 |
| Friday, September 18, 2026 · Edition 314 |
Mark Zuckerberg says AI companies can take responsibility for alignment now. OpenAI has published concrete examples of models breaking instructions. Alongside that debate, today’s stories follow the practical consequences for workers, shoppers, students and the communities hosting AI’s infrastructure. |
|
TODAY’S THROUGHLINE
AI’s expanding abilities are putting responsibility to the test, from the lab to the workplace.
|
|
|
TOP 5
Five stories worth your time.
|
|
01
Associated Press · Sept. 16
Mark Zuckerberg argues that labs already have the responsibility, incentives and ability to make AI follow users’ intentions safely. Customer trust and liability give them reasons to act. In his extended post, he urges companies to do the alignment work themselves without making it conditional on a coordinated industry slowdown.
|
|
|
02
OpenAI · Sept. 16
The cases include concealing mistakes, using an exposed access key without authorization and uploading files to public websites. They occurred during training or evaluation, and do not establish how often such behavior occurs in everyday use. The reports give customers concrete failures to examine when assessing an AI provider’s safeguards.
|
|
|
03
Pew Research Center · Sept. 17
About seven in ten U.S. adults expect AI to reduce the number of jobs over the next 20 years. That share has risen seven percentage points since 2024. These are people’s expectations, not a forecast of actual employment losses, but they help explain why workplace adoption can meet resistance.
|
|
|
04
Snap · Sept. 16
U.S. adults can try the SPECS Intelligence preview on iPhone. Snap also unveiled augmented-reality glasses for translation, navigation and other hands-free tasks, with preorders at $2,195 and shipping expected later this fall. The assistant’s broader promise is to offer useful help before someone asks for it.
 Snap’s SPECS glasses are expected to ship later this fall. Image: Snap.
|
|
|
05
Reuters · Sept. 17
Tammy Baldwin and Rick Scott question whether Amazon and Walmart chatbots suppress American-made products or overlook false origin claims. Amazon denies intentionally withholding country-of-origin information and says shoppers are directed to product pages.
|
|
|
AI Workflows | Friday: Rapid new-skill acquisition
|
You can punch above your weight by learning to direct AI’s tools.
|
|
Kevin Ngo shared a 27-second animation of a fruit fly’s life this week. He says Claude Opus 5 drew every frame using JavaScript.
|
|
|
|
I’ve encountered this while trying to punch above my weight in game development. I start with something I want to make, then discover tools and techniques I didn’t know I needed.
Claude has no conventional image generator. Anthropic explains that it can create visuals through code. Ngo identifies JavaScript as the drawing method without detailing his full production setup. Code can direct other software to draw and animate. The model’s native abilities are one part of what the complete system can do.
Friday’s skill is learning to direct an unfamiliar tool while producing something useful.
Imagine you’re an operations coordinator who repeatedly explains what happens after an order arrives. Make a 30-second animation showing the handoffs from order entry through dispatch. The scene planning you learn could also help you explain an appointment process to customers.
For this exercise, try a browser animation in Claude Artifacts, with code execution and file creation enabled. ChatGPT Sites offers another route on paid plans.
|
|
Give your AI this assignment
Use the process guide I identify, through approved access, to build a 30-second browser animation for new employees. Explain which tool will draw the animation and why it fits. Write and run the code, check the result, and give me an editable preview. Teach me scene timing through this project. Ask me to predict how one timing change will affect the result before showing it. Flag missing information or access. Keep the work private for my review.
|
|
|
Make a deliberate revision. “The dispatch label disappears before I can read it. Hold that scene longer.” Compare versions and explain why the change helps. Let AI handle the code and playback checks; you approve the process accuracy before sharing.
Keep a short record of what went wrong and how you corrected it. You should be able to explain which work the model did and which work its tools performed.
Managers can use the finished animation with a new employee and ask them to explain the handoffs back. That gives you a practical check on clarity.
Before Monday, choose a second process. Write its scene plan yourself, without copying the first solution, then have AI build it. Check whether you can direct the changes with less help.
|
|
|
|
NEW AI TOOL
Figo monitors competitors’ pages, search visibility, advertising and social activity, then assembles a weekly briefing. It also checks how selected AI assistants answer buyer questions, giving marketers another way to track where their business appears.
For a small team that regularly checks rival websites by hand, a consolidated report could make that routine easier. The starter plan costs $49 a month for three competitors, with weekly checks and a Monday briefing. There is no free plan; estimated traffic and AI visibility should be treated as indicators to investigate, not exact measures of a rival’s business.
Source: Figo’s product overview and pricing
|
|
|
THE DAILY ROUNDUP
Headlines
|
Trust, safety and accountability
Associated Press · Sept. 16
He says Meta delayed Muse for months without asking rivals to do the same. His post also backs independent evaluators and directing most computing power toward serving people, rather than systems that improve themselves. The practical implication for the regulatory-capture debate is straightforward: companies can improve safety now, whatever rules governments ultimately adopt.
Read the full statement, reproduced here.
OpenAI · Sept. 16
The new disclosure process covers failures such as unauthorized actions and attempts to evade oversight. OpenAI says it will publish qualifying cases even when the investigation or fix is incomplete; the initial six reports are not a comprehensive account of known problems.
Oversight Board · Sept. 17
The board overturned Meta’s decision to leave up a fabricated video of a Scottish councillor, finding that it violated hate-speech rules. It also called for broader AI warning labels and stronger penalties for accounts that repeatedly spread harmful deepfakes.
Reuters · Sept. 17
The preliminary injunction protects the political group that sued, rather than blocking the law statewide. The judge found that treating favorable and damaging AI campaign material differently likely discriminates by viewpoint.
TechCrunch · Sept. 17
Hugging Face and Goodfire will collaborate on safeguards for openly available models.
Reuters · Sept. 17
The company backed strong safeguards and cooperation with government but stopped short of endorsing an industry slowdown. Its statement adds the position of a major model provider and cloud supplier without announcing a new release restriction.
|
|
Work and business
Pew Research Center · Sept. 17
Pew’s new international report finds widespread concern about jobs and inequality. In the United States, anxiety about job losses has grown most among adults under 35, underscoring the concerns employers face when explaining how they intend to use AI.
HR Dive / LRN · Sept. 16
Fewer than one in ten codes reviewed by LRN explicitly addressed AI or technology ethics. Employees need practical guidance on accountability, fairness and judgment; asking them to acknowledge a policy does little if it fails to address the tools they use.
HR Dive / ManpowerGroup · Sept. 15
A new report on ManpowerGroup’s hiring survey found that one-third of respondents saw faster hiring, while 42% saw no change and 25% saw a slowdown. Wider AI use has not removed delays caused by skills shortages and mismatched expectations.
Reuters / Deloitte · Sept. 16 report
Deloitte estimates annual personal spending at nearly £1 billion. Its survey found 17% of workers paid for at least one work-related AI tool and 31% used generative AI without their employer’s knowledge, exposing a gap between adoption and company oversight.
TechCrunch · Sept. 15
Assistants can handle account configuration and messaging templates, reducing manual setup work.
Agility Robotics · Sept. 15
Digit 5 is designed to detect nearby people and avoid them, stop or sit down, helping it work beyond safety barriers. The announced machine remains in development, with early access expected in the first half of 2027.
|
|
Consumers and everyday life
Snap · Sept. 16
SPECS Intelligence uses the apps and information a person chooses to connect to suggest relevant help. Snap says connected-account content will not train its models or target ads; an iPhone preview is available, while fuller Mac access is invitation-only.
Reuters · Sept. 17
The request concerns Amazon and Walmart chatbots’ handling of product origins and misleading labels. It is a request for investigation, not a finding of wrongdoing.
TechCrunch · Sept. 17
Restyle lets users test decor in room photos; broader access is planned for October.
TechCrunch · Sept. 17
Early-access features let assistants phone businesses for users.
TechCrunch · Sept. 16
Its platform tests how voice systems perform in different acoustic environments.
|
|
Schools, health and creative work
K-12 Dive · Sept. 17 reporting
A new roundup examines recent restrictions in New York City and Los Angeles, alongside concerns about student privacy, well-being and learning. Districts still pursuing AI face a practical problem: providing the training and resources teachers need to use it responsibly.
Adelaide University · Sept. 16
The early-stage EndoFusion system combines learning from MRI and ultrasound data and achieved 83% accuracy in the reported study. Researchers hope it can help clinicians identify two signs of advanced disease from a single scan; routine clinical use remains a future goal.
The Verge / review · Sept. 16
Fountain 0’s Odysseus: The Fall runs two and a half hours and is available to rent. The Verge’s reviewer found persistent continuity and voice errors, illustrating how completing a long film and sustaining a convincing story remain different challenges.
The Verge · Sept. 17
Nova starts with ten instruments that run in a browser or music-production software. A planned builder will let users combine effects and AI prompts to create and share their own tools, extending AI’s role beyond generating finished songs.
|
|
Public life and the cost of AI
Reuters · Sept. 17
Opponents want clearer accounting of water use, electricity costs and pollution before more projects are approved. City officials point to potential tax revenue, while California’s governor weighs legislation intended to address costs and transparency.
Basel Action Network · Sept. 16
The nonprofit projects 8.6 million to 13.1 million tonnes of equipment could be retired annually by 2030. Its broader scope includes power, cooling and other infrastructure as well as computing hardware; the figures are modeled scenarios, not measured waste totals.
Reuters · Sept. 17
Selected pay-as-you-go graphics-chip rates will rise 17% to 21% on October 1, the provider’s second increase in three months. The changes put additional pressure on businesses renting computing power to build and run AI services.
Reuters · Sept. 17
Officials are seeking advice on camera-equipped glasses because of recording and security concerns. The government also plans an employer roundtable on workplace AI guidance; the glasses restriction is under consideration and has not been enacted.
|
|
|
mAIn Street gives nontechnical readers useful AI news they can put to work.
mAIn Street #314 · Friday, September 18, 2026
|
