Daily AI News — 18 Sep 2026

The United Nations announced it is working with Google to make its vast collection of global statistics accessible to AI systems. The new UN System Data Commons, built on Google's open-source Data Commons platform, lets people and AI agents query statistics across UN agencies in natural language, replacing the older UNData portal. Every figure carries its source so answers can be traced back to the original UN agency. Google.org provided $2 million; the platform runs on a UN-controlled instance. It follows a UNICEF benchmark where six leading models averaged just 21.2% accuracy on 133,000 development-data queries.

As people increasingly ask AI for facts, grounding assistants in authoritative, traceable data is how you cut hallucinated statistics — a template for trustworthy AI across public institutions.

An analyst can ask an AI agent to build a dashboard from UN data and trace every number back to its source agency, instead of manually stitching together fragmented datasets.

Newly unredacted court filings in The New York Times' lawsuit allege that Microsoft and OpenAI internally described the practice of training AI on copyrighted content as theft while both scraped paywalled Times journalism, built datasets from it, and warned internally that the practice would gut publishers. The Times argues the companies knew the conduct was legally exposed. The disclosures sharpen one of the highest-stakes questions hanging over the AI industry: whether frontier models were built on unlicensed content and what that means for the outlets they now compete with for attention.

The outcome could reshape how AI models are legally allowed to be trained, forcing licensing deals with publishers and setting the ground rules for the entire generative-AI economy.

A media company can watch this case to decide whether to license its archive to AI labs or litigate — a live strategic choice now facing every publisher.

Factory, which builds autonomous AI coding agents for enterprises, raised $200 million at a $5 billion valuation — more than three times the $1.5 billion it carried in April. The round was backed by Blackstone, Khosla Ventures, Sequoia Capital, Insight Partners and others, taking total funding past $400 million. The San Francisco company said it will spend the money on research, product and global go-to-market as demand for agents that can write and ship software autonomously accelerates across large organisations.

The valuation jump signals that enterprise buyers are moving from AI coding assistants to autonomous agents that own whole engineering tasks — a shift in how software itself gets built.

An engineering team can hand a Factory agent a well-scoped ticket and have it produce, test and open a pull request, freeing developers for higher-level design work.

Emulate, a UK AI startup founded by former Google DeepMind world-model researchers, is in advanced talks to raise up to $700 million in a seed round, per the Financial Times. Index Ventures and Lightspeed are set to lead, with Creandum participating, at a pre-money valuation reported around $3.7 billion. It is the latest in a wave of enormous early rounds for researchers leaving top labs to build world models — systems that learn a predictive model of physical reality — following David Silver's Ineffable Intelligence and Yann LeCun's AMI Labs.

World models are a leading bet for the next paradigm beyond text-based LLMs — AI that can reason about physical cause and effect, key for robotics, planning and science.

A robotics company could use a world model to let machines rehearse tasks in a learned simulation of the real world before acting, cutting costly trial-and-error.

California Governor Gavin Newsom signed SB 1050, which requires clear disclosure when video or audio advertisements use synthetic — AI-generated — performers. He signed it at SAG-AFTRA union headquarters; the actors' union was the primary advocate. It is one of roughly 20 AI-related bills Newsom has signed into law this year, alongside measures creating the nation's first AI auditor registry and new child-safety standards for social media and chatbots. Newsom has until September 30 to decide the fate of 30-plus more AI measures still on his desk.

As synthetic voices and faces flood advertising, mandatory disclosure protects both performers' livelihoods and consumers' ability to know when they're being sold to by a machine.

A viewer will see a clear label when an ad's spokesperson is AI-generated, and a real actor gains legal footing against unlabelled digital replicas of their likeness.

Comp AI, a Miami AI-native compliance and security platform, raised a $34 million Series A led by Roo Capital and Grand Ventures, taking total funding to $37.5 million. Its AI agents write security policies and gather audit evidence; the company now wants to move from compliance paperwork into continuous cybersecurity — real-time monitoring, control validation and security testing across applications and infrastructure. Founded in 2025, it reports 15x year-over-year ARR growth and more than 1,000 customers, and plans to hire across product, engineering and sales.

It reflects a broader shift from point-in-time audits to always-on, agent-driven security monitoring — turning compliance from a once-a-year scramble into a continuous background process.

A startup chasing SOC 2 can let Comp AI's agents draft policies and continuously collect audit evidence, cutting weeks of manual work before a customer security review.

TypeSafe AI, a frontier lab focused on composable, machine-native intelligence, emerged from stealth with $40 million in seed funding led by DCVC. It was founded by Diogo Almeida — a former OpenAI researcher and co-inventor of RLHF and ChatGPT — alongside Erik Gafni and Sasha Sheng. The company argues that today's models are built to imitate human language, and aims instead to build intelligence designed natively for machines to compose and verify, rather than retrofitting human-shaped systems into software pipelines.

If models can be made more composable and verifiable at the machine level, developers get AI that plugs reliably into software systems — addressing the fragility that plagues agents today.

A developer wiring an AI agent into a production pipeline could rely on machine-native, verifiable outputs instead of parsing and second-guessing free-form text responses.

London-based Magentic, which builds AI agents for procurement and supply-chain teams at large manufacturers, raised an $18 million Series A led by Felicis, with existing investors Sequoia Capital and The Westly Group participating. The round comes a year after the company launched in July 2025 with a $5.5 million seed. Magentic's agents take on sourcing and procurement work for manufacturers, targeting the physical-economy processes — negotiating, comparing suppliers, managing orders — that keep factories supplied but have been largely untouched by generative AI.

It pushes AI agents out of the software world and into the physical economy's back office, where automating slow procurement work can directly cut costs and delays for manufacturers.

A manufacturer's procurement team can deploy Magentic's agents to compare suppliers and manage purchase orders at scale, freeing staff to handle exceptions and negotiations.

Agility Robotics unveiled Digit 5, the next generation of its general-purpose humanoid robot, calling it its first engineered for 'cooperatively safe' work at scale — able to operate close to people without the physical safety barriers traditional automation needs. Digit 5 continuously monitors for people using proprietary AI and multiple sensors, then avoids, stops, or sits. It can lift more and lasts longer than Digit 4, which has logged over 65,000 hours at sites including Amazon, GXO and Toyota. Agility reports more than $300 million in multi-year orders, with customer shipments due in December.

Removing the safety cage is the barrier that has kept humanoids out of most workplaces; solving it is what turns them from demos into machines that can scale on real factory floors.

A warehouse can deploy Digit 5 to move totes alongside human staff in the same aisles, without rebuilding the floor around fenced-off robot zones.

Apex Intelligence completed angel and angel-plus funding rounds raising nearly $50 million in total. The angel round was co-led by IDG Capital, LinkX Capital and XtalPi, with the angel-plus round co-led by Zhongguancun Science City Fund and others. Founded in June 2026 and based in Beijing, the company is building next-generation self-evolving foundation models — using recursive self-improvement, where models improve themselves — with a focus on scientific research and discovery. Its stated aim is to move AI from an assistive tool to a co-researcher that can help discover what is currently unknown.

Recursive self-improvement is one of AI's most consequential and debated frontiers; well-funded labs pursuing it — especially for science — signal where the next capability leaps and safety questions will center.

A research lab could pair with a self-evolving model as a co-investigator that proposes and refines hypotheses, compressing the cycle of scientific discovery.