AI News Roundup: Work, Access Rights and Image Verification

The October 18, 2026 news roundup reviews changes to data terms, AI agent access and how businesses choose models. Meta has removed controversial wording from its commercial terms; Replit has switched to ChatGPT as its default while retaining access to Anthropic models. In technical work, financial analysis and theoretical research, reports show notable results while emphasizing the need for people to review and verify outputs. The roundup also covers the AI Thực chiến competition, Jev’s adoption rate on Vercel’s AI Gateway, Zalo’s two awards and an image of a White House column flagged as AI-generated by a Google tool.

Meta removes controversial wording from its terms

Meta has removed a passage from its commercial terms following criticism over the scope of its use of content. The passage referred to Meta’s ability to use content from domains, websites or URLs for product development, research and training. Meta explained that the consolidated data collection terms apply only to advertising and business products. The removal of the controversial wording and the explanation of its scope need to be distinguished: neither amounts to a commitment to stop all use of data for product development.

Anthropic reports fewer junior staff and more time in meetings

Anthropic has reported changes in staffing and collaboration in an AI-integrated work environment. Since the start of 2025, the share of junior employees in technical departments has fallen from 21% to 14%; the share of juniors among new technical hires has dropped from 27% to 9%. Technical staff have also declined as a share of the total workforce, from 61% to 56%. Meanwhile, meeting time per employee has increased by 20% compared with a year earlier, with a 16% increase for technical individual contributors. Anthropic believes human attention and judgment are becoming bottlenecks as Claude generates more text that needs to be read and organized. The figures do not establish that AI caused these staffing changes.

Codex helps build financial analysis tools, with humans retaining decision-making authority

Anant Kumar used Codex to review code for a tracking tool that uses public SEC filings from business development companies to compare how they value the same loan. The first version was completed in one day; total development time, including additions, was about 10–12 hours. A credit agreement analysis tool takes about 20 minutes to produce a preliminary analysis, which still requires at least an hour of human review afterward. Kumar said Codex helps answer committee follow-up questions more quickly, often during the meeting itself. He emphasized that final investment decisions must be made by people.

AI tested on its ability to solve theoretical problems

Thompson reported that the rate at which the best AI models correctly solved theoretical problems rose from about 2% to 35%. The research team compiled problems from scientific papers published online between December 2025 and May 2026. During testing, the team blocked Internet access to prevent the models from finding solutions online. The study covered only fields in which models could find solutions without simulations or experiments. The results are therefore limited to the ability to solve theoretical problems within the scope of the tests. The researchers emphasized that human oversight and verification of AI results are still necessary.

Businesses restrict AI agents’ access to websites

Businesses are placing different limits on AI agents. Yelp allows only paying AI agents to participate in its data licensing program, while Delta Air Lines prevents agents from booking tickets on users’ behalf. JPMorgan Chase and Charles Schwab confirmed that they block AI agents from accessing accounts because of security concerns. Amazon says continued access by Meta’s Muse agent violates its website policies. These cases involve different restrictions, ranging from data licensing to account protection and policy compliance. They limit the tasks AI agents can perform on users’ behalf.

Replit makes ChatGPT the default while retaining access to Anthropic

Amjad Masad clarified that Replit customers can still access Anthropic models, although OpenAI’s ChatGPT is now the default option. According to Masad, OpenAI is the preferred partner for powering Replit Agent, based on performance, economics and the partnership. Before the agreement with OpenAI, Replit relied on Anthropic’s Claude models to power its platform. This change should be distinguished from dropping a provider: the default has shifted to OpenAI, but access to Anthropic models remains available.

AI Thực chiến selects 100 teams from nearly 1,000 applications

Season 2 of AI Thực chiến attracted nearly 1,000 applications and selected 100 teams for the national final round, held online from October 6 to 8. About 80% of these teams have international awards, published research or products deployed in practice. These are different types of achievements; this does not mean every team meets all three criteria. Techcombank announced a scholarship worth one million USD for the first-place team. The final round therefore brings together many teams with research or deployment results, alongside the opportunity for the winning team to receive the scholarship.

White House column image flagged as AI-generated

Trump posted images on Truth Social showcasing renovations to the White House, including a comparison of an old column with a restored one. Google’s SynthID Detector flagged the column image as AI-generated. The post-restoration image also shows different numbers of steps on either side, differing from the earlier image. However, the tool did not detect SynthID watermarks in other posted images of the White House facade and helipad. Actual photographs show people painting the front of the White House, so it remains possible that the column was repainted. The detection results raise questions about the comparison image; they do not prove that repainting did not take place.

Jev sees rapid adoption on Vercel’s AI Gateway

Vercel data shows that 35% of paying teams used Jev within two weeks of its launch, giving it the fastest adoption rate in the first two weeks among models on AI Gateway. Jev is designed as a lightweight tool for choosing between a set of predefined outputs. This figure reflects usage among the service’s customers, not market share across the entire market. AI Gateway data may underrepresent usage of providers with their own APIs, such as OpenAI and Anthropic. The results therefore show Jev’s early adoption within Vercel, but are not sufficient to rank its overall popularity.

Zalo wins two awards at the Vietnam Digital Awards 2026

Zalo’s suite of solutions for expanding access to technology and Zalo video were honored at the Vietnam Digital Awards 2026 on October 14, both in the outstanding digital technology products, services and solutions category. Features Zalo has launched include video call captions for people over 60 and copying text from images. Meanwhile, Zalo video is integrated directly into the Zalo app and has 40 million monthly active users. The two awards recognize two different product directions: helping users access technology and providing video content directly within the app.

Quick summary

Meta removed a passage from its terms referring to the ability to use content from domains, websites or URLs for product development, research and training. The company explained that the consolidated data collection terms apply only to advertising and business products.
According to Anthropic’s report, the share of junior staff in technical departments has fallen from 21% to 14% since the start of 2025; the share of juniors among new technical hires has dropped from 27% to 9%. Meeting time per employee has increased by 20% compared with a year earlier, while the report identifies human attention and judgment as emerging productivity bottlenecks.
Anant Kumar used Codex to review code for a tool that compares valuations of the same loan using public SEC filings. A credit agreement analysis tool takes about 20 minutes to produce a preliminary analysis, which still requires at least an hour of human review afterward; Kumar emphasized that final investment decisions must be made by people.
According to the article on research into theoretical problem-solving, Thompson reported that the correct solution rate of the best AI models rose from about 2% to 35%. The research team blocked Internet access, considered only fields that did not require simulations or experiments, and emphasized human oversight and verification of the results.
Yelp allows only paying AI agents to participate in its data licensing program; Delta Air Lines prevents agents from booking tickets on users’ behalf. JPMorgan Chase and Charles Schwab block agents from accessing accounts because of security concerns, while Amazon says continued access by Meta’s Muse agent violates its website policies.