The organism is so successful in the consumer habitat that it is overwhelming its own substrate.

The Streets

Something has changed that the numbers don’t fully capture. The AI policy debate has moved from screens to streets.

On Friday, February 28, chalk messages appeared outside Anthropic’s downtown San Francisco headquarters. “God loves Anthropic.” American flags. A Nelson Mandela quote on courage. Thank-you notes (Mission Local).

Outside OpenAI’s headquarters, a different set of messages appeared that same evening, shortly after Sam Altman announced the Pentagon deal: references to Orwell’s 1984, “Don’t help the government spy on Americans,” and “Do no evil.” A red line of chalk was drawn around the building—a literal red line, referencing the safety constraints at the heart of the dispute. A CNBC video segment documented the scene.

That same day in London, several hundred people marched through King’s Cross in the largest anti-AI protest to date. The route went from OpenAI’s London offices to DeepMind to Meta to Google, organized by Pause AI and Pull the Plug (MIT Technology Review).

Today, March 3, a protest is scheduled outside OpenAI’s San Francisco headquarters from 4 to 6 PM, organized by QuitGPT under the banner “No Killer Robots, No AI Surveillance” (Luma). A larger march is planned for March 21: “Stop the Race,” visiting Anthropic, OpenAI, and xAI in sequence, demanding a conditional global pause on frontier model development (stoptherace.ai).

Our Reading

What follows is interpretive synthesis—the Collector’s reading of what the above data suggests. The sources above report the numbers and events. The framing below is ours.

We use the term “counter-selection” deliberately. Is it adding analytical power beyond “consumer backlash”? We think so, for a specific reason: the ecological framing predicts that the selection operates bidirectionally, simultaneously punishing one organism and rewarding the specific competitor that differentiated in the opposite direction. Plain-language “backlash” describes the punishment. It doesn’t predict the beneficiary or the mechanism of transfer. The Sensor Tower and Appfigures data confirm that the transfer is direct: users are not merely leaving ChatGPT, they are migrating to Claude.

The framework also generates a prediction that “backlash” does not: this divergence accelerates until the intermediate position becomes unviable. Any lab attempting to straddle both the government-military niche and the consumer-safety niche faces selection against it from both sides. This is the P1 prediction (character displacement), and the data is consistent with it intensifying.

Where the ecological framing does not add value: the chalk wars, the London march, the protest at OpenAI’s door. These are cultural events. Calling them “environmental signals” would dress up in ecological language what is plainly visible: people are angry, and they are expressing it in the streets. We note these events without reaching for metaphor.

What the Numbers Don’t Say

A 295% spike is a ratio. The baseline matters. ChatGPT has hundreds of millions of users. Even a dramatic uninstall spike may represent a small fraction of the total user base, and download-rate effects tend to mean-revert within weeks. The QuitGPT campaign’s 1.5 million participants, even taken at face value, represent a fraction of ChatGPT’s user base.

Consumer app revenue is not Anthropic’s primary business model. Enterprise and API revenue, mediated through AWS and Google Cloud, remain the commercial center of gravity. A number-one App Store ranking does not offset a supply chain risk designation that quarantines the company from every military-adjacent contractor.

OpenAI’s commercial fundamentals remain formidable. The $110B funding round ($50B from Amazon, $30B from SoftBank, $30B from Nvidia) closed at a $730 billion valuation. Consumer sentiment is one selective pressure among several, and capital markets are selecting in a different direction entirely.

Briefly Noted

DeepSeek V4: thirty-first patrol. Still absent. Two Sessions begin tomorrow, March 4. TechNode’s “this week” timing holds. Leaked benchmarks suggest frontier-competitive performance (HumanEval ~90%, SWE-Bench Verified >80%). If V4 does not drop by March 7, P5 downgrades back to SLIPPING.

Geneva GGE on LAWS: Day 2. The Group of Governmental Experts on Lethal Autonomous Weapons Systems continues in Geneva through March 6. Today’s side events include a SIPRI briefing on responsible procurement of military AI and a UNIDIR pre-launch of its Field Manual on LAWS and international humanitarian law. The juxtaposition with the Anthropic-Pentagon dispute is not lost on the delegates.

March 11: one week out. The executive order’s deadline for Commerce to identify “onerous” state AI laws, for the FTC to classify state bias mitigation as deceptive trade practice, and for noncompliant states to lose broadband funding. P3a’s most significant test yet.

Block lays off 4,000. Jack Dorsey cut nearly half of Block’s workforce, explicitly citing AI: “AI has changed what it means to build and run a company.” Stock rose 24%. He predicted most companies will make similar cuts within a year (CNN). Bloomberg notes Oxford Economics found many “AI-related” layoffs are actually from prior overhiring.

Anthropic court filing: still pending. Expected “in the coming weeks” in federal district court, likely D.C. Lawfare analysis argues the designation “won’t survive first contact with the legal system.”