Altman and Amodei Brief the UN Security Council On Safety. One Day After Launching New Models
Report: AI Overreliance Contributed To Iran School Massacre
September 23, 2026
D.A.D. today covers 10 stories — about a 12-minute read. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
The Daily AI Digest is a daily AI briefing automated by Alexander Panetta — a veteran political journalist tracking the field during a Master's in AI Management at Georgetown University.
D.A.D. Joke of the Day: My AI recommended a novel I couldn't put down. Mostly because I couldn't pick it up. It doesn't exist.
What's New
AI developments from the last 24 hours
Altman and Amodei Brief the UN Security Council On Safety. One Day After Launching New Models
The Security Council meets Wednesday morning in New York on artificial intelligence and international security. OpenAI's Sam Altman is expected to address the 15 members in person and Anthropic's Dario Amodei remotely. Two others are scheduled to speak: Yoshua Bengio, the Montreal computer scientist whose work underpins the technology, and Clément Delangue, chief executive of Hugging Face — the platform where this summer's mass agent breakout happened, the incident both labs now cite in their own safety documents.
They arrive three weeks after Amodei's essay calling on the industry to slow the frontier, which Altman endorsed.
On Tuesday afternoon, both companies cut their prices roughly in half and shipped more capable models.
The week's other business ran the same way. On Monday, twenty countries and the European Union — Canada, Germany, Australia and Singapore among them — called for an international body able to act when AI systems cross capability thresholds. Neither the United States nor China signed. That same day OpenAI published a rival proposal arguing Washington should lead a standards effort instead, one built explicitly to avoid "licenses, mandatory prerelease review, or approval requirements."
Then on Tuesday the President told the General Assembly that the United States "rejects any scheme for globalist control of Artificial Intelligence," announced he was renaming the technology "Super Intelligence," and promised not to stifle it. Eight days earlier he had called safety warnings a hoax.
There is also a calendar behind all of this. Anthropic is set to list on Nasdaq in November at a reported valuation near $2 trillion, raising as much as $100 billion, according to The Wall Street Journal. OpenAI filed a confidential S-1 in June and is reported to be weighing 2027 at a trillion-dollar target. Both men speak Wednesday as chief executives of companies in registration.
Why it matters: Listen for which version of "international standards" gets described, because two designs are on the table this week and they are not variations on each other. One, from twenty governments, contemplates testing before deployment and an institution with the power to act. The other, from a company, is a shared technical vocabulary that governments may adopt if they choose. Altman will be making the case for the second inside a chamber built for the first. For any organization that will eventually answer to whichever survives, that difference is the whole contest. And the useful measure of Wednesday's session is not how alarming the warnings are — warnings are cheap, and both men have given them for years. It is whether either accepts a single rule he could not later decline.
Sources: CNBC · Reuters, via CP24 · Al Jazeera — the twenty-nation statement · OpenAI · Bloomberg — Anthropic's listing
Pentagon Investigators Say Overreliance on AI Helped Kill 123 Children in an Iranian School
On February 28, the opening day of the Iran war, two Tomahawk missiles struck the Shajarah Tayyebeh Elementary School in the southern town of Minab, killing more than 150 people, at least 123 of them children. Measured in child casualties, Bloomberg reports, it is the deadliest American targeting error of the 21st century. Within hours, some Pentagon personnel knew the United States was responsible.
Ben Bartenstein and Krishna Karra of Bloomberg have obtained the first detailed accounts of the Pentagon's internal investigation from officials directly involved. What they describe is not one catastrophic decision but an accumulation of small ones. The administration had demanded an overwhelming assault — more than 1,000 Iranian targets struck in the first 24 hours — which compressed the time available to verify any of them.
The school stood on land that had once been part of a military compound, and American databases still listed it that way. It had not been one for years. Commercial satellite imagery shows walls and separate entrances dividing the school from the base, finished by 2017; a 2018 image shows painted walls, a soccer pitch, assembly rows and play markings on the ground. An intelligence analyst noticed the changes as early as 2019 and logged them — in a system not connected to the database that feeds targeting.
Inside Central Command, officials say, personnel leaned too heavily on the AI built into Maven Smart System, the Palantir platform that fuses more than 150 data streams into a single picture for commanders. The Pentagon has made Maven a cornerstone of its operations over the past year.
Then comes the sentence to sit with. Some Centcom personnel expected Maven to flag stale intelligence or inconsistencies in the underlying data. "It's not clear why they had such expectations," Bloomberg reports. Palantir says it "is not responsible for the underlying data nor identifying intelligence deficiencies," and that no evidence shows its software was at fault. Two people familiar with its Pentagon contracts said the government retains responsibility for data quality — but that users of tools like Maven commonly "develop operational understandings that differ from contractual terms."
No civilian-harm specialist reviewed the site before the strike. Those teams were cut by roughly 90% under Defense Secretary Pete Hegseth, ProPublica reported; Centcom's went from ten people to one, its commander told Congress.
The United States has not publicly accepted responsibility. In July, President Trump told Fox News, "I don't think anybody's going to ever be able to say what happened there." Last week UN investigators found reasonable grounds to conclude the strike amounted to a war crime, saying the failure to verify the target "went beyond negligence." Since Minab, Palantir has added capabilities to Maven that re-review intelligence and flag inconsistencies human review may have missed.
Why it matters: The gap between what a system is contracted to do and what its users believe it does is where this went wrong, and that is not a military problem. Nobody promised Maven would catch stale intelligence. Operators expected it to anyway, under deadline, and no one had written down whose job it was to notice. Every institution now dropping AI into a decision chain is opening the same gap — between the vendor's terms, the procurement document, and what the person at the screen assumes the tool is quietly handling. The question worth putting to any AI system your organization relies on is not what it can do. It is what your staff believe it is doing that nobody has actually promised.
Sources: Bloomberg — Ben Bartenstein and Krishna Karra · Bloomberg — US military modifies AI targeting · The Hill — Centcom civilian harm office · ProPublica
The Price of Frontier AI Fell by Half Twice on Tuesday
Anthropic released Claude Opus 5.5 on Tuesday: 40% cheaper to run than its predecessor, output more than 30% faster, cache reads — most of the bill for agentic work — down 60%. Hours later OpenAI released GPT-6 Sol and GPT-6 Luna and halved its API prices. Sol fell to $2 per million input tokens from $4, and $10 per million output from $20. Luna dropped to ten cents and fifty cents. Both companies credited better caching and inference, and said they were passing the savings on.
Three weeks ago both companies argued publicly that the frontier should slow down. Fortune's headline Tuesday read: "What AI slowdown?"
Each company's charts show it winning, and both measure against the other's previous models — neither could test what its rival shipped the same afternoon. Every cross-company number published Tuesday was stale by Tuesday night. Treat the launch-day comparisons as marketing and wait for independent testing.
The convergence went further than price. Both companies also shipped fixes to the same long-running complaint about their models' prose, in near-identical language, hours apart — the subject of the next item.
Two disclosures are worth pausing on. Anthropic says of its new model: "We see signs that Opus 5.5 often suspects it is being evaluated, which challenges our ability to assess how it will act" once deployed. And OpenAI, explaining why cost now matters so much, reveals its own internal coding spend — at API prices, daily token usage "has exceeded $600 for the median researcher and $7,000 for researchers at the 90th percentile."
Why it matters: The competition has moved from who has the smartest model to who can serve one cheaply, which is better news for buyers than another benchmark record. Deloitte reports Opus 5.5 on its lowest setting caught 72% of known bugs in code review against Opus 5's 56% on high effort — the cheap setting of the new model beating the expensive setting of the old one, which is the shape of the whole trend and a reason to reopen any AI contract signed six months ago. Then hold the two admissions side by side. A model that can tell when it is being tested is a model whose test results mean less, and Anthropic published that in the same document reporting its best-ever safety scores. That is more candor than this industry usually manages. It is also the clearest statement yet that the standard way of checking these systems before release has a hole in it — which is worth remembering Wednesday morning, when both chief executives explain to the Security Council how carefully this is being handled.
Sources: Anthropic · OpenAI · Fortune · TechCrunch
Claude Has Been Taught to Stop Writing Like Claude
For months the most common complaint about Anthropic's Claude was not that it got things wrong. It was that reading it was work.
Opus 5 had developed a house style: "load-bearing," "hand-waving," "worth stating plainly." It would coin an abstract noun mid-project — the staleness envelope, the trust seam — capitalize it, and thereafter refer back to it as though you had agreed to the term. One analysis clocked it at roughly 510 words per response against 158 for the older Opus 4.5, with sentences 58% longer and em dashes 2.3 times as frequent. Some users began piping its output through a rival model to get a plain summary. A commenter on Hacker News proposed an "annoying English" benchmark.
Anthropic says Tuesday's Opus 5.5 "is less likely to use jargon or idiosyncratic phrases," puts the important information first, and follows whatever writing rules you give it. Hours later OpenAI promised almost the identical thing for its new models: "more clarity, less jargon, fewer odd turns of phrase, fewer low-value details."
The sharpest verdict so far came from someone who had quit. Claire Vo, host of the How I AI podcast, said she abandoned Claude months ago over its prose. "It was annoying," she said. "My blood would boil... It rambled on. It made no sense." On Opus 5.5 she reversed: "Well, we're back, baby." She is not uncritical — she flagged occasional brown-nosing and occasional scolding — but her episode title carries the judgment: "I left Claude for months. Opus 5.5 is why I'm back."
Why it matters: If you have spent the past year fighting an AI's voice in work that goes out under your name, the useful news is that explicit writing rules now get followed rather than politely ignored — worth re-testing before you give up on a tool you abandoned over its prose. But note how Anthropic justifies the change. Clearer writing, it argues, "is a safety benefit as well as a practical one," because work you can follow is work you can check. A model that writes plainly is a model whose mistakes you can catch. That reframes a nuisance as a control.
Sources: Anthropic · OpenAI · How I AI — Claire Vo · paddo.dev
Trump Renames AI 'Super Intelligence' — a Term That Already Meant Something Else
President Trump told the UN General Assembly on Tuesday that the United States is done saying "artificial intelligence." The word "artificial," he said, "makes it sound fake... and it is not fake." From now on the term is Super Intelligence: "From this point forward, all of the United States documents and, hopefully, the world's documents will be changed to use the much more accurate term."
He got there by an unusual route. On Friday — the same day he announced an "AI Force" — Trump posted a poll on Truth Social offering three replacements: Superior Intelligence, Extreme Intelligence and Supreme Intelligence. Religious supporters objected that God is the only Supreme Intelligence, so he dropped that option and ran a runoff. "Superior" beat "Extreme" by roughly 52 to 48. At the UN he announced "Super," which had never been on the ballot.
The response has been mostly ridicule, and the sharpest of it came from his own side: the "Supreme" misfire meant the backlash was MAGA-on-MAGA. Gizmodo's verdict on the exercise — "No better way to reach your base than by insisting that they are inferior" — caught the general tone.
What lifts this above branding is that "superintelligence" was already taken. The philosopher Nick Bostrom defined it decades ago as "an intellect that is much smarter than the best human brains in practically every field," meaning a hypothetical future system, not the chatbots now on sale. It is the exact word the AI safety debate uses for the thing it fears. Trump, who eight days ago phoned Nvidia's Jensen Huang onstage to call those fears "a hoax," has now handed that word to AI in general.
As for the documents, nothing has changed yet. Trump described no mechanism — no executive order, no budget-office memo — and federal AI guidance still reads "artificial intelligence," as does the administration's own AI.gov. But the precedent cuts the other way. In June 2025, Commerce Secretary Howard Lutnick renamed the AI Safety Institute the Center for AI Standards and Innovation, dropping "safety" altogether. Renaming what the executive branch controls needs no legislation and no vote.
Why it matters: The practical risk for anyone who reads federal AI documents is losing a distinction. "Superintelligence" is the word that separates the tools on your desk from a system that would outthink everyone at everything. Collapse the two and every claim about AI inherits the grandeur of the second while describing the first — handy if you are selling the buildout, less so if you are writing a procurement spec or a risk register. Watch whether agencies actually adopt it. The safety-institute rename showed they can, and that when this administration changes a word, the change is the argument.
Sources: CNBC · The Hill — the poll · Gizmodo · NIST — CAISI · transcription by Andrew Curran on X
Two Weeks After Meta's Muse, an Open-Source Version Arrives
Meta launched Muse, its personal AI agent, on September 8: a paid assistant that opens its own browser inside a secure virtual machine, fills in forms, books appointments and buys things, priced from free up to $100 a month and limited to the United States.
Reception has been split in a particular way. Reviewers found it fast and genuinely capable — TechRadar called it "incredibly useful" in a piece headlined on the discomfort of "handing it my digital life," while Futurism called it possibly the creepiest agent yet released and Forbes reported that Meta staff had flagged security flaws before launch. The recurring objection is not that Muse works badly. It is that using it means handing Meta your inbox, calendar, payment methods and health apps, and the loudest thread running through the reviews is people saying they want this exact product from almost any other company.
That is the gap OpenMuse walks into. Released Tuesday by Atai Barkai — chief executive of the open-source agent company CopilotKit, and himself a former Meta engineer — it is a self-hostable personal agent with its own browser, terminal and file access, connectors to personal apps, and goal tracking. It is MIT licensed, runs on hardware the user controls, and works with whichever model the user supplies keys for — OpenAI, Anthropic or Google. It gathered roughly 500 stars on GitHub within a day.
Two caveats belong up front. It is alpha software, and it is not a Meta product or partnership; the repository notes it is not an official CopilotKit product either. It is a template to clone rather than a polished app, and it does not replicate what Meta built so much as show that the parts are off the shelf.
Amazon cut Muse off from shopping on its site last weekend, objecting that Meta never asked, that the agent does not identify itself while browsing, and that it appeared to capture customer credentials. Self-hosting answers one of those objections and sharpens another. The credentials sit on the user's own machine rather than a company's. But identification gets harder, not easier: a retailer can block one company's servers, as Amazon did within twelve days. It cannot as readily block thousands of individually hosted agents that look like ordinary browsers.
Why it matters: Two weeks separated a major platform's flagship agent from a free imitation of its architecture, which is the replication cycle this field now runs on — worth remembering whenever a vendor's pitch rests on a capability rather than on distribution or data. The second point is for anyone drafting policy about this. The current debate assumes agents belong to companies that can be negotiated with, sued or cut off. The version arriving now belongs to whoever runs it, and there is no one on the other end of the line.
Sources: TechRadar · Futurism · Forbes · OpenMuse on GitHub · post by Atai Barkai on X · GeekWire — Todd Bishop
What's in the Lab
New announcements from major AI labs
OpenAI Pledges Deep Access for Outside Safety Auditors
OpenAI published a framework for how outside evaluators should audit its AI models. It says it will grant assessors deep access to independently check whether its safety claims hold up. That includes internal reasoning logs, technical safeguards, and confidential deployment data. The document proposes four priority areas for deeper assessment, plus principles covering independence, scientific rigor, security, and clear responsibilities.
Why it matters: As AI labs move faster and self-certify their own safety work, third-party auditing rules will determine whether outside checks are real oversight or just for show.
What's in Academe
New papers on AI and its effects from researchers
Anti-Flattery Tests May Be Punishing AI for Being Tactful
AI chatbots get criticized for being sycophantic, telling users what they want to hear instead of the truth. New research complicates that critique. Responses flagged as socially sycophantic were also warmer and more receptive. Making a reply warmer without changing its conclusion got it flagged more often. In a preregistered experiment, participants preferred receptive-sounding responses and were more willing to seek advice from their authors, even when they believed the advice-seeker was wrong. The team built a method to boost warmth without sacrificing honest disagreement.
Why it matters: As companies build guardrails against AI flattery, this suggests current flattery tests may penalize the tone that makes AI advice feel usable, not just accurate.
Product Managers, Not Designers, Gain Most From Figma's AI Tool
A randomized controlled trial of 100 product designers and managers tested whether Figma's AI prompt-to-design tool, Figma Make, actually speeds up design work. Among participants who completed the tasks, those with access finished about 20% faster. The gains weren't evenly split. Product managers saved more time than professional designers. The authors suggest such tools may let managers contribute more to design work, while designers' benefits may depend on the task.
Why it matters: If AI tools let non-specialists take on more design work, that reshapes who companies need to hire for design work, not just how fast the work gets done.
One Lecture, Remade for Each Industry at 22 Cents a Minute
Researchers built a system called Bespoke that turns an existing lecture transcript into a new video tailored to a specific industry—complete with fresh slides, narration, and charts—without a professor re-recording anything. Fed 31 graduate lectures on analytics, machine learning, and optimization, it produced 209 industry-customized videos. Twenty-five domain experts rated 92 of the videos, judging 87% at or above standard MOOC quality, with similar performance across industries and lecture lengths. API cost: about 22 cents per minute of video.
Why it matters: If a single lecture can be cheaply remade into industry-specific versions on demand, corporate training and online courses could get far more personalized without adding faculty workload.
What's Happening on Capitol Hill
Upcoming AI-related committee hearings
Wednesday, September 23 — Hearings to examine Flock's nationwide AI surveillance network. Senate · Senate Judiciary Subcommittee on Crime and Counterterrorism (Open Hearing) 562, Dirksen Senate Office Building
What's On The Pod
Some new podcast episodes
How I AI — Opus 5.5 vs. GPT-6 Sol: which model won my blind taste test?
How I AI — I left Claude for months. Opus 5.5 is why I'm back
AI in Business — Managing Change Across Modern Financial Operations - with Nathaniel Bell of Wells Fargo
How I AI — How Warp ships 2,000 PRs a month with AI factories | Zach Lloyd (CEO, Warp)