On October 14, Anthropic published what it had found and changed ahead of the US midterms. OpenAI has published its own election plan. Read together, the two documents are the clearest picture yet of what these companies will and will not do while you are deciding how to vote.
Start with the useful part. Ask Claude or ChatGPT where to vote or what ID to bring, and the answer now comes from Democracy Works, a nonprofit that maintains official voting information, rather than from the model’s memory. On election night OpenAI says it will show live vote counts from the Associated Press. Both companies still ban using their tools to run campaign messaging at scale.
Sources: Anthropic, Oct 14 · OpenAI election safeguards 2026
Then the uncomfortable part, which Anthropic disclosed itself. Foreign-based operators running a fake news site used Claude to produce US midterm coverage pushing Russian, Iranian and North Korean narratives. Anthropic says it banned the accounts. Separately, outside researchers at the Institute for Strategic Dialogue and the McCain Institute tested the safeguards and found the pattern that matters.
Bars show how reliably the safeguards caught each pattern, not a measured rate · Source: Anthropic, Oct 14 2026
One more change worth knowing. Anthropic’s updated usage policy renamed its elections section “Do Not Undermine Democratic Processes” and dropped its blanket ban on personalised campaign targeting, while keeping bans on deceptive targeting and on misusing voters’ personal data.
You are about to be on the receiving end of a lot of political content, and some of it will be made by these tools. The single most useful thing in both announcements is the admission, not the fix: the guardrails are good at catching one bad request and weaker at catching a thousand patient ones. Treat a striking political image or clip the way you would treat a stranger’s claim at a party. OpenAI is also previewing a public tool that checks whether an image came out of its products, which is worth bookmarking before November 3.
Anthropic analyzed roughly 250,000 conversations from one week in September; it reports requests for direct vote recommendations were rare and policy-violating requests rarer still. It claims 99% on its own political evenhandedness evaluation, which is a self-reported internal benchmark, not an independent audit; external reviewers separately flagged uneven hedging in Sonnet 4.5. New classifiers cover multi-step misuse, foreign state actor detection, and cyber protection for election infrastructure. OpenAI layers SynthID watermarking on ChatGPT images on top of C2PA provenance metadata; C2PA metadata can be stripped by re-encoding, which is why the watermark matters.
Anthropic says it cut model copying by more than 90% Distillation
On October 12 the company described its defenses against “distillation,” the trick where a competitor interrogates a finished model millions of times and uses the answers to train a cheap imitation of it. Anthropic says the rate dropped by over 90%.
This is the AI version of a factory discovering its designs were being copied. It tells you something about the business: these companies now treat the answers their models give as the valuable thing to protect, not just the software.
ChatGPT learned to make slide decks Oct 16
OpenAI added a Slides feature for paying Pro, Business and Enterprise accounts, plus nine new voices for its live voice mode.
Unglamorous and probably the change most people will actually touch. The deck you dread making on Sunday night is now a prompt, with the usual caveat that you still have to read every number on it before it goes in front of anyone.
OpenAI bought a company called Pathway ~$400m
Reported on October 20, in a deal said to be worth close to $400 million paid in OpenAI shares rather than cash.
Paying in your own stock instead of money is what companies do when they think the stock is the stronger currency. Worth filing away as a small signal about how this boom is being financed.
The AI learned the language your bank actually runs on.
On October 19 OpenAI published work on teaching its models old programming languages that almost nobody learns any more. The numbers in its own write-up explain why it bothered.
All run on COBOL, first released in 1959 · Source: OpenAI, Oct 19 2026
Fortran, from 1957, is still doing weather forecasting and nuclear reactor safety analysis. These are what the industry calls “low-resource” languages, meaning there is very little public code for an AI to learn from, so models have always been bad at them. OpenAI built two new tests and reports its newest model migrating old COBOL programs to a modern language at 86.6% accuracy.
There is a real and unglamorous problem here. The people who maintain this code are retiring faster than anyone is replacing them, and banks and government agencies have been quietly worried about it for years. If this works, it is one of the few places where AI is filling a gap rather than taking a seat. If it half-works, the failure mode is a bank’s core system rewritten by something that was right 86% of the time, which is a different kind of problem.
The two evaluations are MultiLangBench, which asks a model to build a SQLite-compatible database engine from scratch across 20 languages, and a COBOL migration benchmark testing behavior-preserving translation. OpenAI reports GPT-6.1 Sol at 86.6% on the migration task at low reasoning effort. The post also reports large MultiLangBench gains between model generations; those figures and the per-language pass rates are measured differently and should not be read as one number.
Who gets to check the labs’ homework?
On October 15 Anthropic said it has given METR, an independent AI evaluation nonprofit, “employee-like access” to its systems. Six METR staff have been onboarded with company hardware, office space and access to internal tools, and METR will publish its risk assessments. Anthropic also names SecureBio and Accenture as evaluators and says it is in early talks with others, including Hugging Face.
This lands on an argument this newsletter has followed since Issue #9, when OpenAI announced a headline score of 99.9% on a reasoning test and the independent organization that runs that test published 62.7% for the same model under its own standard setup. Same model, different harness, a 37-point gap.
“Independent evaluation” has been doing a lot of work in these companies’ safety claims while the evaluators mostly saw what the companies chose to show them. Badging six outsiders into the building and letting them publish is a real step, and more than any rival has offered. The part that has not changed: this is access one company chose to grant and could withdraw, on its own timetable, with no regulator requiring any of it. Good practice is not yet a standard.
When an AI does the math, who checks it?
On October 15 OpenAI committed $5 million to a Simons Foundation initiative to help mathematicians study results produced by AI systems. It follows the company’s October 6 claim of progress on open mathematical problems.
This is the right order of operations, and it is rarer than it sounds. A machine producing a proof is only useful if humans can verify it, and funding the checking is less exciting than announcing the breakthrough. Treat the October 6 claim as unverified until that work happens. The pattern repeats across AI in science: the announcement arrives months before anyone can confirm it.
Anthropic puts some of itself into a fund.
On October 18 Anthropic said it is committing a portion of its own shares to a new Anthropic Public Benefit Fund. The company has still not filed a public IPO prospectus. It filed a confidential draft with the SEC on June 1, and reports since have pointed to a listing as early as this autumn, but nothing public has appeared.
This is a company arranging who owns it shortly before asking the public to buy a piece. The honest summary: we know it filed confidentially in June, we know bankers have been briefing an autumn timetable, and we do not know the date, the price or the exchange. Anyone telling you otherwise is guessing.
Singapore signs up.
On October 14 Anthropic opened an office in Singapore, its fifth in the region, and signed a memorandum of understanding with Singapore’s Ministry of Digital Development and Information covering AI safety, security and model evaluation. A week earlier, on October 8, it committed $150 million over three years to the Genesis Mission, a US federal research effort, part of it to put Claude inside more than 15 government agencies.
These companies are now doing the kind of thing countries do: opening offices, signing agreements with governments, funding national science programs. A memorandum of understanding is not a binding contract, so read it as positioning rather than policy. The pattern to watch is which governments end up dependent on which company’s models, because that is harder to reverse than any regulation.
- November 3The midterms, the first federal US election run with all of this in place. The thing to watch is not whether a deepfake appears, it is whether the slow, many-small-requests pattern the outside testers found gets used at scale.
- November 2The Challenger job cuts report for October, the closest thing to a monthly read on whether AI is still being named as a reason for layoffs. September’s cuts were 43,281, down 18% from August.
- When it landsAnthropic’s public prospectus. That is the first time anyone outside the company sees its real numbers.