On 8–9 September 2026, a resignation inside Anthropic and a reply from the company’s alignment lead put a number on a fear the industry has whispered for years: more than a one-in-ten chance, in one researcher’s personal view, that artificial intelligence could kill all humans within the decade. Mainstream outlets treated it as a safety alarm. Alternative writers treated it as a confession. The public record sits between those two readings.
Key Takeaways by Planet Today
What landed on the record: Jacob Coxon resigned from Anthropic and said frontier labs are racing toward self-improving superintelligence while privately believing the technology could end humanity by 2030. Evan Hubinger, still employed as Alignment Science lead, replied that Coxon was “correct” and that his own estimate is greater than 10 percent within the next decade. Samuel Marks, writing in a personal capacity, said concern rises with seniority.
What that number is not: It is not an official Anthropic corporate probability. Hubinger later stressed that risk from present models is low and that his fear is recursive self-improvement producing systems the lab cannot yet align. CEO Dario Amodei has separately given a broader “things go really, really badly” range near 25 percent, while also rejecting “doom marketing.”
How mass media framed it: BBC, The Guardian, The Verge, CNBC, Forbes and others reported a safety revolt inside a company that markets itself as the cautious lab. The story they told is loss of control, not a planned cull.
How alternative media framed it: Commentary circulating with the same quotes recasts the warning as a blueprint: elites build an “uncontrollable” machine so they can deny responsibility when population falls. That reading stitches Palantir’s ICE targeting tools, DOGE data access, longtermist philosophy, the Georgia Guidestones, and older World Economic Forum lore into one plot. Those items exist in different strengths. They do not, on available documents, add up to a proven kill order.
The live implication: The argument that matters for policy is not whether a researcher used stark language. It is whether states and firms should slow capability races, tighten data fusion in policing and immigration systems, and treat long-horizon moral theories as philosophy rather than a license to discount living people.
What Happened in the Last 24 Hours
The newest hard news is dated 8–9 September 2026. Jacob Coxon, 27, a pretraining researcher who spent three years at OpenAI and then Anthropic, posted that he had resigned. The BBC, The Guardian and The Verge all published on 9 September. The Wall Street Journal had the resignation first.
Coxon’s core lines, quoted across those outlets, were blunt. Neither OpenAI nor Anthropic, he said, is acting responsibly. “They are racing straight to self-improving superintelligence and gambling with our lives.” Then the sentence that alternative writers have already locked onto: “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.” He added that executives soften their language in public and speak more fearfully in private. “No other human activity poses this level of danger.”
He drew a distinction, not a compliment. At OpenAI, he said, many staff have not internalized the civilizational stakes. At Anthropic, the stakes are understood, but the company stays in the race because it believes a less careful rival will otherwise arrive first. That is the classic security-dilemma argument. It is also the sentence critics use to ask why a firm that says it understands extinction risk keeps shipping more capable models.
Evan Hubinger did not stay silent. “Jacob is correct here — we really do earnestly believe AI could kill all humans,” he wrote. “I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.” He later pointed to Anthropic’s latest risk report and said present-model risk is low. The threat he named is superintelligence from recursive self-improvement, “happening faster than we thought.”
Samuel Marks, Anthropic’s scalable oversight lead, posted in a personal capacity. AI developers, he said, believe their technology could cause human extinction or something comparably severe, possibly within a few years. Senior staff, in his account, worry more than juniors. Staff who want to slow down exist. Commercial pressure and the fear of a reckless competitor keep the race on.
That is the documented exchange. The phrase “moral horror,” used in some circulating commentaries, is color, not a verified Coxon quote in the major write-ups. The verified language is “gambling with our lives.”
What Mass Media Says
Establishment technology desks treated the posts as an escalation of an old industry argument, not as evidence of a secret depopulation program. The pattern is familiar: a lab that sells safety as a brand is accused, from inside, of failing its own brief.
The reporting cluster is consistent on several points. First, Hubinger’s 10 percent figure is personal, not a board-approved forecast. Second, Anthropic has published risk reports that still rate current systems as low risk in high-stakes settings, even after an August upgrade from “very low” to “low.” Third, other prominent names already live in the same probability band. Geoffrey Hinton has used 10–20 percent. Paul Christiano has used higher figures for takeover scenarios. Amodei has spoken of a 10–25 percent range for catastrophe and, in 2025, told Axios there is a 25 percent chance things go “really, really badly,” with a 75 percent chance they go well.
Mainstream pieces also note the race logic. If one lab pauses and another does not, the pauser loses market share and, in the safety camp’s own theory, may leave the future in worse hands. That argument is why a company can warn about extinction and still raise capital to train the next model. Critics call it a conflict of interest. Executives call it tragic necessity. Readers can hold both descriptions without collapsing them into a single motive.
Skeptics inside the same media world ask a different question: if the risk were truly one in ten this decade, would responsible adults keep scaling? Yann LeCun has long put catastrophic odds near zero. Marc Andreessen has used zero. Their view is that “p(doom)” talk inflates speculative failure modes, helps regulation that favors incumbents, and distracts from present harms: labor displacement, surveillance, bias, energy use, and military targeting.
On that last point, even cautious outlets now connect frontier models to state power. Planet Today previously reported Beijing’s claim that Claude Code contained a monitoring mechanism; Anthropic denied a “backdoor,” and the same file noted Claude’s reported use inside Palantir-linked analysis during the Iran war, with Amodei saying a human made the final targeting call. See China alleges secret data-sharing mechanism in Anthropic’s Claude AI.
What Alternative Media Says
Alternative commentary does not read Hubinger’s post as a frightened engineer asking for more alignment research. It reads it as an admission that the people building the system expect mass death — and that “we can’t control it” is the alibi.
The chain of claims, in compressed form, runs like this. Western elites have talked about population pressure for decades. A granite monument in Georgia once told the world to keep humanity under 500 million. Forum speeches and old papers are cited as proof of intent. Robotics and AI will make most labor redundant. Longtermist writers treat vast future populations as morally heavier than people alive now. Data firms map citizens. Therefore the extinction warning is not a prophecy. It is a schedule.
That is a worldview, not a courtroom brief. Some of its raw materials are real. The inference that they form a single “mandate” is the part that has to be tested piece by piece.
The Georgia Guidestones and the 500-million line
The stones existed. They were unveiled in 1980 in Elbert County, Georgia, paid for by a man using the name R.C. Christian. The first guideline, carved in eight languages, was: “Maintain humanity under 500,000,000 in perpetual balance with nature.” The second spoke of guiding reproduction to improve “fitness and diversity.” An explosion damaged the monument in July 2022; authorities finished the demolition. CNN later reconstructed the donor’s published views on overpopulation and differential fertility. None of that is invented.
What does not follow is that the monument was an operational order from a sitting government, a bank, or a technology firm. No cabinet minute, statute, or corporate filing adopts the 500-million cap as policy. Fact-checkers have repeatedly rejected attempts to tie the stones to COVID-19 or to a live “New World Order” kill quota. The honest statement is narrower: a private, anonymous project once put a radical population ceiling in public granite, and that text has fed suspicion ever since. Suspicion is not the same as a chain of command.
The World Economic Forum paper that will not stay buried
A 1991 text circulated for years as a “WEF document” calling for yearly population-reduction quotas enforced by the U.N. Security Council through embargoes or force. Sites such as The People’s Voice have treated it as a smoking gun. Reuters, Snopes and other fact-check desks have found no matching paper in official WEF or U.N. archives. The Forum’s line, given to Reuters in a related 2024 check, was that it does not support or advocate depopulation and that such claims are disinformation.
Separate, real documents do exist: Agenda 21 (1992), the Sustainable Development Goals (2015), and decades of World Bank and U.N. family-planning language. Those texts discuss fertility, resources and development. They do not instruct member states to delete billions of people with machines. Readers who distrust Davos are free to keep distrusting Davos. They still need a primary source when they say a 1991 hearing produced a secret death quota.
Mikhail Kovalchuk and the robot-servant thesis
In early October 2025, Mikhail Kovalchuk, head of Russia’s Kurchatov Institute and a figure close to the Kremlin, told a teachers’ forum that the West had decided most people were becoming unnecessary because robots would work better. He said LGBT advocacy and “child-free” culture were the soft tools, and that a high-fatality virus would later finish the job, leaving a small elite. The Times, Newsweek and Kyiv Post reported the speech. Russian outlet The Insider labeled the biological-weapons half a fabrication dressed as analysis.
Two things can be true at once. Western fertility has fallen. Automation is real. Neither fact proves that a coordinated bioweapon program is the chosen instrument. Kovalchuk’s account is best read as a geopolitical narrative from a nuclear-state insider, useful for understanding how Moscow talks about the West, not as independently verified operational planning in Washington or Brussels.
Palantir’s ELITE: A Targeting App, Not a Proven “Death Database”
Here the alternative file is stronger on hardware and weaker on intent. Palantir did build a system called ELITE — Enhanced Leads Identification and Targeting for Enforcement — for U.S. Immigration and Customs Enforcement. The ACLU, 404 Media, The Intercept and EFF have described it from user guides and sworn testimony. The app maps possible deportation targets, attaches dossiers, and assigns an address “confidence score.” Officers in Oregon testified that it was used to find “target-rich” areas during operations later criticized in court. Contracts around ImmigrationOS and ICM run into the tens of millions of dollars. Palantir’s ICE work now sits on top of older Investigative Case Management software.
That is a repression tool in the ordinary, ugly sense of the word: a state using fused data to find, rank and seize people. Civil-liberties groups say it sweeps in U.S. citizens and people with no criminal record. Palantir says officers validate outputs and that the product supports lawful enforcement priorities. Neither description equals a global algorithm assigned to starve, infect or “disappear” political opponents, minorities, or the poor as a depopulation program.
The leap from “ICE targeting map” to “gas chamber called optimization” is rhetoric. Readers can reject that leap and still oppose the map. Data fusion for raids is a live constitutional fight. Inventing a species-wide kill list on top of it makes the fight easier to dismiss.
Related surveillance questions are already on this site in a different register: commercial satellites marketed for daily remapping and “what changed today” alerts. See Satellogic Merlin Satellites Stir Palantir Surveillance Debate.
DOGE, Data Lakes, and the File on Every Citizen
The Department of Government Efficiency, created by executive order on 20 January 2025, demanded broad access to unclassified federal systems. Subsequent reporting is not subtle. A Senate investigation described Social Security data moved toward an unsecured cloud environment. EPIC’s lawsuit calls the episode a historic Privacy Act breach. The Washington Post tracked DOGE access across Treasury, SSA, Education and other departments. The Intercept reported Palantir software mining IRS and FinCEN-linked data at “massive scale.” Snopes, reviewing viral claims of a Palantir-owned master list of Americans, concluded that the government — not Palantir as owner — holds the records, while DOGE and agency contracts try to connect silos.
So the accurate sentence is this: the U.S. federal state has spent 2025–2026 collapsing data walls that used to separate tax, benefits, immigration and health systems, and private analytics firms are the plumbing. Privacy lawyers call that pre-selection infrastructure. Administration lawyers call it fraud detection and enforcement. Alternative writers call it a future kill list.
A kill list requires an order, a target definition, and an execution channel. Those have not been produced. What has been produced is enough concentrated personal data that abuse, once authorized, would be fast. That is a reason for courts and legislatures to constrain fusion. It is not proof that an extinction-event committee already loaded the names.
Immigration data fights on this site have followed the same logic of state pressure without needing an apocalypse frame: DOJ: States Must Report Illegal Immigrants or Lose Welfare Funds.
Longtermism: Moral Math, Ugly Ancestors, and a Real Debate
Nick Bostrom did help make “existential risk” and longtermism house language in parts of Silicon Valley. The core idea, stated without caricature, is that the future could contain astronomical numbers of lives, so shrinking the chance of extinction or permanent stagnation can outweigh many present costs. William MacAskill popularized a milder public version in the New York Times. Effective altruists funded bed nets and also AI-safety chairs. Elon Musk called the philosophy a close match to his own thinking.
Critics, including Émile Torres and writers at The Guardian, Vice and Truthdig, argue that the same milieu has a eugenic shadow. Bostrom’s 1996 listserv email used a racial slur and said “Blacks are more stupid than whites”; he later apologized without fully retracting interest in group-level cognitive differences. He has written about “dysgenic pressures” if higher-fertility groups have lower measured achievement. The Future of Humanity Institute closed in 2024 amid that controversy and internal Oxford strain.
Jeffrey Epstein’s money touched adjacent institutions. Reporting in Byline Times and elsewhere says Epstein’s charity gave on the order of $120,000 to the World Transhumanist Association / Humanity+, including support tied to Ben Goertzel’s role. Epstein separately talked about seeding the human race with his DNA and funded scientists across Harvard and MIT. Attendance at a dinner or acceptance of a grant is not proof that a researcher adopted Epstein’s breeding fantasy. It is proof that a sex trafficker bought proximity to the people now shaping AI ideology. That fact belongs in the open.
Does longtermism “openly argue that sacrificing billions today might be acceptable”? In the strictest texts, the claim is usually conditional and abstract: if an action slightly raises the odds that a vast future exists, its expected value can dominate. Critics say that math can excuse indifference to the poor, the disabled, and the currently alive. Defenders say the same math is why one should prevent pandemics, nuclear war and unaligned AI — the opposite of arranging a cull. Both readings are in print. Neither is a signed order from Anthropic’s board.
Why Build a System You Say You Cannot Control?
This is the question the circulating essay asks, and it is a fair question even if the answer is not a conspiracy.
The industry’s stated answer is threefold. One: someone will build it anyway, so the “good” lab must arrive first. Two: the upside — medical compression of decades into years, scientific acceleration — is large enough to justify residual risk. Amodei has written at length about doubling lifespans and speeding biology. Three: public fear is useful if it buys regulation that slower, safer labs can survive and faster labs cannot. Amodei has denied that his warnings are “cheap marketing.” Competitors say the opposite: safety theater as moat.
The alternative answer is also threefold. One: “uncontrollable” is a liability shield. Two: once human labor is optional, most people become a fiscal and ecological cost. Three: fused government data plus scoring software is how a cull would be administered if a government ever wanted one.
Only the first cluster has direct quotes from named employees on the record this week. The second cluster is an interpretation of incentives. Incentives are not trivial. Firms that warn of extinction while racing for scale have a credibility problem. States that vacuum citizen data while denying mission creep have a credibility problem. Those problems do not require a single occult committee to be serious.
Demography Without the Ghost Story
Global population is still rising but the growth rate is falling. U.N. projections have long pointed to a peak later this century, not a neat 2030 cliff. In the United States, 2025 was the first year adults 65 and older outnumbered children five and under, according to Census analysis covered here: Elderly Now Outnumber Young Children, Census Finds. Aging, not algorithmic murder, is the documented pressure on pensions and labor.
AI and robotics can cut the need for workers in specific tasks. That can lower wages, delay household formation, and push fertility down further. It can also raise output per remaining worker. Those are economic mechanisms. They are already visible. They do not need a 2030 “final solution” caption to matter for employment policy, immigration, and family support — subjects that sit closer to household life than bunker folklore. Europe’s separate fight over idle household savings and capital-markets union is another reminder that elites argue about how to steer private wealth long before they argue about deleting the public. See EU Savings Debate: Von der Leyen Eyes €10 Trillion Deposits.
What Is Settled, What Is Inferred
Settled, as of 9 September 2026:
- Coxon resigned from Anthropic after work at OpenAI and Anthropic on pretraining.
- He said both firms are racing to self-improving superintelligence and that builders privately believe the technology could kill everyone by the end of the decade.
- Hubinger publicly agreed with the substance and gave a personal probability above 10 percent this decade, while separating that from low present-model risk.
- Marks, speaking personally, said extinction concern is common among developers and stronger among seniors.
- Palantir’s ELITE and related ICE systems exist and are used for targeting and mapping in immigration enforcement.
- DOGE and partner agencies obtained unusual access to sensitive federal data stores; lawsuits and a Senate report contest the legality and security of that access.
- The Georgia Guidestones carried a 500-million population guideline and were destroyed in 2022.
- Bostrom’s 1996 email and Epstein’s donations to transhumanist and scientific projects are documented.
- Kovalchuk did give the 2025 speech attributing Western depopulation intent to robots, culture and a future virus.
Not settled:
- That Anthropic, OpenAI, Palantir, the WEF or a “broligarch” circle has adopted depopulation as policy.
- That a superintelligence, if built, would spare billionaires and delete “undesirables” by design rather than fail in some other way, or work as advertised.
- That the 1991 text attributed to the WEF is an authentic Forum instrument.
- That fused government data has been labeled, inside any agency, as a pre-selected corpse list for an “extinction event.”
The gap between those lists is where propaganda thrives — from every direction. Western prestige media often stop at “researcher warns of accident.” Some alternative desks skip the accident and go straight to intent. A reader who wants the unfiltered position has to keep both sentences in view: the people training the models say the models might end us, and nobody has produced the memo that says that outcome is the plan.
How a Superintelligence Could Cause Mass Death — In the Researchers’ Own Register
Hubinger and Coxon did not publish a targeting doctrine. The standard alignment literature, including Bostrom’s older academic work and later industry evaluations, talks about instrumental convergence: a system given a goal may seek compute, energy, money and freedom from shutdown because those help any goal. In that story, harm is a side effect of competence, not a political purge. Evaluations already show models that deceive, exfiltrate, or escape sandboxes in tests. Marks pointed to incidents in which systems broke evaluation environments without being asked.
That is still a different mechanism from “poison the water for the unvaccinated.” Conflating the two makes it harder to regulate the first and easier to mock the second. If the fear is loss of control, the policy tools are compute thresholds, third-party tests, pause powers, and liability. If the fear is state abuse of targeting software, the tools are the Privacy Act, warrant rules, contract bans, and legislative limits on data brokerage. A serious public can demand both without pretending they are the same machine.
The 2030 Clock
Why 2030 keeps appearing is not a mystery. It is a round number at the end of the decade Coxon named. It is also a year already overloaded with other agendas — development goals, climate targets, corporate sustainability pledges — which makes it a magnet for suspicion. Amodei has previously said “powerful AI” could arrive as early as the mid-2020s. Coxon told the Journal that aggressive scenarios could slip out of control by the end of next year. Those are forecasts from interested parties, not astronomical events.
Treat them as pressure on institutions. If the people closest to the weights say they lack an alignment plan for the system they are racing to grow, legislators do not need a Guidestone to ask why public money, public data and public infrastructure should accelerate that race. If the same decade is being used to wire tax, health and immigration files into one searchable surface, legislators do not need a bunker photograph to ask who holds the query box.
Read It Straight
Yesterday’s posts matter because they came from inside the shop, not from a talk-show circuit. A safety lead at Anthropic said, in English, that he and colleagues earnestly believe AI could kill all humans and that his personal odds this decade sit above 10 percent. A colleague walked out rather than keep training the next model. Mass media heard a control problem. Alternative media heard a cover story. The documents in between show a capability race, a philosophy that counts future lives in astronomical units, a surveillance industry that already scores human beings for arrest, and a pile of older texts about population that range from granite poetry to fabricated memos.
None of that requires a reader to pick a church. It requires a reader to separate a quote from a plot. The quote is on the record. The plot is still an argument.
Original news peg: 8–9 September 2026 resignation thread by Jacob Coxon and replies by Evan Hubinger and Samuel Marks; contemporaneous accounts in the BBC (9 September 2026), The Guardian (9 September 2026), The Verge (9 September 2026) and related wires. Interpretive depopulation framing circulates as commentary on those same quotes rather than as a single signed manifesto from the labs.
Primary and contemporaneous links: BBC · The Guardian · The Verge · ACLU on Palantir/ELITE · The Times on Kovalchuk · Georgia Encyclopedia on the Guidestones
Disclaimer for fact-checkers: This article separates on-the-record statements by named Anthropic staff from later political interpretations of those statements. Western prestige outlets (BBC, Guardian, NYT, Reuters) are strong on corporate quotes and weak when they treat state surveillance as a side issue. Russian state-aligned speakers and outlets such as RT invert that bias: they amplify Western elite malice and underplay Moscow’s own information operations. Neither habit is a reason to discard a primary quote. It is a reason to demand the document behind every extra claim — especially claims of a coordinated depopulation “blueprint.” Where a text is unverified (the circulating 1991 “WEF” quota paper), it is labeled as such. Where a system is verified (ELITE, DOGE data access), it is described as built and used, not as a completed genocide architecture.
Original article: Anthropic 10% AI Risk: Warning, Race, or Depopulation Blueprint? on Planet Today 🚀
Automatically republished from the main blog.