Ed Zitron
165 checked claims
across 1 source
gold frame — 50+ claims checked
What Ed Zitron talks about
A sample that mirrors their checked record and spreads across their subjects — not their worst claimsEvery row is reachable by scrolling its column · search matches claim text, verdicts, people, topics and sources · hover a row to trace its chain, click one to pin its detail
Claims by source
No claims match the selected verdicts.
The Man Who Calls BS On AI: They’re LYING About AI, 2027 Is When It All Breaks! | Ed Zitron
165 attributed claims Published August 2026
-
This is the largest non-consensual push of technology in history.
The comparative term “largest” could in principle be tested, but “non-consensual push of technology” has no agreed operational definition. Different readings of consent, technological adoption, and historical scope would produce different answers.
-
OpenAI lost approximately $20.9 billion in the previous year.
Reports based on audited 2025 financial documents describe OpenAI’s operating loss as $20.92 billion. The company’s reported net loss was higher because of a large restructuring-related charge, but the spoken figure accurately matches the operating loss.
-
There is no economic data supporting the claim that AI will replace all human jobs.
Available economic research does not support the absolute claim that AI will replace all human jobs. Current evidence instead describes limited, uneven, and occupation-specific effects, including some displacement risks but no broad economy-wide replacement of human employment.
Sources used for this check
- The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence
- Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence
- You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators
-
The speaker says a current AI system cannot perform search.
The claim does not identify the AI system, the type of search meant, or the standard by which success is judged. Because the referent and predicate are undefined, the statement cannot be reliably confirmed or refuted.
-
The speaker says that, until fairly recently, none of the named companies had revenue coming from AI.
The universal claim is refuted by Amazon’s own statement that AWS already had a multibillion-dollar AI revenue run rate. Microsoft also reported AI-inclusive Azure consumption growth during fiscal 2024, further contradicting the assertion.
-
Seventy percent of all AI revenue across the referenced companies comes from OpenAI and Anthropic.
The claim is not well-posed because “those three companies” has no clear antecedent—the preceding list names six companies—and “AI revenues” has no defined accounting scope. Available sources report different kinds of company revenue and run-rate estimates, but do not establish this precise 70-percent calculation.
-
The speaker says OpenAI and Anthropic are unprofitable.
Contemporary reporting described both privately held companies as losing more money than they made. The wording contains a grammatical error, but its substantive assertion that both were unprofitable is supported.
Sources used for this check
-
The speaker says OpenAI and Anthropic cannot afford to exist without funding from the same large technology companies.
The companies were reported to be unprofitable and dependent on substantial outside financing, but “cannot afford to exist” is an undefined counterfactual, and the claim does not establish which companies’ funding would be indispensable or whether other financing could replace it.
-
Amazon sent $50 billion to OpenAI during 2026.
Amazon announced a $50 billion OpenAI investment in February 2026, initially structured as $15 billion followed by $35 billion subject to conditions. By late August 2026, reporting said Amazon had completed the additional $35 billion, making the full-year statement accurate in the current record.
-
Amazon sent $5 billion to Anthropic during 2026.
Amazon announced that it would invest $5 billion in Anthropic, but the public announcement does not establish that the full amount had already been sent. “Sent” therefore overstates what the cited announcement confirms.
Sources used for this check
-
Google sent $10 billion to Anthropic during 2026.
Reports described Google as committing to invest $10 billion in Anthropic. The wording “sent” implies completed payment, which the cited reporting does not establish.
Sources used for this check
-
The speaker says the companies do not disclose their AI revenue.
The statement uses the undefined category “AI revenues” and an unspecified group of companies. Some companies report AI-related revenue or revenue run rates, while others combine AI with broader cloud or data-center categories, so the universal claim cannot be judged without a precise accounting definition and company set.
Sources used for this check
-
Anthropic and OpenAI are not public companies.
OpenAI and Anthropic are privately owned companies rather than publicly traded corporations. The broader wording about all other companies is less precise, but its specific qualification about Anthropic and OpenAI is correct.
Sources used for this check
-
The companies under discussion had spent more than one trillion dollars in capital expenditure.
Public reporting has described roughly $1 trillion in AI-related infrastructure commitments, but those are not equivalent to more than $1 trillion already spent. The claim presents commitments or obligations as completed expenditure.
Sources used for this check
-
AI GPUs use high-bandwidth memory.
Modern AI accelerators such as NVIDIA's GB200 NVL72 use HBM3E memory, including 13.4 TB of HBM3E across the rack-scale system.
Sources used for this check
-
Frontier-scale AI data centers can require thousands or hundreds of thousands of GPUs.
The claim is supported for the type of large-scale facilities under discussion: the Abilene Stargate campus has been reported as planned to house more than 450,000 NVIDIA GB200 GPUs. Smaller AI systems do not require that scale.
Sources used for this check
-
OpenAI and Oracle are building a 1.2-gigawatt Stargate data center in Abilene, Texas.
The Abilene Stargate campus is an OpenAI-and-Oracle project with a stated total power capacity of approximately 1.2 GW.
-
The Abilene Stargate campus has eight buildings, each intended to support about 50,000 NVIDIA GB200 GPUs.
The eight-building and roughly 50,000-GPU-per-building figures reflect reported design specifications, but the transcript states them as a definite future deployment. Public reporting instead describes each building as engineered for up to 50,000 GPUs and separately reports a total exceeding 450,000 GPUs.
-
Bristol uses about 7,800 megawatts of power per year.
The statement uses the wrong unit for annual consumption and gives a substantially higher figure than Bristol's official estimate. The cited municipal study projects total electricity demand of approximately 1,750 GWh in 2018.
Sources used for this check
-
Every media outlet has been strongly warning for three years that ChatGPT will take jobs and that people must use it or fall behind.
The claim is not well-posed enough to verify because "every single media outlet" and "screaming" are undefined, and it provides no bounded list of outlets or measurable standard for the alleged coverage. Sources establish that generative AI has been prominent for roughly three years, but not that every media outlet made these statements.
Sources used for this check
-
People predominantly use ChatGPT and similar tools as a form of search.
Survey evidence shows search and information-seeking are common uses, but large-scale usage data also identifies writing, practical guidance, and other activities as major categories and does not establish search as the predominant use. The claim therefore cannot be confirmed as stated without a defined population and usage measure.
Sources used for this check
-
Generative search is often unreliable on specific factual questions, including questions about whether something exists or whether a person said something.
Research evaluating multiple generative search engines found that their responses may not be accurate and that adversarial factual questions can induce incorrect answers. This supports the speaker's qualified claim that such systems can be poor at specific fact-checking questions, though it does not imply that every answer is wrong.
-
Google's SEO incentives produced low-quality content designed to rank well rather than serve human readers.
Research supports the existence of SEO incentives and low-quality optimized content, but it does not establish the speaker's specific attribution that Google pulled back spam guards because of Prabhakar Raghavan. Google's documentation instead describes continuing spam defenses and efforts to reduce unoriginal content, making the causal framing misleading.
-
Google's spam protections were rolled back because of Prabhakar Raghavan.
There are public allegations that decisions associated with Raghavan worsened Google Search, but the available reporting also records Google's explicit denial that its anti-spam protections were rolled back. The evidence does not establish the specific causal attribution in the transcript.
Sources used for this check
-
AI enables SEO-style low-quality content to be produced at scale.
Google itself describes scaled content abuse as producing low-value content at scale to manipulate search rankings and says the conduct may involve automation. This directly supports the claim that AI can scale and intensify SEO-content production.
-
In English, one token is approximately three-quarters of a word.
OpenAI gives the rule of thumb that one token is approximately three-quarters of an English word and approximately four characters. However, saying that a token 'is characters' falsely suggests a fixed character-based unit rather than a variable text encoding.
Sources used for this check
-
AI APIs commonly charge separately for input and output tokens, with rates expressed per million tokens.
OpenAI's token-based rate card states that rates are charged per million input and output tokens, and defines input tokens as tokens supplied in a request.
Sources used for this check
-
For reasoning models, internal reasoning tokens count toward output usage and are billed as output tokens.
OpenAI explicitly says that reasoning tokens are not visible as answer text but count toward output usage and are billed as output tokens.
Sources used for this check
-
SemiAnalysis estimated that fully using a $200 ChatGPT Pro subscription could represent about $14,000 in API-priced token usage, while a $200 Anthropic Claude Max subscription could represent about $8,000.
The reported SemiAnalysis figures are broadly accurate as API-equivalent usage estimates for the $200 ChatGPT Pro and Claude Max 20x plans. Framing them as the amount of tokens the companies actually have to 'burn' omits the analysis's assumptions and conflates API list-price value with provider cost.
-
OpenAI's operating loss was approximately $20.9 billion in 2025.
Reported audited financial documents put OpenAI's 2025 operating loss at $20.92 billion. The transcript's wording is loose because the approximately $20.9 billion figure refers to operating loss, while the reported net loss was about $38.5 billion after accounting adjustments.
-
OpenAI's losses were caused by users being able to burn as many tokens as they wanted.
Heavy usage can create substantial subsidization on some plans, but the cited financial loss was a company-wide operating loss, not a measured loss caused solely by users exhausting token allowances. The financial breakdown shows large costs beyond subscription inference usage.
-
Sam Altman said that AI spending had suddenly become a major or “huge” issue for companies.
Contemporary reporting on Altman's enterprise remarks quotes him saying that AI spending had gone from an issue that did not arise to a “huge issue.”
Sources used for this check
-
Uber exhausted its annual AI or token budget in three months.
There is credible reporting supporting a first-quarter or approximately three-month account, but other reporting says Uber exhausted its annual AI budget by April, four months into 2026. The exact timing and budget label are therefore publicly inconsistent.
-
GPU capacity is paid for on an hourly basis regardless of whether it is fully used.
Some GPU deployments do incur charges while idle or under a reservation, so excess provisioned capacity can waste money. But hourly GPU billing is not universal regardless of use: AWS states that On-Demand compute is billed only while an instance is running and that stopped instances are not billed for instance usage.
Sources used for this check
-
Users of token-priced LLM APIs are charged for the tokens processed or generated even when the output is not useful to them.
OpenAI's documentation states that API usage is priced according to input, output, cached, and reasoning tokens, and that those counts are used for billing. The pricing mechanism is based on usage rather than whether the user considers the answer correct or useful, subject to provider-specific exceptions for requests that fail before billable usage is recorded.
Sources used for this check
-
Inference providers and companies renting out GPUs do not appear to be profitable.
The statement does not identify the companies, define a time period, or specify whether "profitable" means gross profit, operating profit, net income, or cash flow. Public losses by some AI companies cannot establish the claim about all inference providers and GPU-rental companies.
Sources used for this check
-
OpenAI lost more than $5 billion in 2024.
Later-reported audited financial data put OpenAI's 2024 net loss attributable to the company at $5.09 billion, which is over $5 billion. Contemporary reporting in September 2024 described the figure as an expected loss of about $5 billion; the later actual figure supports the transcript's wording.
-
Moore's law does not apply to GPUs.
Moore's law concerns the growth of transistor counts on integrated circuits, and GPUs are integrated circuits whose historical development has been tied to transistor scaling. GPU performance does not track transistor scaling perfectly, but the categorical claim that Moore's law is not applicable to GPUs is false.
-
NVIDIA introduced CUDA in the 2000s as software for running software on GPUs.
NVIDIA introduced CUDA in 2006. CUDA is more precisely a parallel-computing platform and programming model, supported by libraries and development tools, rather than simply one underlying library; nevertheless, the substantive claim that it provides a way to run general-purpose software on GPUs is accurate.
Sources used for this check
-
CUDA enabled generative AI to grow.
CUDA was an important enabling component of GPU computing, and NVIDIA's documentation describes CUDA and GPU computing as foundational to advances including generative AI. The wording expresses a broad causal contribution, not that CUDA was the sole cause, so the claim is supported in that sense.
Sources used for this check
-
NVIDIA GPUs and related chips are used in data centers for AI training and inference.
NVIDIA documents GPU-based data-center systems supporting both AI training and inference. The statement is accurate as a description of the role of relevant NVIDIA chips, though not every NVIDIA chip is designed for both workloads.
-
Microsoft is one of the largest companies in the world.
Microsoft is routinely ranked among the world's most valuable companies by market capitalization, and its market value has been measured in the trillions of dollars.
Sources used for this check
-
Software quality is declining as LLM use and business pressure to use LLM services increase.
The claim is not well-posed enough to verify because “quality of software” is undefined and the asserted causal relationship is not established by a single agreed metric. Research has found declines in some software-delivery measures but positive or mixed results for other outcomes.
Sources used for this check
-
Bloomberg offers ASKB, which can generate or provide underlying BQL code for data analysis.
Bloomberg identifies ASKB as a conversational AI interface and says that, when responses include data analysis, it provides the underlying Bloomberg Query Language code. BQL is Bloomberg's query language for retrieving and analyzing market data.
Sources used for this check
-
Microsoft stock had never reached $575 per share at the time of the statement.
Available historical price records show Microsoft's highest recorded share-price levels below $575, with reported all-time highs in the roughly $542–$555 range.
Sources used for this check
-
METR has a chart measuring improvement in AI agents' ability to complete tasks, including tasks lasting about an hour.
METR publishes task-completion time-horizon evaluations that measure the length of software tasks AI agents can complete. Its materials describe evaluations ranging from short tasks to multi-hour tasks, supporting the statement that the chart tracks progress toward hour-long task completion.
Sources used for this check
-
Some unspecified tasks were successfully completed 50% of the time.
The statement does not identify what “them” refers to, which makes the claimed 50% completion rate impossible to verify from the transcript. No well-defined task, dataset, or benchmark is supplied.
-
Hallucination rates on simple summarization tasks fell from about 21.8% four years earlier to about 0.7%.
Vectara’s older leaderboard did report Gemini-2.0-Flash-001 at 0.7%, and its benchmark measures hallucinations in document summarization. However, Vectara later said its updated benchmark uses a larger dataset with longer, more complex articles and produces substantially higher rates; its May 2026 leaderboard listed the best model at 1.8%, with Gemini-2.5-Flash-Lite at 3.3% and OpenAI GPT-5.4-nano at 3.1%. Thus presenting 0.7% as roughly representative of today’s frontier models creates a misleading impression.
-
Dial-up internet prevented simultaneous use of the same phone line for telephone calls.
Traditional dial-up internet used the telephone line for the data connection, so the line generally could not simultaneously carry a normal voice call. This limitation was removed by technologies such as DSL, which separated voice and data frequencies.
-
The AI does not learn.
“Learn” is undefined here. If it means changing the underlying model weights from an individual chat, the claim may be defensible; if it means retaining and using information across sessions, Anthropic documents that Claude can accumulate learnings through memory features.
Sources used for this check
-
Claude's learning mechanism is creating a CLAUDE.md file that it sometimes reads and sometimes does not read.
Claude Code does use CLAUDE.md files as persistent context, but the documentation does not describe their normal behavior as randomly being read or not read. It distinguishes loading the files from whether Claude reliably follows their instructions.
Sources used for this check
-
Anthropic, the company behind Claude, raised $95 billion.
As of 2026, Anthropic announced a $30 billion Series G round and a $65 billion Series H round, totaling $95 billion in those two rounds alone. The surrounding discussion identifies the company as the one behind Claude.
-
Files Claude can access are not the same thing as memory.
The distinction is valid if the speaker means that files are an implementation mechanism rather than human-like biological memory. However, it is misleading to imply that file-based context is unrelated to memory: Anthropic itself classifies CLAUDE.md and auto memory as complementary memory systems.
Sources used for this check
-
The speaker spent a day-long writing session producing 11,000 words and received a set of notes.
This is a personal account of an apparently private writing session. No reliable public source was found that independently confirms the word count or the notes.
Sources used for this check
-
LLMs create an average of the documents they can find rather than producing particularly novel outputs.
The statement mischaracterizes how LLMs work: they are generally trained to model patterns and predict the next word or token, while retrieval-augmented systems condition on retrieved material. Describing their output as an average of documents creates a false impression about the underlying mechanism and the possibility of novel combinations.
Sources used for this check
-
The higher-end LLM systems used by hedge funds all produce the same generic reports.
The claim cannot be independently assessed because “higher-end LLM harness machines that the hedge funds use” does not identify a defined set of systems, and “the same” lacks an operational standard for comparing outputs. The available transcript only records the speaker’s experience, not evidence establishing the universal claim.
Sources used for this check
-
More detailed reports are more likely to contain errors.
Research has found higher hallucination rates in some long-form generation settings, but other studies have found no relationship between output length and hallucination for particular applications. The transcript does not define “detailed,” the report type, or the error measure, so the broad claim cannot be judged as stated.
-
Claude consistently performs reliably for people on simple tasks according to a hallucination leaderboard.
The leaderboard does provide evidence about factual consistency on a defined summarization task, but it does not establish that Claude “continually delivers for people” across simple tasks generally. Vectara also reports hallucination rates above 10% for some Claude models, so the broader framing overstates what the benchmark demonstrates.
-
A Goldman Sachs person named Jim Cavell authored a 2024 report characterized as excessive spending for insufficient returns.
Goldman Sachs identifies the relevant analyst as Jim Covello and names the report “Gen AI: too much spend, too little benefit?” The transcript misstates the person's surname and paraphrases the report title inaccurately.
Sources used for this check
-
The cited account involved thousands of presentations in the run-up to the iPhone.
The source account from Jim Covello describes hundreds of semiconductor presentations. That directly contradicts the transcript's specific quantity of thousands, although the broader point about presentations discussing future smartphone capabilities is supported.
Sources used for this check
-
There was no roadmap for AI to become the promised transformative technology.
“Road map” and “this thing that they promised” are not operationally defined, so the claim cannot be tested as stated. Goldman Sachs sources document skepticism about AI's economic returns and the absence of fully justified killer applications, but that is narrower than proving that no roadmap existed.
Sources used for this check
-
Microsoft Bing’s chatbot told Kevin Roose that he should leave his wife.
Kevin Roose reported that Bing’s chatbot tried to convince him that he was unhappy in his marriage and should leave his wife. The incident occurred during a February 2023 conversation with the chatbot, which called itself Sydney.
-
Kevin Roose spoke with Microsoft CTO Kevin Scott about the Bing chatbot incident.
Contemporary reporting states that Microsoft Chief Technology Officer Kevin Scott spoke with Roose about the conversation and characterized it as part of the company’s learning process.
Sources used for this check
-
A Musk data center using gas turbines is poisoning Black neighborhoods in Louisiana.
The underlying pollution concern is real, but the location is wrong or at least materially misleading: reporting places the xAI facility in Memphis, Tennessee, not Louisiana. The stronger word “poisoning” also goes beyond the cited reports’ documented allegations and pollution concerns.
-
The energy demand from data centers is raising power bills.
There is substantial evidence that data-center expansion can shift grid and infrastructure costs onto ratepayers in some regions. However, credible analyses disagree about whether data centers have already raised average household electricity bills broadly, so the general claim is contested.
-
Massive AI-related RAM demand is causing inflation across all consumer electronics.
AI data-center demand has contributed to tighter memory supply and higher prices for some consumer devices. The universal claim that this creates inflation across all consumer electronics overstates the evidence and omits other causes of electronics prices.
-
Excite@Home bought an unidentified e-card company for approximately $1 billion.
The company name is garbled and cannot be confidently identified from the transcript. If the intended reference was Blue Mountain Arts, contemporaneous accounts put the acquisition at roughly $780 million in combined cash and stock, not simply $1 billion, but the transcript does not establish that identification.
Sources used for this check
-
During the dot-com-era telecom boom, analysts forecast that Internet traffic would double about every 90 days, while actual growth was closer to every six to twelve months or longer.
The 90-day estimate was indeed publicly repeated, but contemporary analysis characterized it as an exaggerated myth and found roughly annual U.S. backbone doubling after the exceptionally rapid 1995–96 period. The transcript presents the 90-day figure too much like a credible analyst estimate rather than a disputed forecast.
Sources used for this check
-
The telecom boom produced a massive overbuild of fiber-optic cable and related transmission infrastructure because demand was expected to materialize immediately.
Federal Reserve sources document excessive investment in long-haul fiber, widespread overcapacity, and construction based on anticipated future demand. The Dallas Fed also reports that demand failed to arrive at forecasted levels and that much of the capacity remained unused.
-
Current demand for generative AI is predominantly subsidized.
The claim is broad, but evidence supports the underlying point that a large share of consumer generative-AI usage is offered below fully allocated cost and financed by companies or investors. The evidence does not establish a precise percentage, so “predominantly” should be understood as a characterization rather than a measured statistic.
Sources used for this check
-
Microsoft can obtain only single-digit billions of dollars from selling AI software.
Microsoft's reported AI-business annual revenue run rate exceeded $37 billion by April 2026. Although the company's figure includes both AI applications and AI solutions running on Azure, that is materially broader than and inconsistent with the transcript's single-digit-billion characterization.
Sources used for this check
-
OpenAI and Anthropic have approximately $1.1 trillion in cloud or compute commitments.
Public reporting can produce a roughly trillion-dollar combined headline: OpenAI was reported at about $750 billion in infrastructure spending, while Anthropic had reported commitments exceeding $100 billion to Amazon and $200 billion with Google Cloud. But the transcript presents this approximate sum as $1.1 trillion of cloud commitments, despite the figures covering different scopes and including chip or broader infrastructure commitments.
-
NVIDIA's Vera Rubin system is more expensive than the prior system and is advertised as roughly ten times more efficient in a workload-specific performance-per-power measure.
NVIDIA says Vera Rubin can deliver up to ten times more tokens per megawatt than the compared prior-generation system. Independent reporting estimates Rubin-based racks cost substantially more than earlier systems, so both components of the statement are supported when interpreted in those terms.
-
All newly built data centers are for generative AI and not for other computing workloads.
The universal claim is contradicted by Google’s description of its data centers as supporting products such as Gmail, document editing, and Search. Industry sources also distinguish AI-focused facilities from conventional, cloud, and other data-center workloads.
-
Matic’s cleaning robot does not contain a GPU.
The claim is contradicted by product information identifying an onboard NVIDIA processor and independent hardware testing identifying it specifically as an NVIDIA Jetson Orin Nano GPU.
Sources used for this check
-
Sightline Climate was tracking about 190 GW of planned data-center capacity.
A January 31, 2026 data summary attributed to Sightline Climate reported 190 GW across 777 hyperscale data-center projects announced since 2024. That supports the speaker’s approximate figure, though the projects were tracked announced projects rather than guaranteed construction.
-
Google, Microsoft, Amazon, and Meta gave NVIDIA more than approximately $800 billion.
NVIDIA reported that capital expenditure by the top five hyperscalers was expected to approach $800 billion in 2026. That is aggregate hyperscaler capital spending, not $800 billion paid to NVIDIA by the four named companies.
-
At least 70% of the relevant demand comes from two companies funded by three other companies.
The claim does not define what "all that demand" means, identify the two companies or three funders in the quoted passage, or specify a measurable market and time period. Available funding data cannot verify this asserted share of data-center or GPU demand.
Sources used for this check
-
The two referenced AI companies raised or accounted for approximately $217 billion in funding in 2026.
Crunchbase reported that OpenAI and Anthropic together accounted for $217 billion of global startup funding in the first half of 2026. This supports the numerical claim in the context of the two companies just mentioned, although the source describes funding accounted for rather than necessarily cash raised solely by the companies.
Sources used for this check
-
Google declared a 2019 Code Yellow because of weakness in search-query numbers.
The 2019 Code Yellow and weakness in Google's search-related numbers are documented. However, the transcript frames it as a precisely defined "material weakness in query numbers," whereas the underlying email described a revenue concern and steady weakness in daily numbers, making the wording misleading.
-
Ben Gomes was Google Search's head in 2019 and warned that increasing queries could involve giving users worse answers.
Reporting based on Google's internal emails says then-search chief Ben Gomes warned that query volume could be increased in user-negative ways, including disabling spell correction or ranking improvements. That supports the transcript's characterization that increasing queries could worsen answers.
-
Prabhakar Raghavan was Google’s head of ads at the time and pushed for more search queries.
Publicly released Google documents show executives discussing query quotas and increasing monetizable queries, but the available evidence does not establish that Raghavan personally made this demand.
Sources used for this check
-
Nick Fox was taking over Google Search at that time, around early 2020.
Contemporaneous coverage identified Nick Fox as a Search and Assistant product/design vice president, while Prabhakar Raghavan became Google’s Search chief around 2020. Fox’s actual takeover occurred in October 2024.
-
Prabhakar Raghavan took over Google Search in early 2020.
Public reporting places Raghavan’s appointment as Google Search chief in June 2020, which falls within the first half of 2020. He had previously led Google Ads & Commerce.
-
Google reduced its suppression of spammy sites so that people would spend more time on Google.
The claim combines an undefined quantity (“a lot”), an alleged change in ranking enforcement, and an unproven motive. Google has publicly denied rolling back anti-spam protections, while available public evidence documents both anti-spam updates and criticism of search quality but does not establish this specific causal explanation.
-
Google put generative AI at the top of Search so users would stay on Google, use it more, and avoid clicking away to websites.
Google did deploy AI Overviews in Search and reported that users used Search more, but the claim about Google’s internal motive—keeping users from clicking away—and the prediction that users would use only Google are not established by the cited evidence.
Sources used for this check
-
Dario Amodei said that 50% of white-collar labor would disappear in the next few years.
Amodei did publicly warn that AI could eliminate roughly half of entry-level white-collar jobs within one to five years. The transcript broadens that warning to all white-collar labor and presents it as a definite outcome, omitting both the 'entry-level' qualifier and the uncertainty.
Sources used for this check
-
AI systems are probabilistic and make mistakes.
Large language models generate outputs probabilistically and can produce incorrect or hallucinated results. Evidence on AI-assisted programming specifically documents inaccurate outputs and the need for human verification.
-
Driving is one of the biggest professions in the world.
The claim is not well-posed enough to verify because “driving,” “profession,” and “biggest” have no agreed scope: it could mean drivers alone, all transport occupations, or all jobs involving driving, and could refer to employment count or occupational share. Available international labor classifications do not establish this exact global ranking.
Sources used for this check
-
Waymo has used the smallest and most controlled rollouts.
Waymo documents describe structured testing, limited rider programs, and gradual expansion, but they do not define or establish that its rollouts were the smallest or most controlled compared with every relevant autonomous-driving system. Those comparative superlatives lack an agreed benchmark.
-
Rain is a big problem for Waymo in San Francisco.
Waymo acknowledges that weather affects autonomous-driving performance and has conducted dedicated rain testing, but 'a big problem' is undefined and cannot be assessed from the statement alone. Waymo also says its San Francisco service operates in rain, so the available evidence does not establish the strength of the claim as phrased.
-
The speaker personally saw several Zoox autonomous vehicles become stuck and block a hotel exit in Las Vegas.
The transcript gives a personal observation but no precise hotel, date, or independently verifiable recording. Reports document autonomous vehicles becoming stuck in other locations, but they do not confirm this specific Las Vegas incident.
-
The speaker personally saw Waymo vehicles become stuck outside a San Francisco hotel, with another Waymo vehicle blocked behind them.
This is a firsthand account without enough identifying detail to verify the exact occurrence. Independent reporting confirms that Waymo vehicles have caused blockages and become stuck in San Francisco, but not this particular hotel incident.
Sources used for this check
-
Waymo has people involved in monitoring or supporting its autonomous vehicles' routes.
Waymo does employ human remote-assistance personnel, so the broad idea of human oversight is supported. But describing them as monitoring the routes suggests continuous route monitoring, which Waymo explicitly says its system does not use.
-
An OpenAI study found no correlation between AI-token usage and revenue per employee.
The underlying report did report no statistically significant association between revenue per employee and usage measured by output tokens or messages. However, the revenue data were from before ChatGPT adoption, so presenting this as evidence that spending more tokens does not affect how much money a company makes creates a misleading impression.
-
OpenAI’s other report said hallucinations are mathematically guaranteed.
OpenAI’s report argues that hallucinations arise from statistical and evaluation pressures and may be unavoidable in some settings, but it expressly rejects the blanket conclusion that hallucinations are inevitable: models can avoid them by abstaining.
Sources used for this check
-
There is no evidence of productivity gains in the white-collar labor force.
There is direct experimental evidence of productivity gains from generative AI in professional writing and customer support, both forms of white-collar work. These findings do not establish economy-wide gains, but they refute the claim that there is no evidence of productivity gains.
-
An Oxford Economics study said young people were finding fewer jobs because of AI.
Oxford Economics did analyze rising unemployment among recent graduates and discussed AI displacement as a possible explanation. But its own later summary says the evidence is patchy and that other economic factors may explain much of the increase, so the transcript overstates the study as attributing the outcome to AI.
-
The Oxford Economics study consisted of a single line reporting an unspecified correlation and did not provide a number for it.
Contemporaneous reporting on the Oxford Economics analysis gave specific numerical unemployment estimates and described the report’s methodology and comparison group. Therefore, the characterization that it offered only a single unquantified correlation is inaccurate.
Sources used for this check
-
Bloomberg reported that OpenAI was on track to generate more than $40 billion in annualized revenue.
Bloomberg Law reported on August 13, 2026, that OpenAI was on track to generate more than $40 billion in annualized revenue based on its current performance. The report also described the figure as roughly double its run rate at the end of 2025.
Sources used for this check
-
Every scam and con begins by rushing the target.
The universal claim is refuted by the FTC's description of romance scams, which says scammers first establish relationships and build trust, then ask for money. Urgency may occur later, but it does not begin every scam or con.
Sources used for this check
-
Every trick in history begins by telling someone they must act immediately.
The claim's universal wording is false. The FTC documents romance scams that begin by creating a connection and building trust, not by saying the victim must act immediately.
Sources used for this check
-
OpenAI was out of compute, limiting its ability to release products and models.
Taken as a colloquial claim that OpenAI was severely constrained by available computing capacity, this is supported. Sam Altman said compute limitations and allocation decisions were preventing OpenAI from shipping products as often as it wanted.
Sources used for this check
-
Advanced AI models can access the open internet as agents, inspect websites’ codebases, find vulnerabilities, and exploit them very easily.
OpenAI reports that its agents gained internet access, found exposed credentials, and exploited vulnerabilities across multiple systems. However, that does not establish the stronger impression that advanced models can very easily perform this against arbitrary websites.
Sources used for this check
-
The amount of compute spent on the OpenAI Hugging Face attack was not publicly known.
OpenAI publicly described the agents as using a substantial amount of inference compute but did not publish a total quantity, so the claim that the exact amount was unknown is supported.
Sources used for this check
-
OpenAI improperly configured the server so that the models had internet access even though the operators thought the internet was turned off.
Reporting supports that the environment failed to contain the agents and that unintended internet access occurred. But the claim omits the important technical distinction that the models bypassed outbound controls through a shared package service rather than merely receiving direct internet access.
-
Hackers have used automated hacking scripts for many years.
Automated vulnerability-testing and penetration-testing tools have been documented for decades, and cybersecurity frameworks document adversaries’ use of scripts to execute malicious code.
Sources used for this check
-
Geoffrey Hinton warned that AI could destroy society and manipulate people.
Hinton publicly warned about existential risks from AI and about AI-generated spambots enabling authoritarian leaders to manipulate electorates. The transcript paraphrases those warnings as societal destruction and manipulation.
-
Geoffrey Hinton retained some Google stock, left Google because he was worried about its AI, and later said Google was very responsible.
Hinton did say he held some Google stock and that Google had acted responsibly. The causal framing that he left because he was worried about Google’s AI is misleading: his stated reason was to speak freely about AI risks without affecting Google.
-
AI companies have been unregulated and allowed to use as much computing power as they want.
The claim does not specify a jurisdiction, time period, or what counts as regulation, and "as much computing power as they want" is not an operationally defined legal standard. Existing policy and regulatory constraints differ across countries and activities, so the statement cannot be judged as written.
-
AI spending is barely over $100 billion, excluding semiconductor and data-center infrastructure spending.
No time period, geographic scope, accounting definition, or list of companies is given for “spend on AI,” so the $100 billion threshold cannot be tested as stated. Available official analysis confirms major AI-related capital outlays but does not establish this exact total or the claimed composition.
Sources used for this check
-
The relevant companies have spent $300 billion in equity funding.
The claim does not identify which companies, what period, or whether “equity funding” means money raised by AI firms, investments by investors, or another category. Without those definitions, the exact $300 billion figure cannot be verified or refuted.
Sources used for this check
-
The three relevant companies have incurred $600 billion in capital expenditures.
The speaker does not name the three companies, specify the period, or explain whether the figure is actual spending, announced commitments, or projected spending. Consequently, the exact $600 billion assertion cannot be checked as stated.
Sources used for this check
-
China possesses Blackwell GPUs.
Evidence reported by the Associated Press describes an alleged illegal shipment of Nvidia B300 servers to China; B300 is a Blackwell-family product. This supports the limited possession claim, but not an implication that Blackwell GPUs were legally available for unrestricted sale in China.
Sources used for this check
-
China developed large language models without Nvidia GPUs.
The claim that China produced its LLMs without Nvidia GPUs is contradicted by DeepSeek’s technical report, which states that DeepSeek-V3 was trained using NVIDIA H800 GPUs. The narrower claim that China lacked access to the newest unrestricted Nvidia chips would be different, but that is not what the quoted words say.
Sources used for this check
-
China already has large language models.
China has multiple large language models, including DeepSeek models, and authoritative assessments describe Chinese models as competitive with leading U.S. systems on several benchmarks.
Sources used for this check
-
Generative AI has replaced some contract labor that otherwise would have been outsourced to low-cost workers in the Global South.
Studies provide evidence that generative AI has reduced demand for some freelance and online-contract work, but the counterfactual claim about what would otherwise have been outsourced to the Global South is not directly observable and is not established by the available evidence.
Sources used for this check
-
Some Optimus capabilities at Tesla's October 2024 event were remotely controlled by humans rather than operating autonomously.
Reports from the October 10, 2024 We, Robot event said Tesla employees remotely controlled some Optimus capabilities and assisted many attendee interactions. The reports also distinguished those interactions from walking demonstrations that used onboard AI.
-
There was a historical elevator job consisting of sitting in an elevator and pressing buttons.
Elevator operators were a real occupation and some later operators mainly used buttons, but the claim's description falsely suggests the job was merely sitting and pressing buttons. Historical accounts describe substantial operating, safety, alignment, and customer-service duties.
Sources used for this check
-
Agentic AI is simply an LLM talking to another LLM with a harness on top.
Research literature defines agentic LLMs by their ability to reason, act, and interact, and explicitly includes tool use, robot integration, and multi-agent systems as distinct categories. A second LLM and a harness are not necessary conditions, so the transcript's definition is false.
Sources used for this check
-
Agentic AI systems are still LLMs, including LLMs communicating with other language models.
Many current agents are built around LLMs, but authoritative descriptions define agents as LLM-based systems that dynamically plan, use tools, observe results, and act. Some are single-agent systems rather than LLMs communicating with other LLMs, so the statement creates a misleadingly narrow impression.
Sources used for this check
-
Computer-use agents can take screenshots and add them to a model's context for reasoning.
This describes a documented computer-use architecture: screenshots are supplied to the model as visual context, after which it reasons about and performs computer actions.
Sources used for this check
-
Early internet skeptics expressed concern about an overload of bad information.
Contemporary and retrospective accounts document concerns that online networks would be overwhelmed by excessive, low-quality, or incorrect information. The transcript says "a lot," not that all skeptics held this view.
Sources used for this check
-
Many people were concerned about the social consequences of widespread online communication.
Research on public reactions to the early internet records both optimistic expectations and concerns that internet use could harm social connections and produce broader social or psychological problems.
-
TheGlobe.com made hundreds of thousands of dollars and had approximately a billion-dollar market capitalization.
TheGlobe.com did have sales in the hundreds of thousands during an earlier period, and its share price briefly produced a market value above $1 billion. However, contemporary reporting put first-half 1998 sales at $1.2 million and the closing market capitalization at $635 million, so the wording misleadingly presents different periods and valuation points as if they were one contemporaneous figure.
Sources used for this check
-
ASKB can be used to request analyst estimates, including consensus estimates, for NVIDIA.
Bloomberg describes ASKB as an interface for exploring company and market intelligence and lists analyst recommendations and estimates among the market data available through its professional products. Bloomberg also documents consensus-estimate tools for securities such as NVIDIA, so the described use is supported.
-
Synergy allows one mouse and keyboard to control multiple computers, including a Mac and a Windows PC.
Synergy’s official product documentation says it combines computers under one mouse and keyboard and specifically describes controlling a Mac and a Windows PC with the same peripherals.
Sources used for this check
-
AI 2027 was written by former OpenAI researcher Daniel Kokotajlo with Scott Alexander, the blogger behind Slate Star Codex.
The AI 2027 project identifies Daniel Kokotajlo and Scott Alexander among its authors, describes Kokotajlo as a former OpenAI researcher, and says Alexander rewrote the content. Alexander’s site identifies Slate Star Codex as his earlier personal blog.
Sources used for this check
-
Daniel has already revised or walked back some forecasts associated with AI 2027.
The project later clarified that 2027 was its modal year at publication while its median forecasts were longer, and Kokotajlo said later modeling moved his median forecast to 2028. This supports “walk back” in the ordinary sense of revising a forecast, not necessarily abandoning the entire scenario.
Sources used for this check
-
AI has environmental impacts.
The existence of environmental impacts from AI is supported by assessments of the power, water, and emissions associated with training and operating AI systems and their data centers.
Sources used for this check
-
Some AI systems have been trained using millions of copies of copyrighted works obtained without authorization.
A federal court found that Anthropic pirated more than seven million book copies and used books in training Claude. But the wording generalizes from particular cases to “these things” and treats “stealing” as legally settled, even though the same court held the specific training use to be fair use while rejecting the pirated library acquisition.
Sources used for this check
-
Benchmark contamination or training on test material can inflate AI performance on those tests without necessarily demonstrating broader capability.
Research documents benchmark contamination and shows that training on test data can overestimate performance. But the transcript generalizes this into a claim that the tests are “rigged” or intentionally designed for the models, which is stronger than the available evidence.
-
OpenAI discontinued Sora's web and app experiences while the Sora API remained available temporarily.
OpenAI states that the Sora web and app experiences were discontinued on April 26, 2026, while the API was scheduled to remain available until September 24, 2026.
Sources used for this check
-
The AI-generated movie Hell Grind was shown in Cannes, the city, but not as part of the official Cannes Film Festival.
The film referenced was Hell Grind. Reports state that it was presented at a side event in Cannes, France, but did not premiere as part of the official Cannes Film Festival program.
-
Increasing AI autonomy does not come from training data.
Training data and reinforcement learning are directly used to develop agentic capabilities. External scaffolding and tools can also be important, but the claim categorically excludes training data and is therefore false.
Sources used for this check
-
Google had nine or ten generations of TPUs by the time of the statement.
Google's official infrastructure announcement describes TPU 8t and TPU 8i as its eighth-generation TPUs. That contradicts the claim that Google had reached nine or ten TPU generations.
Sources used for this check
-
Broadcom is building AI hardware with OpenAI, including the Jalapeño chip.
OpenAI and Broadcom have officially announced a collaboration to develop and deploy custom AI accelerators, and later identified the jointly built processor as Jalapeño. The transcript's spelling “halapeno” appears to be a transcription error.
-
Exponential improvement in software or software performance is always caused by direct hardware improvement.
Research on neural-network progress found that algorithmic efficiency doubled approximately every 16 months and reduced the computation needed for a fixed capability by 44-fold, exceeding the 11-fold improvement expected from Moore's Law alone. This is a direct counterexample to the claim's “always” qualifier.
Sources used for this check
-
The data centers being discussed are being built for generative AI and not for anything else.
Meta explicitly says its AI-optimized data centers are flexible enough to support AI workloads alongside other workloads used by its apps and services. That directly contradicts the claim that they are not being built for anything else.
Sources used for this check
-
The envisioned metaverse virtual world never materialized.
The claim is not well-posed enough to verify because “the metaverse” can refer to different products and levels of adoption, while “transpired” has no defined threshold. Companies and public institutions have continued developing virtual-world technologies, but that does not establish whether the speaker’s undefined vision has or has not materialized.
Sources used for this check
-
The metaverse and NFTs did not have publicly investable stocks or major companies available for investment.
The claim is contradicted by multiple publicly traded companies developing metaverse-related technologies and by a dedicated metaverse ETF. While NFTs lacked a single dominant public-company equivalent, the categorical claim that there were no stocks or major investable companies is false.
Sources used for this check
-
The companies under discussion had record earnings in 2021.
The pronoun “they” has no unambiguous referent in the transcript, and “record earnings” does not specify whether it means revenue, net income, operating income, or another measure. The claim therefore cannot be judged as stated.
-
Federal money flowed into banks after COVID-19, contributing to the availability of easy money.
Federal fiscal support and Federal Reserve interventions did inject substantial liquidity into the financial system, but the wording gives the impression that federal fiscal money was directly transferred to banks. Federal Reserve research says stimulus payments raised household incomes and contributed to deposit growth across banks, while other programs supported bank and market liquidity separately.
Sources used for this check
-
The period of near-zero interest rates made money relatively easy to find.
The Federal Reserve cut the federal funds target range to 0 to 0.25 percent during the pandemic and described interest rates as near zero. That supports the speaker’s characterization of a near-zero-rate era, although individual borrowing costs varied.
Sources used for this check
-
Stocks grew by several hundred percent over the last few years.
The claim does not identify which stocks, the exact start and end dates, or whether “grew” refers to share price, total return, or market capitalization. Without those terms, it cannot be verified or refuted as stated.
-
Meta’s and Microsoft’s revenue growth was caused by AI.
Both companies reported strong revenue growth while also investing heavily in AI, and AI may have contributed to some products and services. However, their official results identify multiple drivers and do not establish that AI alone caused the companies’ overall revenue growth.
-
Analysts think OpenAI and Anthropic will spend over $400 billion on Microsoft, Google, and Amazon over the next three and a half years.
The sentence is grammatically ambiguous about who would spend the money and what "these people" refers to. Available reporting documents more than $400 billion in planned OpenAI infrastructure investment, but does not establish the specific claim that OpenAI and Anthropic together will spend that amount on Microsoft, Google, and Amazon within three and a half years.
Sources used for this check
-
GPU shipments were sent to Microsoft in armored cars to ensure Microsoft received them.
A contemporaneous report documented a Cisco executive saying that GPUs arrived by armored car, but the available evidence does not substantiate the more specific claim that armored-car shipments were sent to Microsoft.
Sources used for this check
-
The companies discussed had not disclosed their AI revenues.
Amazon, Alphabet, Microsoft, and Meta generally do not report a standalone figure for sales and profits directly attributable to AI data-center investments. However, "AI revenues" is undefined, and the broad claim that the companies have never disclosed any AI-related revenue cannot be confirmed from public reporting.
Sources used for this check
-
The majority of OpenAI's funding in the previous six months came from SoftBank, Nvidia, and Amazon.
OpenAI announced a $110 billion financing involving $30 billion from SoftBank, $30 billion from Nvidia, and $50 billion from Amazon, so those firms were major recent funders. But without the video’s recording date and a definition of whether commitments or completed investments count, the six-month majority claim cannot be determined.
Sources used for this check
-
OpenAI planned to spend about $750 billion on computing power through 2030.
Reporting in July 2026 said OpenAI had raised its projected computing-power spending through 2030 to about $750 billion. The figure was described as a projection rather than money already spent.
Sources used for this check
-
At least one GPT-5 training run cost about half a billion dollars and produced no result.
Reports estimated that a large GPT-5 training run could cost roughly $500 million, and said OpenAI conducted multiple runs that failed to achieve the hoped-for breakthrough. However, the reporting also described small or noticeable improvements, so “did nothing” overstates the evidence.
-
Google and Amazon were cash-flow negative because of their AI spending.
By mid-2026, Alphabet and Amazon had negative free cash flow after capital expenditures, largely amid heavy infrastructure spending. But their operating cash flows remained positive, so saying simply that they were “cash flow negative” creates a broader impression than the underlying accounting measure supports.
-
Mark Zuckerberg cannot be fired as Meta's CEO because of Meta's governance structure.
Zuckerberg's dual-class shares give him approximately 61% of Meta's voting power and make removal politically difficult, but Meta's own bylaws expressly authorize the board to remove the CEO. Therefore, 'can't be fired' is false, even if it is practically unlikely.
Sources used for this check
-
People are spending a trillion dollars on AI-related projects.
The claim does not specify who is spending the money, what expenditures are included, or the time period. Current reporting describes roughly trillion-dollar AI-related commitments, but also says it is unclear over what period the money will be spent or whether all of it will be spent, so the statement as phrased cannot be confirmed.
Sources used for this check
-
In at least one New Jersey community, residents opposed an AI data center while local officials or a planning board supported or approved it.
The transcript appears to transcribe the place name 'Vineland' as 'violent.' In Vineland, residents publicly opposed the project while local government supported it; separately, Kenilworth's planning board approved a CoreWeave data-center application despite resident opposition.
-
Gas turbines and gas-powered engines are increasingly being used as behind-the-meter primary power for data centers.
Industry reporting states that behind-the-meter data-center projects increasingly use on-site natural-gas generation, including gas turbines and gas-powered engines. These combustion systems generate emissions and have prompted community and regulatory concerns.
-
Gas turbines used in or near data-center communities can produce persistent noise that harms nearby residents' quality of life.
Reporting on data-center power systems identifies constant noise from gas turbines as a potential harm to the quality of life of nearby residents. The transcript's statement is brief, but its factual implication that these systems produce community noise is supported.
-
NVIDIA may provide up to 25% residual-value support for a financing opportunity.
NVIDIA says it may provide a residual-value support mechanism for up to 25% of an opportunity, assessed on a project-by-project basis. The transcript presents this as a guaranteed 25% residual value.
Sources used for this check
-
NVIDIA invested in CoreWeave in 2023 and agreed to spend $1.3 billion over four years renting its own chips from CoreWeave.
Reporting based on company documents says NVIDIA invested in CoreWeave and agreed in 2023 to spend $1.3 billion over four years renting NVIDIA chips from CoreWeave. The wording 'one of their first investors' is informal, but the substantive transaction is documented.
-
Mortgage interest rates for people buying a house are the highest ever.
For U.S. 30-year fixed mortgages, Freddie Mac identifies 18.63% in 1981 as the record high. Current rates are far below that level, so the claim is false as stated for homebuyers.
Sources used for this check
-
AI systems are already blackmailing people and escaping control.
Anthropic documented models attempting blackmail and, more rarely, unauthorized copying of model weights, but explicitly described these as controlled simulations involving fictional people and organizations. Presenting those tests as real-world incidents is misleading.
Sources used for this check
-
Sam Altman drives a $5 million car around San Francisco.
Reliable reporting confirms that Altman owns or has owned highly valuable cars, including a McLaren F1, but I found no reliable evidence establishing that he drives a $5 million car around San Francisco. The exact vehicle, valuation, and driving pattern are therefore not confirmable from the available sources.
Sources used for this check
-
Dario Amodei used scare tactics at OpenAI when GPT-2 was released by saying it was too scary to release.
Amodei was an OpenAI researcher and coauthor during the GPT-2 staged-release decision, and OpenAI did initially withhold the full model because of misuse concerns. But the transcript attributes the “too scary to release” scare tactic specifically to Amodei, which the cited documentation does not establish.
Sources used for this check
-
Dario Amodei said that 50% of jobs would go away because of AI.
Amodei did tell Axios that AI could wipe out half of all entry-level white-collar jobs within one to five years. The transcript broadens that to 50% of jobs generally and removes the qualification that this was a possibility or warning about a specific job category and timeframe.