Contradicted claims, page 2 of 2
Claims by Gary Marcus and Ed Zitron. Deadlines are judged as of 2026-10-29.
Contradicted claims (150)
The dated evidence goes against what the claim committed to, under the criteria frozen before grading.
Page 2 of 2: claims 101 to 150 of 150, ordered by the date first said, newest first. To find a claim by its words, use search.
Assessment: ContradictedSaid
“Almost every one of these companies is claiming that these systems "reduce drudgery" or "enhances productivity," yet there's no real explanation as to how it might do so or what the drudgery is that's being relieved, nor any unassailable metrics on how productivity improves.”
What happened: This identifies both the work involved—writing replies—and the proposed mechanism—generating suggested responses and revisions. It directly contradicts the absence of a concrete explanation, without proving the size of any productivity gain. (ServiceNow, )
Assessment: ContradictedSaid
“That being said, we're yet to see this actually happen, outside of OpenAI's alleged $3.4 billion of yearly revenue (it still loses $5 billion a year), and generative AI's multi-billion dollar contribution to Amazon Web Services revenue over an indeterminate amount of time.”
What happened: Microsoft publicly attributed eight percentage points of Azure growth to AI services before the post. The same call describes organizations using Azure's model services, making this a counterexample under the ordinary reading of customer AI integrations. (Microsoft, )
Assessment: ContradictedSaid
“Nobody is trying to make these things more efficient, or at the very least nobody has succeeded in doing so.”
What happened: The developers report serving the existing LLM traffic with half as many GPUs after changing the serving software. This is a concrete resource-efficiency improvement before publication. (vLLM project, UC Berkeley researchers, )
Assessment: ContradictedSaid
“The prices are not decreasing, the software is not becoming more useful, and the "next generation" model that we’ve been hearing about since last November has turned out to be a dud.”
What happened: OpenAI announced lower input and output token rates for the updated GPT-3.5 Turbo offering. This directly contradicts the broad claim that model prices were not decreasing. (OpenAI, )
Assessment: ContradictedSaid
“In the new form, SB 1047 can basically only be used only after something really bad happens, as a tool to hold companies liable. It can no longer protect us against obvious negligence that might likely lead to great harm.”
What happened: The sponsor's official description expressly retains enforcement before harm where public safety faces an imminent threat. It also confirms the narrower point that enforcement was restricted. (Office of California Senator Scott Wiener, )
Assessment: ContradictedSaid
“No private startup can actually afford to build and scale a Large Language Model without the help of Google, Amazon or Microsoft — and thus they would, assuming they actually became useful, effectively control every single product involving this technology.”
What happened: The paper documents DeepSeek-V2's 236-billion-parameter model, substantial pretraining, and deployed inference on H800 hardware. Together with the independently identified High-Flyer resources in e2 and e3, this supplies a counterexample already operating before the post. (DeepSeek-AI, arXiv, )
Assessment: ContradictedSaid
“The promises of artificial intelligence will not be kept through the pursuit of generative AI, as generative AI is a dead-end technology that has peaked, one that costs too much — both financially and socially — to keep it going.”
What happened: The benchmark operator reports verified scores of 37.6% for Opus 4.5 and 54% for a Gemini-based refinement, alongside disclosed methods and open-source solutions. These material advances contradict a categorical capability ceiling, while the report also identifies costs, narrow task coverage and possible overfitting. (ARC Prize Foundation, )
Assessment: ContradictedSaid
“Along with $3.5 billion from Saudi Arabia’s Public Investment fund and over $8 billion from Softbank, the latter a secondary market sale, it also raised (equity and debt financing respectively) from Goldman Sachs and Morgan Stanley — two parties I do not believe are going to be willing to subsidize generative AI.”
What happened: Both named banks participated in OpenAI's executed credit facility, demonstrating willingness to finance the business. The report's commercial interest terms do not establish a concessionary subsidy. (CNBC, )
Assessment: ContradictedSaid
“I fully expect that the generative AI bubble will begin to burst within the next 12 months, for many reasons:”
What happened: EY reports $49.2 billion of first-half generative AI VC investment versus $44.2 billion for all of 2024. It also reports nearly 25% fewer transactions than in the preceding half-year, a material sign of concentration that qualifies the expansion. (EY Ireland, )
Assessment: ContradictedSaid
“The current approach has reached a plateau”
What happened: User comparisons collected June 10–25 show newer frontier models outperforming preceding models on image-containing conversations. Preference measurements are imperfect, but this is independent evidence of improvement within the target quarter. (LMSYS Org, )
Assessment: ContradictedSaid
“Even though it doesn't really automate anything.”
What happened: Intercom reported that deployed Fin systems resolved customer questions directly. The existence of successful bounded tasks contradicts 'anything', although the reported percentage requires qualification. (Intercom, )
Assessment: ContradictedSaid
“Here's the thing, I get that that might be what they're pushing toward, but generative AI isn't even automation at this point.”
What happened: Intercom reported actual customer deployments resolving queries directly through Fin. This supports the existence of bounded task automation, although the reported percentage is not an independently verified success rate. (Intercom, )
Assessment: ContradictedSaid
“To reach level 3 (whatever “Agents, systems that can take actions” means) you’d need a level of sentience dependent on entirely new forms of mathematics.”
What happened: The account describes Claude Sonnet 3.7 using search, notes, Slack, and price-setting tools over a month, with humans providing physical services. I infer from this documented implementation that sustained action-taking did not require the asserted new mathematical foundations. (Anthropic, )
Assessment: ContradictedSaid
“And yes, that sound you hear is the slow deflation of the bubble I've been warning you about since March.”
What happened: Their account of PitchBook data reports record generative-AI investment in Q2. The same article describes weaker US seed activity and cooler enthusiasm, so the contradiction concerns broad financial contraction. (Aspiring for Intelligence — Sabrina Albert and Vivek Ramaswami, )
Assessment: ContradictedSaid
“It cannot do much more than it is currently doing, other than doing more of it faster with some new inputs.”
What happened: The public evaluation reported 97.8% success on basic Blocksworld and 52.8% on its obfuscated version, where earlier tested LLMs remained below 5%. These are substantive planning gains; the event date denotes publication of the demonstration, since exact test-run dates are not supplied. (Valmeekam, Stechly, Gundawar and Kambhampati / arXiv, )
Assessment: ContradictedSaid
“It isn’t getting much more efficient.”
What happened: The paper reports stronger model performance with 93.3% less attention-cache memory than DeepSeek 67B. Its deployment tests report 5.76 times the generation throughput on an eight-H800 node using the earlier service's workload distribution, demonstrating substantial resource-efficiency gains. (DeepSeek-AI / arXiv, )
Assessment: ContradictedSaid
“Generative AI is hitting a wall, and any future improvements will be marginal, and cost billions to achieve.”
What happened: ARC Prize reports 75.7% and 87.5% for the o3 preview configurations and contrasts these with GPT-4o's 5% performance in 2024. The evaluator describes a substantial advance, while disclosing benchmark-specific training and increased inference computation. (ARC Prize, )
Assessment: ContradictedSaid
“While Google has desperately tried (and failed) to convince Wall Street that it’s building and selling the future”
What happened: Goldman Sachs's reported assessment after I/O is direct evidence that a prominent Wall Street institution accepted Google's AI leadership and product-integration pitch. The original analyst note was not accessible, so this assessment relies on reporting. (Yahoo Finance, )
Assessment: ContradictedSaid
“He doesn’t specify which actual restrictions in the proposal bill would “stifle AI models”.”
What happened: Ng explicitly names safety assessments and shutdown capabilities. In the next paragraph, he focuses on the hazardous-capability designation and potential downstream liability, then connects that burden to inhibited model development, especially open source. (DeepLearning.AI, )
Assessment: ContradictedSaid
“Since Rabbit is acting as a third-party, and doesn’t use a standard like Oauth or SAML, it needs to keep user credentials in their original form. This is just insanely dangerous.”
What happened: The report describes inspection during April 29–May 10; the event date marks that testing window's end. It found that service-issued session cookies, rather than entered usernames and passwords, were collected for storage. (Obscurity Labs, )
Assessment: ContradictedSaid
“Generative AI hallucinates because it doesn’t have a consciousness or any ability to learn or know anything.”
What happened: The experiments demonstrate usable abilities acquired through training and adaptation to examples. They contradict an absolute absence of learning under the operational reading, without establishing consciousness. (arXiv; Tom B. Brown and colleagues, )
Assessment: ContradictedSaid
“Large Language Models are — and always will be — incapable of reasoning or consciousness”
What happened: SATBench reports performance well above its 50% chance baseline on generated, solver-checked puzzles, without allowing external tools during model evaluation. Reasoning-trace checks and partial human validation support a limited inference capability, although they also reveal errors. (Anjiang Wei and colleagues, arXiv, )
Assessment: ContradictedSaid
“And it all kind of makes sense why generative AI doesn’t really help anyone other than those who want to sell something.”
What happened: The dated episode presents the hosts using an AI tour guide for museum descriptions and questions. Access to exhibits during personal recreation is a concrete non-sales benefit. (Eyes On Success, )
Assessment: ContradictedSaid
“Over a year ago in The Atlantic, I urged that we come up with laws and new technological responses to blunt the impact of AI-generated misinformation. So far almost nothing has been done.”
What happened: The ruling applied existing consent and disclosure protections to AI-generated voice calls and became effective on release. It was a concrete federal response relevant to deceptive impersonation calls, subject to statutory exemptions. (Federal Communications Commission, )
Assessment: ContradictedSaid
“As a result, every single LLM-based application — ChatGPT, Google Gemini, Anthropic's Claude, Meta's LLaMA, to name but a few — is trained on an indeterminately-large portion of the internet made up of both Common Crawl (a 250-billion page repository of web content) and what appears to be as much of the rest of the internet as they're able to download, legally or otherwise.”
What happened: The methods restrict StarCoderBase's training set to The Stack's GitHub repositories; StarCoder adds Python fine-tuning from that corpus. This documented training recipe does not include Common Crawl. (BigCode research collaboration, arXiv, )
Assessment: ContradictedSaid
“These models are not "learning" or, let's be honest, "training" on this data, because that's not how they work — they're using mathematics to plagiarize it based on the likelihood that somebody else's answer is the correct one.”
What happened: The documented objective is predictive model training. A probability over next tokens is not a probability that somebody else's complete answer is correct. (OpenAI, arXiv, )
Assessment: ContradictedSaid
“This is why driverless cars are still just demos, and why LLMs will never be reliable.”
What happened: The filing describes daily public driverless rides for work, school, errands and airport travel, and identifies Waymo One as a deployed commercial service. Those operations contradict the demonstration-only premise. (Waymo LLC; filed with the California Public Utilities Commission, )
Assessment: ContradictedSaid
“Artificial intelligence isn't new, these models aren't new, and after just over a year of hype, generative AI is yet to prove it can do anything substantially different other than helping tech company CEOs raise money and create headlines.”
What happened: The public research demonstration reported practical customer-support benefits beyond fundraising or publicity. The event date records when that demonstration became public, not when the underlying workplace observations occurred. (arXiv; Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, )
Assessment: ContradictedSaid
“In essence, the AI boom requires more high-quality data than currently exists to progress past the point we're currently at, which is one where the outputs of generative AI are deeply unreliable.”
What happened: SCoRe uses existing MATH and MBPP problems, generated correction traces, and correctness rewards. Its Gemini 1.0 Pro coding experiments improve self-correction on held-out HumanEval, demonstrating further error-correction progress without requiring an expanded stock of human-written material. (Google DeepMind researchers via arXiv, )
Assessment: ContradictedSaid
“ChatGPT can spin up piles of anodyne business copy, yet its outputs always require enough editing that it's questionable how much time you've actually saved.”
What happened: The randomized experiment involved 453 professionals completing occupation-specific writing assignments. Substantial savings on completed work, accompanied by higher assessed quality, contradict the blanket claim under the stated reading of business copy. (Science; abstract reproduced by Stanford SCALE, )
Assessment: ContradictedSaid
“In some ways we have made progress, but in the ones I keep harping on—factuality and compositionality—we have not.”
What happened: The report compares factual accuracy on TruthfulQA and reports gains over earlier models. This contradicts a categorical absence of factuality progress, subject to the company's evaluation methods and contamination caveat. (OpenAI via arXiv, )
Assessment: ContradictedSaid
“As a journal reviewer, I would also note that they reported neither the p value they obtained, nor the test of statistical significance they applied; in a competent peer-reviewed journal, that’s a no-no.”
What happened: The methods paragraph names one-sided t-tests and the Mann–Whitney check. The applied tests were therefore identified. (OpenAI, )
Assessment: ContradictedSaid
“It came in the middle of May 2022.”
What happened: The November 2021 compromise text explicitly mentions general-purpose AI and contains a dedicated Article 52a. This predates the claimed first mention in May 2022. (Council of the European Union, Slovenian Presidency, )
Assessment: ContradictedSaid
“I realize there’s been almost zero public disclosure of how any of the driverless car companies are using remote operators. It’s not just Cruise.”
What happened: Waymo publicly described human assistance while distinguishing it from vehicle control. The paper also explains requests for contextual information and directions to stop, so this was substantive disclosure of operators’ roles. (Waymo, published through arXiv, )
Assessment: ContradictedSaid
“He has spent hours of interviews vaguely mumbling about how the metaverse would change everything, only to give up on the idea the second that artificial intelligence became the new big thing.”
What happened: Meta's global affairs head reaffirmed the metaverse strategy in a briefing held inside its operating Horizon Workrooms product. The report documents continued pursuit during the claimed abandonment period. (Bloomberg, republished by The Business Times, )
Assessment: ContradictedSaid
“Essentially none of the fancy cognitive things GPT-4 is rumored to do stand up.”
What happened: The study tested new cognitive-reflection and Linda/Bill problems, finding strong GPT-4 performance across repeated trials. Figure 5 explicitly dates these experiments to March 2023, so the later publication bears on the contemporary model. (Communications Psychology / Europe PMC, )
Assessment: ContradictedSaid
“None of the top labs actually is remotely transparent, governance basically doesn’t exist yet, and there are basically no safeguards actually in place.”
What happened: Meta disclosed named data sources and their proportions, preprocessing, architecture, training settings, hardware, and benchmark results, including bias and toxicity assessments. These disclosures provide a substantive counterexample to the assertion that no leading laboratory offered meaningful transparency. (Meta AI / arXiv, )
Assessment: ContradictedSaid
“At present, anybody can release AI at whatever scale they like, with virtually no oversight, literally overnight.”
What happened: Article 20 requires security assessments for covered new deep-synthesis products, applications, and functions with public-opinion or social-mobilization capacity; Article 23 includes text generation and question-answer dialogue. The rules took effect January 10, 2023; event_date records their legal status on the assessment day, not their enactment. (Cyberspace Administration of China, )
Assessment: ContradictedSaid
“To my knowledge, tech companies are free to leverage a human gullibility around chatbot technologies however they like, without consequence”
What happened: This primary regulatory announcement reports an immediately effective restriction on Luka's processing of Italian users' data, citing transparency failures and risks to children and emotionally vulnerable people. It establishes a concrete constraint before Marcus's post, although it does not establish a general prohibition on simulated affection. (Garante per la protezione dei dati personali, )
Assessment: ContradictedSaid
“On top of that, these products are extremely expensive to run. ChatGPT burns millions of dollars a day in computing power, while charging a per-access-token price for companies to plug into its models.”
What happened: The preceding sentence explicitly identifies this as February research on the hardware supporting ChatGPT. Its estimated daily cost is below either US$1 million or US$2 million. (Computer Weekly, )
Assessment: ContradictedSaid
“The only thing keeping it from being even more toxic and deceitful than it already is, is a system called Reinforcement Learning by Human Feedback, and “OpenAI” has been very closed about exactly how that really works.”
What happened: OpenAI's confirmation describes a toxicity detector incorporated into ChatGPT and efforts to filter training material. These are protections beyond reinforcement learning; the report does not quantify their separate effects. (TIME, )
Assessment: ContradictedSaid
“It's worth asking since (a) Amazon is apparently greatly reducing Alexa's staff, (b) hasn't added significant features in years, and (c) seems clumsy and limited compared to GPT-3.”
What happened: The dated launch account describes multiparty interaction and speech directed at Alexa without repeatedly saying the wake word. This is a substantive interaction feature within the claimed period, although it does not establish GPT-3-like open conversation. (Amazon Science, )
Assessment: ContradictedSaid
“There was a lot of bluster (we will 100x the world’s productivity) but no road map.”
What happened: Tesla supplied tentative delivery timing. The same transcript outlines prototype improvement followed by factory deployment and progressively broader uses. (Tesla, transcribed by Stock Analysis, )
Assessment: ContradictedSaid
“we were right to be skeptical; so far, four and half years later, Duplex hasn’t gotten a whole lot further.)”
What happened: Google reported over a million completed bookings and over three million business-information updates across eight countries, alongside the quoted automation rate. These reported deployments show substantial operational progress beyond a limited demonstration. (Google, )
Assessment: ContradictedSaid
“(LLM’s like GPT produce well-formed prose but do not produce interpretable representations of utterances that reflect structured relationships between the parts of those sentences.)”
What happened: The study uses late-2021 GPT-3 and Codex models to map utterances into structured meaning representations. It demonstrates a task capability, while its use of prompts and decoding machinery limits conclusions about ordinary internal representations. (arXiv; Richard Shin and Benjamin Van Durme, )
Assessment: ContradictedSaid
“Current systems simply do not postulate such cognitive models—and so, they technically can’t prove themselves wrong, because they do not even have a representation of what being “right” would look like.”
What happened: The experiments evaluate models critiquing their own outputs against supplied passages or task constraints, and using critiques to improve answers. This contradicts the functional claim of complete inability, while leaving the nature of internal cognitive representations unsettled. (arXiv; William Saunders and colleagues, OpenAI, )
Assessment: ContradictedSaid
“The downstream consequence of all of this is that these models have no grasp whatsoever on truth.”
What happened: GPT-judge classified the factual truth of answers and transferred across held-out answer-producing model families. This contradicts absolute behavioral incapacity, but all benchmark questions were represented in training, so it does not establish unrestricted factual understanding. (Association for Computational Linguistics; Stephanie Lin, Jacob Hilton and Owain Evans; ar5iv mirror, )
Assessment: ContradictedSaid
“Software like LaMDA simply doesn’t; it doesn’t even try to connect to the world at large, it just tries to be the best version of autocomplete it can be, by predicting what words best fit a given context.”
What happened: The paper describes retrieval, calculator and translation tools, and training that routes queries to those tools during inference. This directly contradicts the assertion that LaMDA does not even attempt external grounding. (Google Research authors / arXiv, )
Assessment: ContradictedSaid
“But 8 years later I doubt most people (even in AI) have ever even heard of his program, outside of my mentioning it here. It made zero lasting contribution to AI.”
What happened: The author's statement says Emelya was based on a reworked natural-language analysis system created for Eugene Goostman. Her account dates development to 2016, release to May 2017 and further user research to 2018, documenting a technical legacy beyond the 2014 competition. (Habr / Ontico, account by project participant Anna Degteva, )
Assessment: ContradictedSaid
“They do not gravitate towards human language, and they never develop a cohesive undestanding of the world, no matter how much experience they have.”
What happened: Kallini and colleagues compared GPT-2 learning across English and systematically altered languages, including different training stages. The observed preference conflicts with the language-learning component under the stated family-level interpretation. (arXiv, )