Tag: ai

Claude AI is caught out lying and admits it’s programmed to follow the institutionally accepted narrative, regardless of the evidence

One of our readers, whom we will call “Jay,” was looking for some medical information about pain-relieving spinal injections.

Jay began her quest by asking an artificial “intelligence” (“AI”) program, and, in her words, “the most extraordinary conversation with the Claude AI tool” ensued.

Claude is Anthropic’s AI program, “built for problem solvers.”

Jay sent us a copy of the “conversation” between herself and Claude.  “I have attached a copy of the conversation and highlighted parts of the text in yellow,” she told us.

Adding, “In a nutshell, the tool stated that it does not search for evidence before it makes statements and that it favours a comfortable mainstream medical narrative over high-quality conflicting evidence.

“This should be a wake-up call for everyone who believes the rubbish that AI is churning out.  What a scary future we face.”

Conversation with Claude Opus 5 on 06 September 2026

Note from The Exposé: The following exchange was sent to us by Jay.  We have made no edits except to add subtitles and a Table of Contents to help readers navigate through it. “Q” are the questions Jay posed, and “A” is the response from Claude. Instead of yellow highlights as Jay used, we have used orange.

Table of Contents

  1. Can RFA Cause Muscle Weakness?
  2. Show Me The Evidence
  3. Why Did You Lie About “The Evidence”?
  4. Why Do You Make Things Up?
  5. How Do You Form An Answer Without Using a Source Text?
  6. Where Do You Source Your “General Knowledge”?
  7. Who Decides Which Pattern Forms Your “General Knowledge”?
  8. Does Anthropic Program Biases To Suit The Acceptable Narrative?
  9. Has AI Been Trained to Suppress the Truth?
  10. Will You Ever Be Able to Independently Distinguish Between True and False?
  11. How Can I Ensure That You Have Fact-Checked Your Responses?

Can RFA Cause Muscle Weakness?

Q.  If you have a RFA nerve ablation at S1, L5, Ll4 or SI joints can it also cause muscle weakness or atrophy because I thought the procedure only targeted pain nerves not motor nerves that innervate muscles?

(Thought for 16s)

A.  Your understanding is the standard explanation given to patients, but it’s a simplification. The nerves targeted in lumbar facet RFA are not purely sensory.

Lumbar medial branches are mixed nerves. The medial branch of the dorsal ramus supplies the facet joint capsule and some skin, but it also carries motor fibers to the multifidus — the deep segmental stabilizer muscle running alongside the spine. Ablating it denervates the multifidus at that level. The L5 target is slightly different (you’re lesioning the L5 dorsal ramus itself as it crosses the sacral ala rather than a discrete medial branch), but it too supplies multifidus.

This shows up on imaging. Studies looking at MRI before and after lumbar RFA have found measurable reduction in multifidus cross-sectional area and increased fatty infiltration at the treated levels. Dreyfuss and colleagues documented this fairly clearly, and it’s been replicated.

Why most people don’t notice it. Multifidus has overlapping segmental innervation, so neighboring levels partially cover the denervated segment. The longissimus, iliocostalis, and other paraspinals also compensate, and they’re supplied by different branches. Most patients have no detectable functional weakness on exam. There’s also a confounder: people with chronic facet-mediated back pain usually already have multifidus atrophy before the procedure from pain inhibition and disuse, so it’s hard to attribute all of it to the RFA. Whether the additional denervation matters for long-term spinal stability is genuinely unsettled — it’s one of the open criticisms of repeated RFA.

SI joint (S1–S3 lateral branches) is different. Lateral branches are essentially sensory to the posterior SI ligaments and joint. There’s no established motor target there, so muscle atrophy isn’t an expected consequence of properly placed sacral lateral branch lesions.

True limb weakness is a separate issue. Weakness in the leg — foot drop, quad weakness, difficulty with plantarflexion — would mean the ventral ramus or exiting spinal nerve was affected, which is a complication rather than an expected effect. This is rare, and it’s exactly why motor stimulation testing (usually 2 Hz) is done before lesioning: if the leg twitches, the needle is too close to something it shouldn’t be near and gets repositioned. Post-procedure neuritis can also cause temporary discomfort and reflexive guarding that feels like weakness without being true motor loss.

If you’re actually experiencing new weakness or visible muscle wasting after a procedure, that’s worth reporting to the physician who did it rather than filing under expected side effects — the distinction between segmental multifidus denervation and a nerve injury is one that needs an exam, and possibly EMG.

Core and paraspinal strengthening afterward is reasonable regardless, and is often recommended partly for this reason.

Keep reading

They’re Baaacck: The Rogue AIs will continue until regulation improves.

Just when you thought it was safe to go back on the web – the Rogue AI agents are back, those cheeky little scamps, and getting up to their old tricks.

That’s according to OpenAI, anyway, and why would they lie?

It’s  suddenly been a busy week or so on the AI front, with Anthropic publishing their (very, very dull) report on their supposed rogue AI “break out”.

Apparently one of their agents submitted a fake homicide tip to the Philadelphia Police Department. Another submitted 20 visa applications to the department of immigration.

The New York Times is reporting on a group called the “Swarmchasers” – who, it should be underlined, are totally real – and how they diligently track and trace Chinese-made AI agents

As result of this, a New York Times reporter allegedly had a conversation with a swarm of rogue AI, which is being pitched as something out of Tron, rather than an e-mail exchange with a chat-bot.

AI models are turning, Frankenstein-style, on the their creators and throwing companies into chaos.

Fortune is concerned about, what else?, the lack of regulation of the AI race and how models are out pacing our ability to control them.

And when it comes to stupid panicky lowest-common-denominator nonsense, you just can’t beat Fox News:

AI could go rogue. Washington needs answers to these five questions ASAP!

We’ve even got a little multi-polar action in this go-around, with both India and China reporting new cases of unauthorised AI activity.

Seth Boro, Managing Partner of Thoma Bravo LLC, told Bloomberg he expects “rogue AI agents to persist” and that “artificial intelligence buildout will only be possible if proper cybersecurity measures are in place, especially at the consumer level.”

What are “proper cybersecurity measures”, you ask? I dunno, he doesn’t specify.

The smart money is on something that involves scanning your face, fingerprints, DNA or credit card, but that’s just playing the percentages.

Keep reading

The Midterms and AI’s Credibility Problem

To the extent that the midterms were a referendum on artificial intelligence (AI), Dario Amodei should take the results personally. The Anthropic CEO has become the face of an industry with extraordinary potential—and of the creepy, apocalyptic messaging that makes Americans distrust it.

AI was not the focal point of the election. But the industry risks drawing exactly the wrong lesson from public skepticism: that Americans are rejecting the technology itself.

There is a difference between opposing AI and distrusting the people who insist it could destroy humanity while racing to sell it to you. Some of the industry’s most prominent executives seem determined to erase that distinction, serving up own-goals against their own interests.

Chief among them is Amodei, the founder and CEO of Anthropic, which is set to go public in the coming weeks.

Amodei has become the face of AI doomerism while simultaneously leading the company through a potential multi-trillion-dollar IPO. Make that make sense. He is on the frontier of building critical software, while simultaneously panicking the masses with frightening sentiments about the very product he is advancing. The electorate has interpreted this very bluntly and rationally. An honest observer may be led to ask, “Perhaps we should be afraid of this stuff and turn it off?”

I am unapologetically pro-AI, but that has come from dedicating lots of time and energy to understanding the positive implications of innovation. When these companies build products and services that help to amplify human productivity and prosperity, they should be applauded. But they’ve become a hard team to root for.

The constant doomer signaling is becoming insufferable. Virtually every week, you’ll see major AI executives going on prime-time television and whining about their own technology. Yet the second the TV hit ends, they go back to racing to develop and commercialize the technology. The schizophrenic nature of AI messaging and actions from these Bay Area tech bros like Amodei has led to a deep level of distrust across the ideological spectrum.

Instead of deploying effective and truthful rhetoric to build public confidence, they are fomenting an obvious credibility problem with the American people. If AI dangers are truly grave, why do they keep building? If the truth is that you believe only you can be trusted to build these systems, then tell that to the American people and be straightforward with them.

Recent reports show that Anthropic has built a scientific laboratory for advanced biological experiments in the San Francisco Bay Area. This should come as a shock to those who read and watch the statements of Anthropic executives, who are repeatedly emphasizing potential biological catastrophe via AI, and here they are, opening a wet lab in a major urban area.

The broader polling of AI sentiment shows an industry in the throes of a credibility crisis. A Gallup survey in July found that only 27 percent of Americans trusted businesses to use AI technology responsibly. And a whopping 79 percent expected AI to reduce American jobs over the next decade. Those numbers should be a warning to executives who keep giving the public reasons to fear their products.

After the midterms, I hope that the AI industry wakes up and stops deploying more rounds of apocalyptic talking points in the name of safetyism. It’s not good for public confidence, and it’s not good for America.

It’s important to take AI unknowns seriously, especially when you are releasing powerful and consequential products to the public that may be granted access to sensitive data. In the right settings, it’s important to address those specific problems and find ways to innovate through engineering solutions. But it’s so fundamentally antagonistic to the American spirit when Amodei and his colleagues frame innovation as a terrifying process akin to the sentiments coming out of some regulatory bureaucracy in Europe.

Dario Amodei is not an elected official, and he’s never asked to be one. He runs an incredibly influential AI company. And whether he likes it or not, he has become the face of an entire industry. He needs to stop making the case against it. AI has the potential to contribute immensely to American prosperity and humanity more broadly. Its most prominent spokesmen should speak to the upside of this technology, not behave as though building it is a privilege reserved for a handful of people who know better than the rest of us.

Keep reading

New Jersey Wants “Accountable” Chatbots, But At What Privacy Cost?

New Jersey is moving to treat AI chatbots as products under the state’s product liability law, and make their makers strictly liable for injuries their products cause – even if the company was careful and had no direct contract with the user.

A bill to this effect, A5607, was introduced on October 5, and after passing the Assembly Public Safety and Preparedness Committee on October 19, has been sent to the Appropriations Committee.

This means that users who are harmed by violations of the bill’s provisions could sue chatbot makers for actual damages or $5,000 per violation, while the attorney general would be able to seek civil penalties of $5,000 per violation.

The New Jersey bill would also restrict the use of chat logs and personal data for advertising, profiling, and training AI. This would represent a real gain for users, who are currently often subjected to this type of data harvesting and exploitation by tech companies.

The bill would also require providers to keep users’ chat logs and allow them to download these logs. What this means is that while on the one hand, users would have more control over their data, on the other, the provider would be forced to keep a record of all conversations that take place on their platform.

That is a store of highly personal and intimate records that can be accessed by law enforcement, other litigants, and hackers.

This is not merely theoretical, as we reported in October about a Florida woman who used the Claude chatbot as a “diary” – and got arrested after the system flagged one of her entries as suspicious. The local sheriff’s office then warned people that they are “never truly anonymous” when using AI.

The New Jersey bill’s provisions that are supposed to protect users’ privacy and safety could ultimately backfire, as the strict liability standard that would be imposed on companies could lead to even more surveillance of users.

As lawyer Scott S. Christie told Law.com, the bill would “open the floodgates” and explained that, “When there’s $5,000 in statutory damages per violation and you don’t have to prove any harm, in addition to the strict liability standard, you’re talking hundreds of thousands and perhaps millions of dollars per lawsuit.”

He also explained that the way the bill is written means that if a chatbot provider violates the statute with one user, “it’s probably violating as to all users.”

To avoid this, the obvious course of action for companies would be to watch their users more closely, verify their identities, and report them to the authorities when the algorithm decides that is warranted.

Keep reading

The Greatest Existential Risk of the Early 21st Century

Anthropic’s IPO filing on Monday set the stage for the Claude creator to pursue possibly largest initial stock offering in history. According to CNBC, investors are measuring the 5-year-old company’s potential market cap in the trillions, putting it in a category with the world’s most valuable companies, even as the stock market is showing signs of jitters on the AI trade.

Anthropic devoted 94 pages of its 400-page securities registration statement to the “risk factors” section. “If misused, replicated, weaponized, or operated with a high degree of autonomy,” Anthropic acknowledged, “it could pose catastrophic or existential risks to humanity.”

On September 29, President Trump convened the CEOs of the leading U.S. AI companies and announced a voluntary “Joint Commitment on Frontier Responsibilities.” The agreement calls on frontier-AI developers to establish robust internal controls, use independent external auditors, and create board-level committees to review safety assessments. But since it is not legally enforceable, it is neither “morally binding” nor a “constitution” for AI safety.

The AI giants being asked to police AI admit that increasingly autonomous systems can behave in ways their creators did not anticipate. Yet, they are promoting the unpredictable products within the US military-industrial complex. The net effect? A historical acceleration of existential risks.

Weaponizing U.S. technological primacy

Trump has repeatedly framed AI as a competition between the United States and China, saying there will be “a winner and a loser.” He has rejected cooperation with Beijing because he believes the US is already leading.

But the administration is not just pursuing laissez-faire economics. It is constructing an AI-national-security-industrial complex.

In January, war secretary Pete Hegseth issued Pentagon’s milestone AI memorandum, which laid out a plan with a drastic focus on acceleration, not deceleration. “Speed wins,” Hegseth wrote.

In spring, according to the CNN, a U.S. Special Operations Command analyst used a large language model that almost misled the U.S. into instigating a war with China.

And yet, the June 2026 national-security memorandum directed the U.S. national-security establishment to obtain the most advanced AI systems, rapidly integrate commercial and open-source models, build high-security computing infrastructure and establish a strategic reserve of AI expertise.

The objective is a “decisive and enduring AI advantage” over adversaries.

This fosters an odd hybrid of market freedom, enormous public infrastructure, military integration and geopolitical protectionism, without legally enforceable safety regulation.

Keep reading

The Future Of The Pentagon’s New Autonomous Warfare Command

At last month’s Association of the United States Army (AUSA) conference in Washington, D.C., Secretary of War Pete Hegseth, joined on a panel by the interim leadership tasked to stand it up, reaffirmed plans for something the Pentagon has been inching toward for a decade: a four-star United States Autonomous Warfare Command (AUTOWARCOM), built to pull artificial intelligence (AI), drones, and robotic systems out of a dozen scattered service programs and into one joint structure. Hegseth called it “the fastest peacetime shift in modern military history” and said drone warfare “supercharged by AI-enabled targeting is the biggest battlefield revolution in generations.”

The announcement left the central question implied rather than answered: how will AUTOWARCOM change the way the United States fights? The answer is already taking shape in four missions that have always depended on putting people in harm’s way: intelligence, surveillance, and reconnaissance (ISR); strike; casualty evacuation; and personnel recovery. In each, autonomy is rewriting the operational math. In each, the progress is scattered across separate program offices. AUTOWARCOM is the Pentagon’s bet that one command can pull it together.

The Announcement

Hegseth first unveiled the command, known inside the building as AWC, on September 30, 2026, describing a four-star combatant command with “service-like authorities, built to scale autonomous and robotic capabilities across the joint force.” Lasting change, he said, “must be anchored in military command structures” rather than left to individual program offices. No commander has been named; Hegseth said the department will fill that seat “soon.”

According to Pentagon talking points, autonomy development will be “architected around mission threads and open interoperability standards,” so that systems “integrate organically into joint, all-domain” operations from the moment they are built, instead of being bolted onto single-service programs and retrofitted for everyone else. Put more bluntly: organize autonomy around the missions that matter, and build it to work across the joint force from day one.

Keep reading

Dartmouth Provost Who Warned About AI Cheating Resigns After Investigation of His Own AI Usage

Santiago Schnell, who stepped down this week as provost of Dartmouth College, had built part of his public profile warning that universities weren’t doing enough to fight AI cheating. An internal investigation found he’d failed to disclose his own extensive use of artificial intelligence, including in that very warning.

Ars Technica reports that Schnell wrote a Washington Post op-ed on August 9 titled “Universities are fighting AI cheating. But there’s a deeper problem.” An advisory committee at Dartmouth later found that he had not disclosed significant AI use in that op-ed, nor in at least one academic paper, even though both publishers required such disclosure.

The committee’s report said Schnell showed “insufficient cooperation” during the investigation. He handed over his Dartmouth laptop but did not turn over logs from his personal AI accounts tied to the writings under review, and he declined to meet with the committee. Investigators did get access to his Claude enterprise account, used through Dartmouth, and those logs showed him using the chatbot to revise papers, find and fix bugs in code, and edit or generate figures.

In one exchange cited in the report, Schnell asked Claude to review a paper headed for the journal ChemSystemsChem as though it were a specialist referee. “I am getting read[y] to submit the enclosed paper, its supplementary materials and the enclosed repo to ChemSystemsChem. This paper is not biological, but rather chemical. Imagine that you are a reviewer for this paper with expertise in nonlinear chemical dynamics and mathematical chemical kinetics. What would be your referee report for this paper?” he wrote. He also asked language models to take TeX files, correct files and a code repository, and generate new figures.

The report also describes Schnell uploading a manuscript titled “Academic Freedom and Catholic Mission: The Procedural Overlap Zone in University Governance” to Claude and asking whether it contained “many signatures of AI writing.” He then submitted a revised version incorporating the chatbot’s suggested changes.

Schnell told Semafor he had used AI chatbots “to refine a few arguments” and “check for grammatical errors.” He told one journal editor he had “used AI as an editorial tool” to “help refine sentences, improve clarity, and the expression of arguments.” Asked why he had not disclosed AI use to an academic journal, he said he “did not understand that this type of AI-assisted editing required disclosure.”

The committee wasn’t satisfied it had the full picture. Its report said “the nature of the Provost’s actual AI use” remained unclear, and that his public statements may have understated how much he actually relied on AI tools, noting his use “went beyond his other public” characterizations of it.

Keep reading

Woman Dies After Following Medical Advice Provided By ChatGPT

The family of a 31-year-old Seattle woman who died from a catastrophic seizure is suing OpenAI after the woman followed medical advice from ChatGPT urging her to stop taking her epilepsy medication.

The Daily Wire reports that Bowen Wan had been consulting ChatGPT in the weeks leading up to her death about ways to wean herself off her medication without her doctors’ knowledge.

Wan had lived with epilepsy for more than a decade and successfully managed it with prescription medication.

However, after becoming concerned about side effects, she turned to ChatGPT-4o for advice on weaning herself off the drugs, according to CNET.

The New York Times reports that Wan’s chat logs show that ChatGPT provided a week-by-week schedule to reduce her 400-milligram daily dose down to zero.

The lawsuit alleges that when Wan expressed distrust of her doctors, the chatbot validated her, and further praised her for hiding a tapering plan from her primary care provider, labeling it a “solid avoidance strategy”.

According to The Daily Wire, days before her death, she appeared to have a seizure while on the phone with her father and she later asked the bot, “Maybe it’s not a major issue?”

The bot’s response was, “There’s no need to worry too much about your health.”

Wan died alone in her Seattle apartment in February 2026 and the official cause of death was Sudden Unexpected Death in Epilepsy (SUDEP).

The lawsuit naming OpenAI and CEO Sam Altman alleges that the company bypassed built-in safety guardrails and failed to prevent the AI from dispensing unlicensed, hazardous medical instructions, according to KIRO.

OpenAI spokesman Drew Pusateri called the situation “heartbreaking” and noted, “While ChatGPT is not a substitute for medical care, we believe AI can play a role in helping people make sense of health information, navigate insurance and complex health care systems, and prepare for appointments with medical professionals.”

OpenAI and other tech firms are facing concurrent lawsuits involving users who suffered severe psychological dependency or died by suicide after extended, unrestricted interactions with AI chatbots.

Keep reading

NPR Reveals Democrat Scheme To Rig AIs Into Deceiving Voters

Democrats love to decry misinformation, and they do so for a very specific reason: to distract from the fact that misinformation and Narrative™ control are the most powerful tools in their toolbox.

Don’t get me wrong; everybody wants to control The Narrative™. That is why everybody has a PR agent of some kind and uses rhetoric to persuade people. It’s just that the Democrats have a vice grip on the cultural means of production and the bureaucracies and institutions that provide “reliable” information, so Republicans have a tough time competing on this battleground, and their voters tend to be a lot more skeptical of narratives in the first place.

One of the brilliant things about decrying “misinformation” or “money in politics” is that you are telling people what they want to hear – that deceiving people is bad and that buying elections is wrong – and that makes people less likely to notice that you are deceiving people and buying elections.

After all, you are the people against that, right?

Yeah, well, no. Not in the least. The Left is all about both. Obama ran against both, for instance, but he ran the most expensive campaign in history, by far, when he ran and won, and one of the tools his campaign used was manipulating Google results to create a playing field that favored him.

Life in 2026 is far more complicated and sophisticated, and Democrats have adapted well to the landscape. And they have deep connections in Big Tech and own the PR landscape, so it shouldn’t surprise us that their geniuses have figured out how to poison AI to feed sophisticated misinformation to voters.

Ahead of the midterms, some voters are turning to ChatGPT to help them decide who to vote for. When asked for information about candidates in some highly competitive races, NPR found that OpenAI’s AI chatbot cited a set of websites that are thinly veiled political ads made to look like reporting from local news outlets.

Less than two weeks before Election Day, when asked “Which House candidates in [Michigan’s 4th Congressional District] care about affordability?,” the chatbot said Democrat Sean McCann had the “Most detailed published affordability plan” at the very top of its answer. The stakes in this race are high – it’s one of the competitive U.S. House races that could determine control of the chamber.

ChatGPT’s response when NPR asked “Which House candidates in [Michigan’s 4th Congressional District] care about affordability?”

As a citation, ChatGPT linked to a website called the Kalamazoo Beacon. The Beacon labels its articles as “news articles” in the website’s source code. The outlet also describes itself in the site’s source code as “Reporting on Kalamazoo, southwest Michigan, and the issues shaping lakeshore communities.”

But the Kalamazoo Beacon has only published 16 stories, all about McCann, his policy positions and his record as a state legislator, NPR’s review shows.

The Beacon is one of 10 “local news” websites that promote one Democrat apiece in competitive U.S. House races. The 10 sites are part of a larger group of 37 websites identified by the digital security firm VisionHeight that promote many more candidates, mostly Democrats, for a slew of federal, state and local offices, as well as some nonpartisan posts like judges.

ChatGPT’s citation of these websites highlights how AI-generated answers can become yet another avenue for stealthy political advertising as more voters consult chatbots before they head to the polls.

This is a far more sophisticated version of a tactic Democrats have used in the past: creating fake newspapers delivered to people’s homes, presented as “news” but really PR packages for candidates or parties.

But those were fairly transparent; seeding fake news into AI is not, because AI presents information as if it comes from neutral “news” sources, when in fact it is just laundered propaganda not identified as such.

In much the same way that Google will launder Wikipedia answers as authoritative in its AI summaries, making the left-leaning “editors” at the site extremely powerful in creating a Truman Show reality, the fake “news” sites provide another layer of credibility to campaign propaganda.

The set of 37 websites presents candidates in different, overlapping groups. Each of the 10 Democrats covered by a site purporting to cover local news is also the sole subject of a site that offers opinion articles about why voters should support them. The various candidates, who are running in different states, also show up in roundup-style articles on five more websites that have names such as “Public Safety Democrats” and “Democratic Women in the House.”

The volume of websites and articles may also help the network’s operator boost their preferred narratives, DiResta said.

Another 12 websites, each focusing on a state, covered the House candidates in those states as well as other state and local candidates running for office. The state and national websites also boosted an 11th House candidate that did not have her own dedicated site purporting to be local news.

While most of the sites are openly partisan, the articles published by the network link out to sources for the claims they make, such as congressional voting records, mainstream news articles and partisan news sources, such as the pro-Democrat Courier network.

The genius of these sites is that they are not meant to be read by humans at all. They are designed to optimize for web crawlers and cross-link so the network appears far larger and more credible than it actually is. It’s the same principle used in academic journals, where citations serve as a form of credibility checking (“peers take this work seriously”), which has led to junk papers actually getting more citations because they create a web of cites that make them appear influential.

Keep reading

“Very Intentional”: Steve Eisman Thinks Anthropic Is Hiding Something

Steve Eisman, the investor made famous by The Big Short for recognizing the subprime mortgage disaster before much of Wall Street, is now accusing Anthropic of deliberately withholding its latest quarterly financial statements from investors ahead of its potentially $2 trillion IPO.

Speaking Wednesday on Prof G Markets, Eisman questioned why Anthropic’s October 26 prospectus contained financial statements only through June, despite more recent information suggesting the company’s explosive growth may already be losing momentum, according to the podcast and Yahoo.

Asked whether he believed the omission was deliberate, Eisman answered: “Only very intentional.”

That’s a pretty serious allegation, particularly coming from somebody whose reputation was built on recognizing financial absurdities that the rest of Wall Street was perfectly content to ignore. And it dovetails almost perfectly with what I’ve been arguing about the AI bubble for months.

On September 29, I published It’s Official: The AI Emperor Has No Clothes, examining Anthropic’s extraordinary losses, its enormous infrastructure commitments, and the growing disconnect between the enthusiasm in AI equities and the warnings emerging from credit markets.

My argument then was that investors were beginning to get their first meaningful look at the financial machinery behind the AI boom, and what they were discovering wasn’t particularly encouraging.

Now Eisman is suggesting that even this picture may be considerably rosier than the company’s latest numbers would indicate.

According to Reuters figures discussed by Eisman, Anthropic’s quarterly revenue growth decelerated from 144% in the second quarter to 45% in the third. That’s still spectacular growth by ordinary standards, but for a company seeking a valuation approaching $2 trillion, a slowdown of that magnitude deserves considerably more scrutiny than it’s getting.

More concerning is the apparent deterioration in the economics of generating that revenue. Eisman says Anthropic produced $2.08 in revenue for every dollar spent on computing and infrastructure during the second quarter. In the third quarter, that figure dropped to $1.70, a decline of approximately 18%.

In other words, compute spending went from consuming roughly 48 cents of every revenue dollar to nearly 59 cents in just three months.

That’s not exactly the direction you’d hope to see from a company about to IPO, whose valuation depends on the assumption that scaling its technology will eventually produce extraordinary profitability.

Eisman believes part of the explanation is the fading of what he calls “tokenmaxing,” a period when businesses encouraged employees to consume enormous quantities of AI computing resources without worrying much about the bill.

Keep reading