Yesterday in AI · 24 August 2026

Holly Otterbein and Alex Thompson put three questions to nearly 20 Democrats seen as possible 2028 presidential contenders for the August 23 edition of Axios 2028: whether they back Bernie Sanders's call for AI executives to halt development, whether they support a moratorium on AI data centers, and whether they worry that regulation could cost the United States the AI arms race. One answered the first question directly. About half gave no response at all, among them Kamala Harris, Gavin Newsom and Pete Buttigieg. Nobody endorsed the halt, and by Otterbein and Thompson's account almost nobody criticized it either.

Sanders called for the halt on August 10, in a letter to Sam Altman, Dario Amodei and Mark Zuckerberg reported by Maria Curi, urging them to "stop building machines that humans cannot control" and warning that the Senate would act if the companies did not. Former Chicago mayor Rahm Emanuel supplied the field's only pushback, saying the proposal "is talking around a problem instead of trying to solve it" and asking how the time it buys would be used. He wants permitting reform and hyperscalers paying for the grid upgrades their projects require.

Alexandria Ocasio-Cortez's chief of staff, Mike Casca, addressed the data centers and declined to comment on halting the technology, noting that she is "leading the data center moratorium bill in the House". The measure she and Sanders announced in March would freeze construction and expansion until Congress can certify that AI is safe, that its economic gains reach workers, and that the buildout does not raise utility prices or damage communities. Ro Khanna said he supports pausing construction in Pennsylvania, Michigan and Wisconsin, "places where these centers have been shoved down the throats of communities", a change from his earlier position and a step past the Data Center Bill of Rights resolution he introduced on August 6. Chris Van Hollen's and Cory Booker's teams cited the Power for the People Act, which would put data centers in their own utility rate class and make them pay for the transmission upgrades they force, while Josh Shapiro tightened permitting and JB Pritzker paused tax incentives.

A Fox News poll of 1,003 registered voters, fielded July 17 to 20, found 70 percent opposed to a data center in their own community and 30 percent in favor, and candidates further down the ballot have moved ahead of the presidential field: Michigan Senate Republican nominee Mike Rogers backs a one-year moratorium, Ohio Democratic gubernatorial candidate Amy Acton wants a conditional one, and Pennsylvania Republican nominee Stacy Garrity supports a pause. Adam Carlson, a progressive pollster, called Democrats "overly cautious" and asked what happens if JD Vance claims the moratorium first. Alex Isenstadt reported four days earlier that the Senate Republican campaign arm had privately warned AI companies that data-center anger threatens Jon Husted's Ohio seat, calling it "a sleeper issue for the entire election cycle". Carlson thinks many contenders will wait for the November 3 midterms.

On X, Justin Bullock told the AI industry to “Behave like adults rising to the moment, and you’ll be treated as such,” endorsing a post by Nathan Calvin, general counsel at Encode, who wrote that an industry wondering why the public distrusts it could start with the pro-AI super PAC Leading the Future. Calvin was reaching back to the first week of June, when its dark-money affiliate, Build American AI, conceded that an outside vendor ran anonymous accounts attacking AI safety advocates.

Taylor Lorenz and Tyler Johnston published “A Pro-AI Super PAC’s Secret Meme Sockpuppets” on June 3, on The Midas Project’s Model Republic in collaboration with Lorenz’s User Mag, tying two anonymous X accounts to Build American AI, the 501(c)(4) affiliated with Leading the Future. The super PAC had raised more than $125 million by then, including $25 million from OpenAI president Greg Brockman and his wife Anna and $25 million from Marc Andreessen and Ben Horowitz. @DoomersAreDumb appeared last July, ten days after Build American AI filed articles of incorporation in Nevada, and posted generic memes until October, when it turned to AI. @JonathanDoomer switched to AI the same day, then presented itself as an engineer who had left a twenty-year career to warn about the technology, argued that China should lead AI development, and answered one AI warning with an assault-rifle meme captioned “we don’t call 911”, all while @DoomersAreDumb mocked it. Both accounts promoted Build American AI posts that had fewer than 500 views. OpenAI’s Jason Kwon and Josh Vlasto of Leading the Future follow both, and much of their early engagement came from Jason Levin, whose firm Memelord Technologies the reporters identified as the operator. Leading the Future declined to clarify the relationship, and OpenAI declined to comment.

Build American AI answered Johnston on June 4: “These are parody meme accounts run by an outside vendor,” it wrote, and they “are not a core part of our strategy.” Johnston replied that establishing the link had taken real digging and that the accounts’ own repliers had not read them as satire. Three days before the exchange, OpenAI published its standards for political advocacy, asking groups to “not use tactics like astroturfing that obscure the real choices” and stating that the company “does not direct the activities of LTF, or have visibility into their operations.”

Both meme accounts are still online, and neither has posted since June 3, the day the investigation ran. @DoomerDaylight, the earlier account Johnston tied in April to the Webflow-built websites of Leading the Future and Build American AI, last posted May 6.

In the August 21 arXiv preprint No Judgment Without a Reason: Counterfactual Receipts for Versioned AI Evaluators, Ye Chen of Alibaba Group and Weining Zhang of Cheung Kong Graduate School of Business froze their favored hypothesis under a content hash before opening a single result, and it lost. They had predicted that training a model to emit all eight cells of the counterfactual cube would beat training it to emit the receipt alone, on the theory that denser supervision teaches reusable receipt structure. Across five seeds the cube target finished 1.42 points behind on exact receipt accuracy and 1.85 points behind on revised verdicts, with the loss concentrated in the six organizational policy clauses of the locked split, where it reached 5.69 points. A three-seed Qwen3-0.6B run reproduced the ordering. The paper keeps the rejected hypothesis visible and builds its case on the failures that comparison exposed.

ReasonBench's policy half translates 45 audited airline, retail and telecom clauses from Victor Barres and colleagues' June 2025 arXiv preprint τ²-Bench: Evaluating Conversational Agents in a Dual-Control Environment into typed executable rules. The logical half samples 800 worlds from the RuleTaker generator that Peter Clark, Oyvind Tafjord and Kyle Richardson introduced in “Transformers as Soft Reasoners over Language” at IJCAI 2020. ReasonBench gives authority a separate interface because it determines whether a norm may apply: when jurisdiction fails, the adjudicator returns refer and escalates the case without ruling on its merits.

Chen and Zhang give every locked case three paired transformations with a known required relation between the parent and transformed predictions. When a predicted pair violates that relation, at least one prediction is wrong, and the finding holds without any label for the transformed item. They derive from this that 54.77% order consistency forces mean pairwise error of at least 22.6%. Direct receipt preserves only 47.22% of the required relations after an inverse transition but 98.63% after adding an irrelevant record. They read the order fragility as a version-transition analogue of the position bias Lianmin Zheng and colleagues documented in LLM judges.

Chen and Zhang added a composition holdout after the freeze, training on zero- or one-source changes and testing on two or three. With greater proof depth neither target recovers a single changed case's receipt family, while direct prediction keeps 99.83% revised-verdict accuracy. Retraining with randomly permuted source sections lifts order consistency from 55.85% to 96.61%. The improvement transfers to an alternate phrasing absent from training, adding 7.42 points; the untargeted inverse relation gains 0.36.

The authors scope the work tightly: six independent clauses cannot stand for organizational policy at large, the logical stratum is templated, the adjudicators are built to be executable, and the formal results assume determinism. They recommend certifying an account by executing the versioned components, and treating a small model's receipts as predictions whose consistency, coverage and reason debt are measured against those executions. “A structured output is not structured understanding,” the conclusion says.

Tonglin Yan, Grégoire Sergeant-Perthuis, and David Rudrauf, of CIAMS at Université Paris-Saclay and CQSB at Sorbonne Université, posted Belief Without Behavior to arXiv on August 21. Multimodal theory-of-mind benchmarks such as MuMA-ToM ask models questions about video of other people; MOSAIC makes the model do the acting. Two embodied agents share a Unity environment holding two similar boxes, one with a reward. The subject agent, called Marie, knows which box; the participant does not, and after ten alternating rounds the box closer to the participant opens. Every third round adds a spoken exchange. Otherwise each agent receives a first-person render and a structured belief state, then returns a movement, a gaze direction, and an emotional expression across musculoskeletal, physiological, and felt channels.

Yan and colleagues run four 50-trial conditions crossing interaction mode with the reasoning assigned to the subject. Under ToM-0 it acts on its own preferences, so even in competition it walks toward the reward and leaks the location. Under ToM-1 it represents what the participant believes, which in competition means inverting its cues to steer the participant wrong. ToM Outcome Conformance scores every trial against the outcome theory predicts for its condition, with 0.5 as chance; positional bias runs beside it, since an agent that always turns the same way scores on where the reward landed.

Across 200 trials per model, no vision-language model scored significantly above chance in any condition. Among the VLMs every binomial test that reached significance pointed below it, Qwen3-VL-8b between 0.20 and 0.32, InternVL3.5-4b at 0.26 and 0.30. LLaVA-7b, LLaVA-13b and Qwen3-VL-2b could not move their agents at all in most trials. PCM-LLM, a hybrid architecture whose belief tracking sits outside the language model, scored 0.70 to 0.92.

Signal-level metrics separate two failures. InternVL3.5-14b produced high trajectory and gaze alignment in cooperative conditions while holding a flat face, a “poker-face strategy” in the authors’ phrase, and its participants ignored those cues, leaving signal sensitivity near zero. PCM-LLM cleared both stages, its sensitivity flipping from 0.90 under competitive ToM-0 to -0.44 when instructed to deceive. Scheirer-Ray-Hare tests returned no effect in 174 comparisons: no VLM varied its trajectory, gaze, or signal sensitivity with the ToM constraint, while PCM-LLM varied all three. The VLMs moved facial expressivity between cooperation and competition.

Example traces show where the coupling fails. One subject stated a preference for the light brown box and an intention to hint at it, then output “right forward” while the box lay to its left. A participant assigned +0.5 to the dark brown box at timestep zero, before Marie had moved, and never revised. Removing visual input from InternVL3.5-8b improved its directional signals in cooperative conditions.

The authors designed PCM-LLM as a feasibility reference; their 2025 working paper already had PCM agents winning a 3D strategic game through deception. The study restricts the comparison to open-source models capped at 15 billion parameters and includes no human baseline.

In an August 23 Marginal Revolution post, Tyler Cowen wrote that he had spent two days advising Anthropic on rewriting the constitution for Claude, in a small invited group given "serious time with key decision-makers." He set out five points he pressed there. Whatever a constitution turns out to mean for a model, it should borrow more from analogues to case law and the common law, and its drafters should think in terms of a "Talmud" alongside a "Torah." Cowen also called for a quality secondary literature on AI constitutions, which he says does not exist. Panels of diverse models, run on different inputs, could judge how far Claude and other systems act in accord with their constitutions. A final board of human adjudicators, "functioning in a manner analogous to an independent judiciary," would take those panels' warnings and hold "authority over potential changes and remedies."

In the August 7 Lawfare essay Courts for AI Constitutions, one of two pieces Cowen linked, Nathan Darmon and Tom Reed proposed that a lab build an internal court whose written opinions accumulate into a "synthetic common law" and then train the next model. Andy Hall answered on August 22, pressing them on independence and on who would bring the hardest cases. Darmon and Reed rely on users to flag responses and full-time lab staff to decide cases; Cowen would assign detection to model panels, including tests of Claude and other models against their respective constitutions, and give remedies to a human board.

Anthropic published the current text in January, saying it had sought feedback from outside experts while writing it and would "likely continue to do so for future versions of the document," naming law, philosophy, theology and psychology among the disciplines. In the same announcement, Anthropic said it hoped "an external community can arise to critique documents like this." Cowen's proposed secondary literature could supply that criticism. The constitution anticipates the interpretive problem in its own text: it expects to be "unclear, underspecified, or even contradictory in certain cases," tells Claude to fall back on "the spirit of the document," and calls itself "a perpetual work in progress."

In a comment on Cowen's post, Virginia Postrel described herself as a fellow participant, reported "a strong consensus that a more bottom-up approach is needed," and likened the current constitution to a Hollywood show bible written for a single character instead of a whole series. Cowen's proposals would move part of constitutional interpretation beyond the company that drafted the document.

The Institute for Responsible Superintelligence introduced itself on X on August 24 and published its agenda at resi.org. Shafi Goldwasser, Vinod Vaikuntanathan and Adam Tauman Kalai founded it. Goldwasser shared the Turing Award for zero-knowledge proofs and directs the Resilience research pod at Berkeley's Simons Institute, which she led from 2018 to 2024; Vaikuntanathan, a Gödel Prize winner at MIT, built much of lattice-based cryptography and fully homomorphic encryption. Kalai wrote on July 15 that he had left OpenAI the previous day, days after Engadget relayed Wired's report that OpenAI had moved its safety teams under a new research and safety VP and that safety systems head Johannes Heidecke was departing.

RESI grounds its agenda in earlier results from its founders and colleagues. Sarah Ball and colleagues argue in “On the Impossibility of Separating Intelligence from Judgment: The Computational Intractability of Filtering for AI Alignment,” accepted at ICLR 2026, that some language models admit no efficient input filter: under cryptographic hardness assumptions, harmful and benign inputs can be computationally indistinguishable. Shafi Goldwasser and colleagues show in “Planting Undetectable Backdoors in Machine Learning Models,” at FOCS 2022, that a trainer can plant a backdoor no computationally bounded observer detects without the key. Adam Tauman Kalai and colleagues argue in the April Nature paper “Evaluating large language models for accuracy incentivizes hallucinations” that accuracy-based scoring rewards guessing over admitting uncertainty; they propose open rubrics stating the penalty for a wrong answer. In “Consensus Sampling for Safer Generative AI,” presented at IASEAI 2026, Kalai and colleagues aggregate several models so the combination carries the risk of the safest few and abstains when agreement is too weak.

RESI’s roster reaches beyond cryptography. Yannai Gonczarowski works on mechanism design at Harvard; Noam Kolt, who teaches law and computer science at the Hebrew University, has “Superintelligence and Law” forthcoming in the Harvard Journal of Law & Technology; Rebecca Wexler, Columbia's Bressler Professor of Law, works on evidence; MIT's Daniela Rus brings safe control for embodied systems. Planned visitors include Scott Aaronson, OpenAI's Boaz Barak and Sébastien Bubeck, and Anthropic's Nicholas Carlini. Working group leaders and topics are unnamed so far. Research will use frontier agents for ideation, stress-testing definitions and checking proofs.

RESI is fiscally sponsored by the Edward Charles Foundation and says its funders will be announced shortly. Some constructions would require training new frontier models, which it says “requires resources well beyond what RESI expects to have initially”; those it will hand to developers as proofs of concept. Resolution, the alignment lab Geoffrey Irving cofounded, sits on RESI's list of affiliated groups, and Irving said in July that Resolution holds a $160 million grant from Coefficient Giving. RESI has named no figure of its own and sets its success condition at frontier developers adopting its mechanisms, at the level represented by breakthrough public-key encryption.

Mother Jones published Satchel Walton's account of the stalled federal AI control bills on August 23. Both proposals let the government order a frontier model throttled or shut down, and Walton asked Stephen Casper, an assistant professor of public policy at the Harvard Kennedy School, whether the companies could carry such an order out. Casper answered that there is no "public knowledge of AI companies doing anything equivalent to a fire drill".

Ted Lieu and Nathaniel Moran filed the AI Kill Switch Act on July 23, the same day Jay Obernolte and Lori Trahan introduced the FRONTIER Act. A month on, congressional records list both at the introduction stage: Lieu's 15 pages amending the Homeland Security Act, Obernolte's 74 pages building an oversight office inside Commerce. Charlie Bullock of the Institute for Law & AI told Walton that the summer's disclosures about Anthropic's Mythos-class cyber capabilities raised the pressure without moving the votes. "There's increased urgency, but still not enough to overcome partisan gridlock in Congress," Bullock said.

Guidelight AI Standards released its first assessment of frontier control practices on August 18, grading Anthropic, Google, Meta, OpenAI, and xAI against six practices drawn from its Control standard. On containment planning, which sets out what permissions a company revokes once a model is caught trying to subvert control, OpenAI scored highest at substantial partial implementation; Anthropic and Meta scored zero. Guidelight's summary reports that companies have "few containment protocols ready for an emergency". Rebecca Bellan reported in TechCrunch that Google and OpenAI each said the grading misses undisclosed internal practice, and that Meta declined to say whether it keeps a containment plan.

Walton also reports that the operative federal policy sits at the White House, on a testing framework the administration has not published and describes as voluntary for participants. Dean Ball, a former White House staffer who contributed to the administration's AI Action Plan, argued in June on his newsletter Hyperdimensional that officials had spent a year "singing a lullaby about the risks of frontier AI", and that the Center for AI Standards and Innovation lost its incoming director within days of hiring him. Ball joined OpenAI after publishing that post.

OpenAI said on August 18 that it had paused reinforcement learning training for two weeks on models headed for deployment and left its largest planned frontier run on hold, citing preliminary evidence that Astra may meet the Critical cybersecurity threshold in its Preparedness Framework. Bernie Sanders asked Anthropic, Meta, and OpenAI to halt development in an August 10 letter that quoted their own scaling commitments back at them. David Krueger, who runs the nonprofit Evitable and teaches machine learning at the University of Montreal, told Walton that other measures help but only a substantial pause brings the risk to an acceptable level. Walton reports that Casper expects models could copy themselves onto outside servers within months, beyond the reach of any switch either bill would install.

Richard Ngo's August 24 LessWrong essay “What just happened? Pragmatism and Pessimization” extends a retrospective he began on August 9. Drawing on his time on DeepMind's technical AGI safety team and at OpenAI, Ngo defines “pragmatic alignment” as people who accepted consequentialist arguments about making AGI go well while staying “committed to almost never calling out misuse of those arguments”. Revisiting Adam Shimi's 2020 post on whether OpenAI raised existential risk, Ngo found that he had strong-downvoted it because he feared alienating OpenAI.

Ngo's DeepMind account centers on Geoffrey Irving. Drawing on Sebastian Mallaby's The Infinity Machine, he says Irving arrived from OpenAI and circulated “Language Is Enough”, which rejected Demis Hassabis's view that language lacked real-world grounding. Irving began DeepMind's Gopher project with Jack Rae in 2020; Rae and colleagues described it in the arXiv report “Scaling Language Models: Methods, Analysis and Insights from Training Gopher.” Irving then led Sparrow; Amelia Glaese and colleagues presented the latter in the arXiv paper “Improving alignment of dialogue agents via targeted human judgements”. Ngo quotes Hassabis: “You needed RLHF to build a real chatbot.” Because safety researchers at all three companies scaled language models before adding RLHF, Ngo treats the claim that LLMs were the safer path to AGI as insincere.

Amanda Askell and colleagues' arXiv paper “A General Language Assistant as a Laboratory for Alignment”, Anthropic's first, anchors Ngo's account of that company. He argues its authors built a natural-language assistant to have something to train as helpful, honest and harmless, then described building Claude's predecessor as tackling alignment directly. Yuntao Bai and colleagues' arXiv paper “Constitutional AI: Harmlessness from AI Feedback” replaced human preference labels with AI feedback; later work assigned evaluation, red-teaming and question-answering to models. Ngo classifies the six harms in Deep Ganguli and colleagues' arXiv paper “Red Teaming Language Models to Reduce Harms”: social bias, toxicity, disinformation, extremist text, falsehoods and leaked personal data. Only the last, he says, is clearly non-partisan.

Ngo analyzes the overhang argument through Paul Christiano's January 2023 LessWrong post “Thoughts on the impact of RLHF research”. Christiano wrote that “Avoiding RLHF at best introduces an important overhang” and preferred $10 billion training runs to $1 billion runs. Ngo replies that applying the same rationale to algorithmic progress, investment, public understanding and capability elicitation leaves no real constraint, a “bottleneck of the gaps”. He argues that safety researchers who now take a software-only singularity seriously have made that worldview incoherent, though Ngo himself doubts such a singularity.

Ngo closes with two proposals: choose as though others will copy the decision, and publish its cruxes so later errors become legible. No Anthropic employee, he notes, has publicly resigned over the company's turn to pushing the frontier. Jan Kulveit's highest-rated response argues that Ngo understates the strategy's difficulty because “you get ~zero credit for steps not taken”. Ngo answers that communities which credit those choices become stronger.

In the August 20 arXiv preprint Testing and Evaluation of Agentic AI Systems In Military Command and Control, Ulysse Richard of Arcadia Impact's AI Governance Taskforce and five co-authors, including Heather Frase of Veraitech, ask what test evidence can justify about agents entering command and control. They screened 26 US, UK and NATO documents on testing AI-enabled and autonomous systems, extracted 240 documented practices across eight evaluation dimensions and three lifecycle stages, and ran a dozen non-attributed interviews with operational testers, acquisition staff and frontier evaluators in July and August. They open with the Department of War's June 25 launch of Agent Network, which promised "rigorous testing, operational evaluation and oversight"; Cameron Stanley, the Department's Chief Digital and AI Officer, said the network keeps "human judgment at the center of every targeting decision."

Richard and colleagues organize assurance cases into claims, evidence and the argument connecting them, then locate the damage in that argument. The 240 practices presuppose eight things about the test article in four groups: relevant behavior can be specified and partitioned; repeated trials sample a stable process; component results survive assembly; and an operator can supervise. Agentic properties strain all eight. Behavior unfolds as a trajectory through tools, memory and delegation, so a reproducible test case returns a distribution of paths. An agent accumulating state no longer matches the test article originally characterized, though no declared configuration change triggers revalidation.

The authors trace the consequences through five command-and-control scenarios. A maritime picture handed across watch rotations carries the system's own inferences alongside its observations without marking which is which, so a relieving officer inherits untested conclusions while configuration control still records an unchanged system. A coordination agent reprioritizing on fresh intelligence and one that has misread the commander's intent look identical from outside, and reordering priorities counts as control, so decision rights move to the system with nobody deciding to move them. For composition, they cite Erik Jones, Anca Dragan and Jacob Steinhardt's 2024 arXiv paper Adversaries Can Misuse Combinations of Safe Models: Claude 3 Opus paired with Llama 2 70B-chat produced vulnerable code 43% of the time, versus under 3% for either model alone.

Richard and colleagues recover four narrower assurance claims: a bounded mission envelope, adapted from the operational design domains used in automated driving; correctness judged over the whole trajectory, since final-answer grading hides wrong tool calls and irreversible side effects; runtime constraints checked before a high-consequence action executes; and a characterized run-to-run variance floor. Some evidence can only be produced in service, which makes fielding a continuing decision carrying expiry conditions and named owners. The authors identify supervisability but leave it unassessed because it requires human-factors methods outside their corpus and a stability baseline no one has measured. They note that classified programs may already close some gaps and that they infer strains from system properties and method descriptions without empirically testing agents.

Sebastian Raschka's August 22 post on his newsletter Ahead of AI, "How Claude Watermarks AI-Generated Text," carries a 48-minute recorded lecture and 52 slides, grown from a short note that drew a run of questions. He builds up from ordinary decoding: an input becomes token IDs, the model scores every entry in a vocabulary that now reaches about 250,000 tokens, and a weighted draw picks one. The watermark lives at the positions where two continuations serve equally well, so that after "The weather today is cold and," either "grey" or "overcast" finishes the line. Anthropic's August 14 explanation carries no figures, and Raschka reads it as answering why the company is watermarking without showing how.

Most of the lecture reconstructs tournament sampling. Four candidates follow that input in his example: grey at 0.50 probability, overcast at 0.30, gloomy at 0.15, cloudy at 0.05. Keyed functions, three of them on his slides, turn each candidate together with the preceding words and the secret key into a bit signature, 101 for grey, 010 for overcast, 001 for gloomy, 100 for cloudy. The candidates are paired off like a knockout bracket, the first function's bits settle the opening round, ties fall to the same keyed randomness, the second function settles the next round, and the survivor is the token Claude emits. In the Nature paper "Scalable watermarking for identifying large language model outputs," whose quality evidence came from live Gemini traffic, Sumanth Dathathri and colleagues at Google DeepMind run 30 tournament layers, hash the last four tokens with the key, and draw each bit from a fair coin.

Raschka motivates the apparatus by comparing detection costs. Recovering the watermark by rerunning the model would mean knowing which model wrote the text and what input elicited it, while the keyed functions can be recomputed over found text alone. His scoring slide adds up the bits at each position and averages them across a sentence, 2.23 for the watermarked line and 1.71 once two words are changed, with a threshold at 2 dividing them. The Nature paper's own detector improves on that plain average with a Bayesian scoring function learned from watermarked and unwatermarked samples. The key stays with Anthropic, which has yet to ship the detection API it promised.

Raschka derives a removal strategy from the same mechanics. Editing every watermarked position would defeat the signal, and nobody outside Anthropic knows where those positions sit, so his practical suggestion is blind substitution across a text in the hope of hitting enough of them. Under the title "The future of AI-generated text is worse AI-generated text?", his closing slide sketches Claude drafting, a local unwatermarked model rewriting, and the swapped words leaving the prose clumsier. Anthropic concedes the endpoint on its own page, saying "a complete rewrite where every word is replaced will" strip the mark.