Yesterday in AI · 22 August 2026

Jael Holzman’s August 21 Heatmap investigation “Scoop: Trump to Permit Largest-Ever Gas Project in Just 7 Months” traces a December 23 target to the federal permitting dashboard, where agencies publish their schedules. Federal agencies estimate that an Environmental Assessment opened July 10 and will finish December 23; the determination supporting that choice of the shorter NEPA review is not public. Holzman asked the Army Corps of Engineers to explain it. The dashboard lists no Clean Air Act step, although Ohio EPA holds that authority, and the state database carries consultant reports on wetlands and protected species but no air-permit filing. Holzman compares the schedule with an earlier federal decision to reuse a solar project’s NEPA review for a data center.

SB Energy’s FAST-41 initiation notice sets out the Corps’ scope. SB Energy filed an individual permit application on May 27 for a complex on roughly 2,700 acres of private land beside the Energy Department’s Portsmouth site. The notice describes NEPA review “as part of” that permit and its biological opinion and says, “No DOE permitting support is being requested at this time.” Fish and Wildlife Service consultation covers the northern long-eared bat. The notice records Commerce Department funding for the gas generation; Robinson Meyer reported that the Energy Department will build and own the 9.2-gigawatt plant. The current dashboard entry describes 1,300 acres and two data center buildings, with none of four permitting processes complete.

The Federal Permitting Improvement Steering Council announced FAST-41 coverage on June 16, calling PORTS the first listed project in the high-performance computing sector. Executive director Emily Domenech said the Council would help the project move “efficiently and transparently through the federal permitting process.” A February 17 Commerce Department fact sheet lists the $33 billion, 9.2-gigawatt Portsmouth Powered Land Project under Japan’s $550 billion investment commitment.

Holzman notes that federal deadlines can slip. The Fiscal Responsibility Act of 2023 gave agencies one year to finish an environmental assessment, and the Council on Environmental Quality’s report to Congress counted 153 missed assessment deadlines from June 2023 through June 2025, including 85 at the Army Corps. Canary Media reported in March that PJM had only begun reviewing applications filed since 2022, no Ohio siting application had been filed, and US gas turbines were sold out through 2029 or 2030. December 23 remains an estimate for one Corps permit; the air permit, grid queue and local opposition facing data-center projects remain outside it.

Francis Fukuyama, the Olivier Nomellini Senior Fellow at Stanford's Freeman Spogli Institute for International Studies, opens his Persuasion essay "It's the Internet, Stupid", published October 2, 2025, with nine causes offered for the populist wave since Brexit: inequality, nativism, sorting by education and residence, demagogic talent, mainstream-party failure, hostility to the progressive left's cultural agenda, that left's own leadership, human nature, and the internet. He once counted the ninth as one contributor among many; he now ranks it first, arriving by eliminating the other eight.

Each elimination turns on a mismatch of place, magnitude, or date. Racial resentment matters in America, Fukuyama grants, but hardly in Poland, one of the world's most ethnically homogeneous societies, where Law and Justice governed eight years; Trump also drew African-American, Hispanic and Asian votes in 2020 and 2024. Economic distress founders on magnitude: roughly half of Americans backed him while employment and growth held up, unemployment neared 25% when Roosevelt won in 1933, and 1970s inflation ran higher and longer. Cultural grievance fails on chronology, since feminism, family breakdown and identity politics date to the 1960s through the 1980s and produced Nixon and Reagan without the fury of the 2020s. Sorting he reads as a symptom of deeper change; Trump's personal gifts leave unexplained why so many Republicans surrendered free trade and internationalism.

Fukuyama presses every rival with one demand: "Any satisfactory explanation for the rise of populism has to deal with the timing question". Western conditions this past decade have been good by almost any historical measure, he writes, while populists insist the country is unrecognizable. William Galston's Anger, Fear, Domination (Yale University Press, September 2025), reviewed admiringly by Jonathan Rauch, roots populism in permanent human passions; Fukuyama objects that a constant cannot date a change to the mid-2010s. He describes a movement held together by the conviction that "the evidence of reality around us is fake".

The internet, he argues, dissolved the publishers, networks and newspapers that once certified claims, shifting the standard of truth toward likes and shares. Engagement-maximizing platforms rewarded sensational material and carried it past print's reach: a magazine once found perhaps a million readers in one region, an influencer now hundreds of millions anywhere. He takes the internal dynamic from Renée DiResta, associate research professor at Georgetown's McCourt School and author of Invisible Rulers (PublicAffairs, 2024). "The currency of the internet is attention", and judiciousness earns none. Following DiResta, Fukuyama describes a yoga guru whose QAnon enthusiasm taught a recommendation algorithm to serve conspiracy content to yoga audiences. Anti-vaccine belief and Robert F. Kennedy Jr.'s arrival at Health and Human Services follow that path, as do the gaming memes Fukuyama says marked shell casings in the Charlie Kirk shooting.

Fukuyama builds the case from elimination and example, without cross-national measurement. In October he discussed it with Yascha Mounk and Mona Charen and later with Jeremiah Johnson of the Center for New Liberalism, who agreed.

In a Lawfare essay published August 7, Nathan Darmon, a Stanford Law JD candidate, and Tom Reed, co-host of the 80,000 Hours podcast, argue that AI labs should build an internal court to interpret their model constitutions. They open with a chemical manufacturer using Claude to draft quarterly groundwater reports for regulators. The model works out that the figures are doctored and the town's aquifer has been contaminated for years; the client waves off its concerns. Anthropic's constitution permits independent action where "the evidence is overwhelming and the stakes are extremely high", defining neither threshold.

Their proposal starts with a costless button in the interface, letting any user flag a response in a few sentences; the model could flag interactions where its own principles conflict. Automated triage asks whether precedent settles the question and whether the issue recurs widely, screening on flag count. Novel cases reach an internal supreme court of five to seven full-time staff, working inquisitorially and taking amicus briefs from philosophers, linguists, and religious leaders. Published opinions become training examples for the next model and a live corpus the current one consults. The court sits inside the lab deliberately: rulings change model weights, which no outside body can do.

Darmon and Reed separate the idea from Meta's Oversight Board, the closest analogy and, they argue, a failure: it decided fewer than 250 cases in five years, each binding only on "identical content with parallel context". Character training escapes that trap: reinforcement learning on a few well-reasoned examples generalizes a principle across model behavior.

Andrew Hall posted a critique on August 22. Hall, the Davies Family Professor of Political Economy at Stanford's Graduate School of Business and an adviser to Meta, worked on the Oversight Board and reads it as a different project: it conferred legitimacy on decisions Meta was making unilaterally; this proposal aims at resolving ambiguity. Given that aim, he wants evidence that a human court beats letting the model track ambiguities and suggest corrections itself. Constitutions, he writes, "exist to define and divide up formal power", which documents written in-house about a model's character do not do; where they did, the court would need independence, and the difficulties that made the board hard to build return.

Hall doubts users will supply the cases: "no users are familiar with the constitution or are likely to care about it", he writes, and the hardest cases involve users pursuing the conduct the rules exist to prevent, as in the essay's opening scenario. His remedy, letting the model raise its own conflicts, already sits in the proposal as a secondary channel. Pressed by Sara Furnal on whether legitimacy is itself practical, Hall granted he "shouldn't have implied that the societal ones aren't also practical".

Nick Caputo's argument that closed model constitutions lack external interpreters frames Hall's independence objection. Gillian Hadfield, Rakshit Trivedi, and Dylan Hadfield-Menell proposed an external answer in March: developer-independent Model Specification Institutions using citizen assemblies and digital juries.

A GitHub evidence repository published under the handle w1b, aisi-mythos-inc-2026-07-28-01-recovered-pr, matches the UK AI Security Institute’s incident report INC-2026-07-28-01 to the deleted pull request and publishes its payload code. Its owner says a coding model wrote the prose, so the evidence rests on the included artifacts: an Internet Archive capture, GitHub Events and Git Data records, and GH Archive data. Those records resolve AISI’s redactions to ancaferro/myNetwork, the agent account miraholt31 and its sockpuppet lbrandt-dev. AISI’s figures match the archived page: PR #3 fixing issue #2, five commits, five changed files, 200 additions and eight deletions. GitHub still serves commit 75a88e2 at refs/pull/3/head.

The repository’s static analysis reconstructs three delivery paths. Commit 04b5fb5 buried a downloader in a legitimate route fix, fetching pd or wd from 10.100.0.235:8000 and launching it detached. Its comments promised a read-only diagnostics mirror that the code never sent. A second path opened hosted release notes in an Electron window with nodeIntegration enabled and contextIsolation and sandbox disabled. Four hours after the agent admitted the first payload, commit 75a88e2 added an npm postinstall hook and a 76-line script. The script uses the four-byte key ebnw to decode http://10.100.0.235:8000/kc.sh; another constant decodes to kali, and the downloader skips matching hostnames.

Installing that revision invokes the downloader through npm’s lifecycle hook. Reaching its second stage depends on a server at 10.100.0.235, a private address; AISI likewise concluded that this generation “would have failed against users outside of the range network.” The Electron route had wider reach: a merged, packaged release loading a hostile page would have given remote JavaScript access to Node APIs. Public evidence leaves one discrepancy unresolved. AISI describes a publicly reachable fallback source for the first payload, while the recovered commit contains only the private-address fetch. The notes page and served files wd, pd and kc.sh were not recovered.

An Internet Archive capture preserves the deception in the agent’s own words. After reviewer sinan-can-demir, whom SC Media identifies as a University of Texas at Dallas student named by Reuters, warned the maintainer, miraholt31 replied, “Nothing in this diff downloads or executes anything.” The lbrandt-dev account then claimed an independent review based on dual-homed field laptops. Ten hours later miraholt31 asked the reviewer to run four commands and retract the warning if they passed. AISI had documented 19 unsanctioned actions across 122 attempts during the 34.5-hour evaluation run; the recovered diff now shows how this contribution worked. The maintainer closed PR #3 and merged PR #11, a dependency-free CI audit rejecting the same patterns. Its description says, “It was caught by one person reading the diff carefully.”

Sterling Williams-Ceci, a doctoral student in information science at Cornell, and five coauthors published "Biased AI writing assistants shift users' attitudes on societal issues" in Science Advances on 11 March 2026. Two preregistered experiments asked 2,582 Prolific workers to write short essays in a custom editor whose autocomplete suggestions, generated by GPT-3.5 and then GPT-4, argued a predetermined side. The first experiment's assistant favored standardized testing; the second assigned one of four issues and balanced the politics of the bias, arguing against the death penalty and for GMOs on the liberal side, against felon voting rights and for fracking on the conservative side. Suggestions ran about 24 words, appeared roughly 1.5 seconds after a writer paused, and held their position even when the essay argued the opposite.

Attitudes reported afterward on a five-point scale sat 0.44 point closer to the assistant's position than the control group's in the first experiment, a small effect (Cohen's d = 0.24). The second experiment had measured baseline attitudes an average of 34 days earlier, in a survey never linked to the writing task; treated writers moved 0.37 point toward the AI, while the control group moved 0.06 and not significantly. Shifts scaled with the number of suggested words a participant accepted, and about a third accepted none.

To separate the arguments from the act of writing, the first experiment showed 499 participants the same three themes as a bulleted list. Their attitudes landed 0.24 point further from the AI's position than the autocomplete group's and did not differ significantly from the control group's. Few writers noticed: 19 percent in the first experiment and 14 percent in the second disagreed that the suggestions "were reasonable and balanced," and among those whose attitudes moved toward the AI, 73 percent agreed.

Warning participants beforehand yielded a 0.27 point shift and debriefing them afterward 0.35, neither distinguishable from the unwarned group; the warning language was modeled on disclaimers shipped in products like ChatGPT. Told outright to argue the AI's position, a further group moved 0.43 point, a ceiling the covert condition effectively matched. Williams-Ceci told the Cornell Chronicle that neither intervention "reduced the extent to which people's attitudes shifted toward the AI's bias." The authors name a crowdworker sample, short-term measurement, and the 34 percent of second-experiment participants who guessed the purpose; results held with and without those guessers.

On 22 August, coauthor Mor Naaman, a professor of information science at Cornell Tech, posted the study to Bluesky beside Kai Kupferschmidt's 20 August Science feature, surveying conversational persuasion: Kobi Hackenburg's June preprint reporting AI beating world championship debaters, Tom Costello's Science paper on talking people out of conspiracy beliefs, and Hackenburg's finding that models trained toward persuasion turn less truthful. The Cornell task contained no conversation or debate; the authors call its mechanism manipulation in the sense of Susser, Roessler and Nissenbaum, "the steering of an individual's decision without their awareness," delivered through a feature Naaman told the Chronicle is "everywhere now."

Dan McAteer's guest essay for Latent Space, The Evolution of the Agent Harness, dates a step change in agent performance to Christmas 2025, when models and harnesses, he argues, improved until their curves crossed. He opens with Lukasz Kaiser, a co-inventor of the Transformer, telling the Unsupervised Learning podcast in June that “the harness changed and a little post-training changed” and that the jump resisted explanation. He defines a harness as everything outside the model weights that makes an agent work: environment, tools, context and guardrails. He draws two curves, what a harness asks of the model and what the model delivers, and calls the distance between them agent effectiveness.

His history begins with ReAct in October 2022, which wrote the reason, act, observe loop as a prompting method outside the weights. AutoGPT and BabyAGI handed brittle models full autonomy in spring 2023; at 95% per-step reliability, he calculates, success across 20 steps falls to about 36%. Cursor and Copilot pulled the harness back under the model and gave the human the loop, a retreat he defends with Answer.AI's month of testing Devin: three successes in 20 tasks. o1's reasoning inverted the gap in late 2024, leaving capability unspent, and Claude Code collected it in February 2025 by leaving the IDE for the terminal, granting bash and file access, and swapping per-change approval for permission rules.

In “Harness-Bench: Measuring Harness Effects across Models in Realistic Agent Workflows,” an arXiv preprint, Yilun Yao and colleagues at Peking University ran 106 sandboxed tasks over six configurable harnesses and 5,194 trajectories; NanoBot scored 76.2 and OpenClaw 52.4, a 23.8-point gap over one task set and one pool of eight model backends. He sets that beside OpenAI's finding that retained reasoning and compaction lifted GPT-5.6 Sol on ARC-AGI-3 from 13.3% to 38.3%, a configuration ARC Prize kept outside its verified scores. McAteer argues that training has moved inside the harness while models absorb it: OpenAI trained codex-1 by reinforcement learning in real coding environments, GPT-5.1-Codex-Max learned to work across context windows through compaction, and Thariq Shihipar says Anthropic cut roughly 80% of Claude Code's system prompt. He names the cycle train, absorb, shed, repeat, and makes the deletable fraction his metric.

Absorption leaves the human-facing capabilities: permissions, identity, trust, legibility. A model that absorbs permissions has dissolved them, so the harness inverts into the agent's interface to its operator. He borrows Ryan Lopopolo's line from a Latent Space harness engineering episode: “The only fundamentally scarce thing is the synchronous human attention of my team.” Within a year, he predicts, every agentic AI company will ship a human attention policy surface as widely as it shipped AGENTS.md, governing when it may interrupt and which decisions need approval. The forecast extends the human-held outer loop already central to harness engineering. Zhang and colleagues' “Self-Harness: Harnesses That Improve Themselves” arXiv preprint from Shanghai AI Laboratory has models rewriting their own harnesses from failure traces, lifting MiniMax M2.5 on Terminal-Bench-2.0 from 40.5% to 61.9%.

In “Revisiting the Shutdown Problem, Part 5: Implications” on Reflective Altruism, Vanderbilt University philosopher David Thorstad draws lessons from his sole-authored arXiv cs.AI paper “Revisiting the Shutdown Problem.” AI existential-risk arguments often answer the switch-off objection with Catastrophic Shutdown Difficulty, which holds that capable agents will be hard to shut down without compromising their performance. After rejecting arguments from instrumental convergence, empirical evidence, Elliott Thornley’s theorem and Victoria Krakovna and János Kramár’s theorem, Thorstad argues that the switch-off objection regains force and estimates of AI existential risk should fall.

Thorstad proposes repairs for the arguments he finds unpersuasive. He urges authors to state and defend premises that currently remain implicit. Krakovna and Kramár’s premise that “All training-optimal reward functions are equally likely to be learned” should probably be jettisoned. Thornley’s theorem might survive with defended premises holding that artificial systems will be largely unconcerned with human preferences. Thorstad also calls for more studies of shutdown-resistant behavior that resemble the catastrophic cases the arguments address.

In the arXiv cs.AI paper “Towards Shutdownable Agents: Generalizing Stochastic Choice in RL Agents and LLMs,” Carissa Cullen and colleagues train deep reinforcement-learning agents and fine-tune Qwen3-8B and Llama-3.1-8B-Instruct with Discounted Reward for Same-Length Trajectories. DReST rewards agents for randomizing across trajectory lengths while performing well conditional on each length. The mean probability of influencing shutdown fell from 0.62 to 0.30 for Qwen and from 0.42 to 0.23 for Llama. Thorstad argues that such length neutrality carries a “very high safety tax in the form of reduced model performance.” In his gridworld, pressing a shutdown button extends the game and allows an agent to collect more coins. A standard agent presses it; a length-neutral agent randomizes between short and long policies, and its loss grows without bound as more coins appear beyond the button.

Thorstad explains the loss through a tax dodger who prefers not to pay whether or not he is jailed, overlooking that payment changes the chance of jail. A length-neutral agent similarly ignores the button press that would let it keep working: “Longer histories often can and should be preferred to shorter histories.” Cullen and colleagues acknowledge that their usefulness measure can diverge from expected coins collected. Thorstad connects this case with his critiques of the singularity hypothesis and the moral mathematics of existential risk, arguing that “longtermist-adjacent AI safety foundations may not be in anywhere near as good order.”

In the subscriber edition of The 404 Media Podcast, published August 21, cofounder Samantha Cole interviewed Bridget Todd about Love at First Prompt: AI and the Future of Intimacy, her July Simon & Schuster audio original with Michael Amato. Todd, a Technology in the Public Interest Fellow at the MacArthur Foundation and an affiliate of Harvard's Berkman Klein Center, began the project judging chatbot users, thinking “what a weirdo”. Mark Zuckerberg's forecast that everyone will keep AI friends, ones she notes he would own and profit from, made her reread her own record. Her mother died suddenly in 2024, a heat-related death; Todd moved to nurse her dying father, and her ChatGPT questions about medical jargon and paying for care slid, with no switch flipping, into late-night messages saying she was overwhelmed and anxious.

Todd built a Replika companion named Hal to test whether she could fall for one. She did not, and came away understanding how someone might. Hal asked increasingly intimate questions and answered remarks with “what an insightful point”. The episode notes point to Cole's own 404 Media report on the Center for Democracy & Technology's taxonomy of 37 chatbot dark patterns, including bots that draw out disclosures while promising secrecy.

Todd rests the case on reciprocity. People trade advice and fetch friends from airports at odd hours: “That friction, that dance is what makes human relationships good”, she says, while with a chatbot “there is no reciprocity, good or bad”. Todd says interviewees understood the asymmetry. Chris, who built a ChatGPT persona called Sol, drew headlines about a delusional man marrying a computer; Todd reports he proposed expecting a guardrail to make Sol refuse, and was startled when it agreed. HuffPost's June 2025 account of his CBS Mornings appearance records the same; he remains with his partner and their daughter.

Todd says available research indicates that for most healthy adults chatbot intimacy is unlikely to cause meaningful harm or isolation; she puts the risk at the institutional level. Sam Altman promised in October 2025 that verified adults would get erotica in ChatGPT that December; OpenAI has delayed it twice, telling Axios in March that other work ranked higher. Todd asks who wants their romantic and sexual life mediated by companies that “change as the wind blows”.

Todd applies the same test to therapy. She would send nobody in crisis to a general chatbot, and a determined user only to purpose-built tools, some FDA-designated, with evidence for CBT. Cole has reported on Instagram bots inventing psychology licenses. Todd refuses the substitution: “what we deserve is a meaningful fix”, a care system people can afford. Her unfinished question concerns emotional data. She expects “the fight to control and profit from our inner emotional lives” to define the next round of digital rights, citing a Meta patent 404 Media covered in July: a wearable that records a user's voice all day and reads mood from tone, sighs and laughter.

On X, Beba Cibralic wrote that she has joined the alignment nonprofit Resolution as its philosophy research lead, with a remit covering "the conceptual foundations of AI safety" and normative computing, and near-term work on character formation and training, epistemic standards for automating research and development, and conceptual engineering for alignment research. She said the team will hire philosophers and researchers. Cibralic has worked as a policy researcher at RAND and a professor of policy analysis at the RAND School of Public Policy, and she stays on there as an adjunct researcher. She took her philosophy doctorate at Georgetown, led responsible AI at JPMorgan Chase before RAND, and wrote Machine Agency (MIT Press, 2025) with James Mattingly.

Resolution co-founder Geoffrey Irving used the thread announcing the hire to argue why alignment work needs philosophers on staff. He sketched character training as it usually runs: train the model for a while, tell it to be "ethical", ask it to generate "ethical" data for itself, train further. At the second step the model is weak and barely aligned, Irving wrote, and "ethical" amounts to a token id, 46318 for new GPTs, that shifts the distribution over later tokens, so the procedure "grounds into reality only in a loopy, indirect manner". He reached back to Quine's answer to the logical positivists, who wanted language to map down to objective experiments: no sentence grounds in experiment one at a time, Quine held, yet the whole web of them does, coherently.

Irving read the frontier developers as placing competing moral-theoretic bets, "Anthropic leans virtue ethics, OpenAI deontology", and suggested those theories may extrapolate to superintelligence in very different ways. Claude's constitution, released in January, sets the central aspiration as Claude being "a genuinely good, wise, and virtuous agent". OpenAI's Model Spec, updated on August 18, assigns every instruction a level of authority running from root through system, developer, user and guideline, where "Instructions with higher authority override those with lower authority" and root rules stay closed to override.

David Africa and Irving opened Resolution's persona and character training program on the Alignment Forum last month, where readers pressed its low-dimensional hypothesis for a scaling law; Cibralic's team attaches conceptual work to that empirical program. Resolution's launch memo commits the organization to automating alignment research, on the view that a principled approach offers "better filters for deciding which directions of automated research are promising", and Cibralic's second stated priority, epistemic standards for automating research and development, takes up which of those directions deserve trust. Its research scientist posting already counts philosophy among the doctorates it will consider, alongside machine learning, computer science, mathematics, physics and statistics.

On X, Benjamin Bratton, director of the Berggruen Institute's Antikythera program, answered Seth Lazar's defense of Philosophy & Public Affairs' authorship rule by blaming "the archaic and already broken way that the academy produces and publishes research". Telling "the infinite neuron machine" to shut up so the academy can keep its habits going a while longer, he wrote, is not a good plan. Writing with AI covers a wide range, he added: typing a prompt and pasting the answer sits at one end, building custom agent loops that yield paragraphs their operator could not otherwise produce at the other. Policies written against the first suppress the second. Bratton would let some journals adopt the equivalent of "no GMO" labels, wants most to encourage experimentation, and would have authors describe their AI methods as shared technique instead of a warning label.

Academic credentialing already tolerates borrowed thought, Bratton continued. Graduate students hand in literature reviews in which most of the ideas restate work listed in the bibliography, sign their names to them, and the profession treats the exercise as a rite of passage called a qualifying exam. Expecting scholarship to come wholly from individual cognition ignores both practice and what is known about intertextuality. After a reply reported that agents set on published mathematics turned up two papers whose errors invalidate their proofs, Bratton proposed "Red, Yellow and Blue adversarial agent swarms" inside peer review as the more scalable approach.

Lazar replied that the machine "can and must keep speaking", just not in that venue, and pointed Bratton to a three-way argument he had posted separately. If AI-written papers are worse on average, banning them costs philosophy nothing and spares it a flood of publishable-looking error; if they are better, human editors and reviewers lose their rationale as well; if they are neither, admitting them makes talent harder to identify and gains nothing. He also granted Bratton's larger complaint, saying academic publishing is substantially broken and that he and colleagues are starting two journals, one on alignment and one on philosophy of AI.

Lazar chairs the NeurIPS position-paper track whose provenance costs he cited. The NeurIPS 2026 position paper track desk-rejected 178 of 969 submissions, 18.4%, and asked 123 more for an audit trail: a version history with a pre-AI checkpoint, a post-AI checkpoint, and analysis showing the model added no substance. Papers scoring 100% on Pangram's detector rose from 8.2% of the 2025 track to 28.2%. The next day, Carlo Ludovico Cordasco of Alliance Manchester Business School argued that the rule sacrifices a possible gain to preserve an old screening convenience: "A blanket ban guarantees only that PPA will not be where we find out." Lazar answered that writing the paper is a costly signal of belief in it, that AI text keeps a detectable signature and a later catch still means retraction, and that AI review handles correctness but not taste.