New chatbot protections, late earnings reports and useful AI tools for work.
 |
| Friday, September 11, 2026 · Edition 309 |
|
Today’s throughline
AI is taking on everyday tasks as states demand stronger protections for the people using it.
|
|
|
Top 5
Five stories to start your day.
|
|
01
Gov. Gavin Newsom signed laws requiring independent child-safety audits, annual risk assessments and stronger protections for young chatbot users. The package also restricts addictive social-media features for children under 16.
|
|
|
02
OpenAI’s new Data agent answers business questions using approved company systems and an organization’s own definitions. It follows existing access permissions and can turn an analysis into a dashboard that teams can share and refresh.
|
|
|
03
Gusto linked survey responses from 2,262 businesses with payroll records and found that AI adopters had about 7% higher employment after a year than comparable nonadopters. The observational study cannot establish whether AI caused the difference.
|
|
|
04
Oracle reported $19.3 billion in quarterly revenue and more than $30 billion in new AI cloud contracts. Its results show growing demand alongside the heavy cost of building data centers.
|
|
|
05
The open-source model combines observations from lunar missions to identify craters, study volcanic terrain and locate promising areas for ice research. Better maps could help scientists plan where future astronauts explore and establish a base.
 Yellow and white mark areas the model considers more likely to contain ice. Image: NASA, via IBM Research.
|
|
|
AI Workflows | Friday: Rapid skill acquisition
You can learn to test the AI projects you’re proud of.
OpenAI’s latest Astra update describes AI that can carry out complex tasks and check its work. That gives us a testing partner while we learn to judge the results.
It’s easy to get excited about what we “created” with AI. A working game or website can feel like proof that we’ve done something impressive. We spend too long admiring the result and too little time checking how it behaves.
I’ve run into this while building video games with AI. Characters have body types that don’t match their descriptions. Their movements look wrong when you watch closely. Small details that barely register at first become hard to ignore once you start using what you made.
Friday’s skill is learning to test those details systematically. Start with one feature and write down what should happen. Give ChatGPT Work access to the project, then ask it to inspect the feature and reproduce the problem. Have it explain the mismatch and what else a fix could affect.
Give the small features a turn.
| Feature |
What to try |
| Appearance |
Compare the character with its written description. |
| Movement |
Watch starts, stops and turns, including quick changes. |
| Controls |
Change one setting. Check that the intended behavior changes. |
| Saved choices |
Reload. Check which settings should persist. |
For a project in ChatGPT Sites, request a saved version for review before deployment. Every Sites deployment is live. Keep publishing as a separate decision.
|
Give your AI this assignment
Help me learn to test this project. Use its existing files and requirements. Reproduce the problem I identified and compare it with my expected behavior. Explain the cause in plain English. Make the fix in a saved version, then test it and any connected features. Keep a short log of what failed, what changed and what passed. Separate verified results from checks you couldn’t complete. Ask me to explain why the fix works. Wait for my approval before publishing.
|
Try the corrected feature yourself. If a character’s appearance changed, check whether its movement still works. Add your observations to the test log. Understanding how one change affects another is part of what you’re learning.
At work, apply this to a customer form or dashboard filter. Managers can make the test log part of the handoff, giving a colleague something specific to review.
Before sharing your next project, choose a different feature. Write its expected behavior and design a test before asking AI for suggestions. Run it, compare notes, and explain what would count as a passing result.
|
|
|
|
AI Tool Spotlight
Give your audience captions in their language.
|
|
Subanana produces live captions in the speaker’s language and up to five translated languages. Attendees can scan a QR code or open a link, then choose captions on their own phones; organizers can also show captions on a projector or livestream.
It is a practical option to test for community meetings, training sessions and events with multilingual audiences. The site offers a five-minute free trial, enough to check the microphone, names and specialized vocabulary before a longer rehearsal.
|
|
|
Headlines
And that was only the Top 5
|
People, skills and public services
Among businesses with fewer than 10 employees, employment was roughly 10% higher after a year for AI adopters. Gusto checked the result using several comparisons, although differences between businesses could still influence the findings.
In an iCIMS survey of 1,000 U.S. job seekers, 30% said they taught themselves AI skills, up from 22% a year earlier. Employer-provided training remained near one in six respondents, leaving many applicants to prepare on their own.
Deloitte and the Manufacturing Institute argue that AI could help workers in adjacent industries learn manufacturing tasks by making technical knowledge easier to find. The study’s figures describe the potential worker pool for this proposed training approach.
In a recent company analysis of 2.5 million submissions, 84% reviewed with BusyBee received written feedback, compared with 29% reviewed without it. Those figures measure feedback coverage; teachers still approve every grade.
The partnership supports a campus AI platform, research infrastructure and skills credentials at the Baltimore university. Planned work includes an AI center of excellence and applications such as studying the city’s traffic.
AskCA organizes help around needs such as finding work, starting a business and recovering from disasters. Pilot registration opened September 9 for California residents 18 and older, with a related state-job matching feature planned for September 30.
The Sanas rollout combines speech processing with transcription, call summaries and compliance alerts. The company says processing on each agent’s device will support clearer calls while reducing the need to move sensitive audio between systems.
The nonprofit platform recommends nearby programs, classes and events based on people’s interests. It reports participation from more than 1,000 nonprofits and is developing an assistant to answer questions about civic life.
|
|
Tools for daily work
The Data agent lets employees ask questions across connected business systems without writing database queries. Its answers depend on the organization’s approved sources, shared definitions and existing access permissions.
Existing customers can develop campaigns in Claude with desktop and mobile previews, then finish and send them through Constant Contact. The connection reduces the copying and reformatting involved in moving a draft into a marketing platform.
Its new workflow analysis examines tasks and handoffs across applications so teams can choose where to make changes and measure the results. Select customers have early access, with general availability planned for January 2027.
New connections let authorized assistants query records such as tracked hours and missing timesheets. Managers can ask a business question directly and use the underlying records to check the answer.
Its planned Walk-and-Talk feature connects spoken updates and images to locations on a job site. The broader preview also includes indoor positioning on building plans, with selected capabilities available through early-access programs.
GPT Live 1 gives developers more natural conversation handling, including interruptions, pauses and background noise. Customer-service and tutoring applications could use it to reduce the awkward turn-taking that makes automated conversations difficult.
|
|
Safety and accountability
The September 10 package adds parental controls and crisis-response protocols for companion chatbots, alongside independent safety checks. It also strengthens protections for student information used in AI systems and addresses AI-generated child sexual abuse material.
Its September 9 assessment describes a January incident and revisits three previously reported cases. The cybersecurity tests had unintended internet access and lacked standard product safeguards; METR will conduct an independent investigation.
Its September report includes five cases of biological work that could support weapons development, alongside cyber and surveillance operations. Anthropic banned the accounts involved in those biological cases but explicitly stopped short of alleging the scientists intended harm.
Two laws signed September 9 establish a framework for independent verification organizations and a state registry for AI auditors. They set standards for outside assessors examining AI systems for compliance with state law.
Among 700 technology professionals at large enterprises, 76% believed they could disable an agent within 15 minutes, while 33% reported an instant shutdown control. The commissioned survey covered organizations already testing or using agents across five countries.
The announced capabilities tie agents to human owners and give administrators ways to track permissions and revoke access. A further connection for managing privileged access is due before the end of the fourth quarter.
The service combines identity verification with background and rental screening and is available nationwide. In a company survey of 1,400 U.S. landlords and property managers, 67% said AI-driven fraud made genuine applications harder to distinguish from fake ones.
|
|
Business and infrastructure
Capital spending reached $28.5 billion, including $11.36 billion covered by customer prepayments. The results put the financing of AI infrastructure alongside its sales growth.
The reported talks with Fluidstack concern support for the U.S. data-center supply chain. A loan has not been announced, and Reuters said it could not independently verify The Wall Street Journal’s report.
The offering combines financial data sources with tools for research, models and presentations that follow a firm’s templates. Its initial focus is investment banking and equity research, where teams need to trace figures back to source material.
Analysts and portfolio managers will use Claude to synthesize information and support research, while developers use Claude Code to build investment tools. The firm says business teams retain responsibility for each application and the judgment behind investment decisions.
|
|
Science and creative work
Researchers can adapt the model to several mapping tasks using the same combined observations. Its ice predictions identify places for further investigation, while its terrain analysis could help future crews navigate difficult lighting and surface conditions.
Adobe reported more than 150% year-over-year growth in its AI-first annual recurring revenue. The company is expanding AI features across its creative software and customer-experience products.
The planned DeepSky constellation will combine five types of instruments to observe storms and atmospheric conditions, especially over poorly monitored oceans. The company says the system is in production; Thursday’s announcement details the design.
The partnership plans to let fans create remixes and other interpretations using participating artists’ music. The service is still in development, with its own licensed catalog and permissions separate from ElevenLabs’ existing music products.
VideoClaw helps users identify topics, develop ideas and continue into scripting, visuals and editing within the same service. For a small marketing team, it offers a shorter route between researching a subject and preparing a video.
|
|
|
mAIn Street gives nontechnical readers useful AI news they can put to work.
mAIn Street #309 · Friday, September 11, 2026
|
|
|
|
|
