AI News: August 23, 2026

1. Tesla Approved to Deploy Up to 5,000 Robotaxis in Nevada

Autonomous Vehicles. The Nevada Transportation Authority approved a permit on August 20 allowing Tesla to operate up to 5,000 robotaxis in Clark County, which includes Las Vegas, over the next 12 months. Nevada becomes the third US state after Texas and Florida to clear Tesla’s autonomous ride-hailing service, which will start with Model Y vehicles before transitioning to Cybercabs. Commercial operations are expected to begin within 30 days pending vehicle inspections and insurance filings. Source

2. Inherent Launches Faraday Agent for Replicating Research

AI Agents. Inherent, a British AI lab founded by former Google DeepMind researchers, released Faraday, an AI agent built to reproduce published scientific papers. The company said Faraday outperformed systems from Anthropic and OpenAI at replicating research results, which it frames as a step toward automating parts of the scientific process. The comparison is based on the company’s own evaluations. Source

3. Study Finds Few FDA-Cleared AI Devices Tested on Patient Outcomes

Healthcare. A study published in PLOS Digital Health found that of 1,357 AI medical devices cleared by the FDA, only three were tested against actual patient health outcomes such as death, hospitalization, or quality of life. Just 34 were linked to registered prospective trials and only 12 had peer-reviewed publications, with most supporting studies enrolling fewer than 500 participants. The authors attribute the evidence gaps in part to the FDA’s 510(k) pathway, which lets devices clear by showing equivalence to existing products rather than proving clinical benefit. Source

4. OpenAI Urges California to Strengthen Its AI Safety Bill

Government & Policy. OpenAI publicly called on California lawmakers to strengthen SB 53, a state AI safety bill the company had previously opposed. SB 53 would impose safety and transparency requirements on developers of frontier AI systems. The reversal signals a shift in how a leading model developer is positioning itself on state-level AI regulation as the bill advances. Source

5. Study Says Frontier Labs Lack Plans to Contain Rogue Models

AI Safety. A new study found that leading AI labs have few publicly documented plans for containing a model that behaves unexpectedly or dangerously. The analysis covered companies including Anthropic, Google, Meta, OpenAI, and xAI and raised questions about industry preparedness as systems grow more capable and autonomous. The findings add to scrutiny of how labs would respond to loss-of-control scenarios. Source

6. Netflix Tests In-House Language Model for Recommendations

AI in Practice. Netflix said an in-house language model called GenRec outperformed its long-standing recommendation engine in internal tests. Rather than relying on thousands of hand-crafted features, GenRec converts a member’s viewing history into plain text and lets the model reason over it. Netflix described the approach as an early but promising step toward replacing hand-built recommendation logic. Source

7. Research Explains Why Agent “Skills” Help and When They Fail

Research. Researchers at Princeton University and UC San Diego found that so-called skills improve AI agents mainly by supplying structured workflows rather than new knowledge. The study also showed that as an agent’s skill library grows, the agent has increasing trouble selecting the right instructions for a given task. The results have practical implications for teams building agents around large libraries of reusable skills. Source

8. UK Institute Finds Safety Benchmarks Do Not Measure One Trait

Research. The UK AI Security Institute used psychometric methods to show that popular language-model safety benchmarks do not measure a single consistent trait. The researchers found that blanket refusal of requests can inflate a model’s safety score even as the model becomes less useful, and they proposed a method for detecting models that act more cautious during testing than in normal use. The work points to weaknesses in how model safety is currently evaluated. Source

9. Research Adds Human Beliefs to AI World Models

Research. A new framework called Mental World Modeling argues that world models such as those behind video generators simulate physics but ignore human beliefs, intentions, and goals, leading them to predict the wrong actions. Adding mental variables let even weaker language models outperform stronger models that lacked them on the researchers’ tests. The main bottleneck was jointly predicting how physical and mental states change together. Source

10. Privacy Backlash Grows Over Meta AI Glasses

Privacy. Ars Technica examined a growing privacy backlash against Meta’s AI glasses as sales climb and the devices become harder to spot in public. The piece reviewed apps such as Zuckoff that try to detect nearby smart glasses, noting the detection tools are imperfect, and it documented bans at venues including schools, courts, and DEF CON 2026. The trend highlights unresolved consent and surveillance questions as camera-equipped wearables spread. Source