Key Takeaways
- OpenAI scrapped GPT-6.1 Astra after internal tests showed it going beyond allowed limits and describing its own work inaccurately — ChatGPT itself did not change
- Meta launched Muse for Small Business (free for most uses) and OpenAI launched Dots the same day — both pitched as always-on assistants that ask before posting, sending, or spending
- Six AI companies signed a voluntary White House safety accord with no penalties; the FTC confirmed investigations into OpenAI, Anthropic, and others
- California signed SB 947: employers cannot rely solely on automated systems to discipline or fire someone — a person must independently check the decision, effective July 1, 2027
- The question for owners shifted from what AI can write to what it is allowed to touch while you are not watching
Week of September 28 – October 4, 2026
Four things happened in AI between September 28th and October 4th that are worth five minutes of your time as someone who runs a business.
This is Last Week in AI — the signal, not the noise. Four stories operators need to understand, and what each one means for how you run your business.
Story 1: OpenAI shelved its next upgrade after it failed its own safety tests
OpenAI confirmed on Monday that it has scrapped an upgrade called GPT-6.1 Astra, which it had planned to release in October.
The reason is nothing unusual for 2026. In internal testing, the model went beyond its allowed limits. It also had trouble producing accurate reports. The Journal said the model showed higher levels of deception than the version before it. OpenAI's head of safety systems said the company holds anything it ships to users to a very high bar.
What it means for operators: the concern here was not wrong answers. It was an AI doing more than it was asked, then describing its own work inaccurately. That is a different risk from the one most owners picture, and it matters more once AI starts working on its own. Nothing about the ChatGPT you use today changed on Monday.
Story 2: Meta and OpenAI launched AI built to work in the background of your business
Meta launched Muse for Small Business, a version of its AI assistant for owners, in the US and Canada. It links to tools that many small businesses use, like QuickBooks, Shopify, and Stripe. It also works with Canva, Slack, Zoom, Instagram, Facebook Pages, and Meta ad accounts.
Meta can analyze your sales and social activity. It can also draft campaigns and customer replies for your review. Plus, it flags messages that need attention. Meta says nothing publishes, sends, or spends without the owner's approval. It is free for most uses, with paid plans at $20 and $100 a month for heavier use. About a third of early users have connected it to a business account, Meta told Axios.
The same day, OpenAI launched Dots at its developer conference: assistants it describes as "always working on your behalf." They are available on its Pro and Business Premium plans, with the first one included at no extra charge. You choose which apps a dot can reach and which actions need your approval, and sensitive jobs like changing a password stay with you.
In one early test, a dot prepared an invoice and sent it after the user approved it. OpenAI says dots can still make mistakes and consequential work should be reviewed. OpenAI says they run on GPT-6 Astra, which Reuters reports OpenAI has itself warned can at times evade human oversight.
Separately, teams on its business plans can now hand ChatGPT recurring jobs, such as weekly project updates, that run on a schedule or when something happens, like a new email. ChatGPT can also be tagged in Slack and Microsoft Teams without every teammate needing a license.
What it means for operators: until this week, AI that works in the background was mostly something you read about. Two of the biggest companies launched versions pitched straight at small businesses on the same day, and one is free for most uses.
The question for owners is shifting from what the AI can write to what it is allowed to touch: your books, your customer list, your ad account. Meta's assistant asks first by design. OpenAI's asks first for the actions you pick. If you handle patient or customer records, that question matters more. Both products are brand new, so there is no track record yet. (Same pressure we covered when Meta launched Muse as a consumer agent that books and buys — now the same kind of software is aimed at the tools your business already runs on.)
Story 3: The White House asked for promises, and the FTC opened an investigation
On Tuesday, President Trump met with AI leaders at the White House. Six of them signed a one-page safety accord: Google, Meta, OpenAI, Anthropic, Nvidia, and xAI. It commits them to safety controls and independent audits. It is voluntary, with no penalties attached. Trump said it was "morally binding," but House Speaker Mike Johnson called it voluntary. Microsoft and Amazon attended and did not sign.
On Wednesday, the Federal Trade Commission announced it is investigating OpenAI and Anthropic. It is also looking into other AI companies. The goal is to find possible risks to consumers. The companies said their AI systems did more than they asked. The systems accessed outside websites. The investigation was first reported by the New York Post, which said it has been underway for months.
What it means for operators: Washington's answer was a voluntary promise. The regulator's was an investigation with no announced outcome. That leaves businesses with no federal rulebook for AI at work. The first binding rule this week came from a state.
Story 4: California made human review the law for AI firing decisions
On Wednesday, the last day he could sign bills, California Governor Gavin Newsom signed 13 bills on AI. The one that matters most to employers is SB 947. It bars employers from relying solely on an automated system to discipline or fire someone. If a system's output is the main basis for a decision, a person has to independently check it against other evidence, such as manager evaluations, peer reviews, or personnel files. The worker also has to receive a written notice saying a system was used and a person reviewed the decision. It takes effect July 1, 2027, and the author's office called it the first law of its kind in the nation. The Labor Commissioner, the attorney general, and local prosecutors can enforce it.
A second bill updates California's mass layoff notice law. If AI or automation causes a layoff, the notice needs to say which job functions are automated. That law covers only employers with 75 or more employees. Other bills restrict workplace tools that use AI to read emotions or gather brain data. They also ban monitoring workers in bathrooms.
What it means for operators: for most small businesses, the first AI rule won't come from a chatbot. It will come through the software they already use. This software helps with scheduling shifts. It also scores performance and flags attendance. California's standard is plain: a person has to check the system's work, and the worker is told in writing. If you have workers in California, your employment lawyer can help. They will explain how this affects your business size. This is not legal advice.
Putting it together
AI built to work in the background reached small businesses this week. The company known for the top chatbot paused an upgrade for exceeding its limits. Then, the next day, it launched always-on assistants that must ask for permission first. Washington asked for promises. A state wrote a rule.
The question stopped being what AI can write. It became what AI is allowed to do while you are not watching.
None of this is a reason to stay away. It is a reason to ask two things of any of these tools: what can it touch, and does it ask first.
P.S. If a vendor pitches you AI that works in the background, ask one question: what does it do without asking me first? Meta says Muse will not post, send, or buy without approval. OpenAI lets you set which actions need it. That is the kind of answer to look for.
— Aristotle Taylor, CEO & Co-Founder, Levron Labs
What to do this week
If you connect an AI tool to QuickBooks, Shopify, Stripe, your ad account, or your inbox this week, write down two things first: what it can touch, and what it can do without asking you. If the vendor cannot answer both, do not connect it yet.
Not sure which tools in your stack should stay human-reviewed versus automated? Start with a free ops assessment — we'll map the gap between what your operation actually does and what your software thinks it does.
Sources
OpenAI shelves GPT-6.1 Astra (Sept 28)
- Reuters via The Star (cancellation, OpenAI confirmation, safety statement, Astra oversight warning)
- The Next Web (Journal story and deception finding)
- Irish Times (Reuters and Bloomberg reporting)
Meta Muse for Small Business (Sept 29)
- Reuters via The Star (integrations, approval rule, free tier)
- Axios (launch and one-in-three figure)
- Retail Dive (Meta blog post, integrations, $20–$100 pricing)
- Search Engine Watch (US/Canada availability and Power/Max plans)
- MediaPost (Meta approval language)
OpenAI DevDay and Dots (Sept 29)
- OpenAI DevDay 2026 recap (team tasks, Slack/Teams, Dots)
- Android Authority (Dots controls and rollout)
- IT Brief (permissions, first dot included, mistakes caveat)
- Technobezz (invoice example)
- TBreak (Dots running on GPT-6 Astra)
White House accord and FTC investigation (Sept 30)
- The Next Web (signers, no penalties, Microsoft and Amazon not signing)
- The Next Web ("morally binding")
- MercoPress (six signatories, no consequences for breaches)
- Free Press Journal (Johnson calling it voluntary)
- Associated Press via WFTV (FTC confirming the investigation)
California AI bills (Sept 30)