University of Washington study finds accent bias in AI models

AI news spans market swings, research, product controls and public policy. Asian markets gave mixed signals Wednesday after AI stocks led Wall Street lower: shares in South Korea, Japan and Taiwan rebounded in one account, while another reported most Asian markets fell and South Korea’s Kospi moved between gains and losses. U.S. futures edged higher and oil prices fell.

Joseph F. Krause, co-founder and CEO of Radical AI, says materials discovery can take 10 years and cost up to $100 million, holding back applications in semiconductors, batteries, data-center power, fusion and other industries. He argues self-driving labs can search research, make and test materials, then learn from experiments. Krause points to a New York lab that discovered a material in a campaign for hypersonic applications, and says faster research could help the United States turn AI advances into products sooner.

OpenAI says Codex agents run on users’ machines, with workspace profiles restricting where they can write and a harness enforcing command permissions, including for child processes. Users can turn network access off, allow direct access or proxy specified domains, and customize profiles. OpenAI demonstrated a risk involving an agent changing an ignored dependency such as node_modules/prettier, which a host editor could run outside the sandbox; the company did not publish figures on bypasses, approval prompts or sandbox overhead.

Researchers at the University of Washington found accent-based differences in scores from Google Gemini 2.5 Pro and Flash, OpenAI GPT-Audio-1.5 and Alibaba’s Qwen3-Omni. Testing synthetic and human voices reading the same words, they found American and British accents ranked highest, Indian accents in the middle, and Chinese and Nigerian accents lowest. The models understood the speech but judged delivery and content differently; explanations acknowledged accent bias only 10% to 30% of the time. The study was presented in October at the 2026 Conference on Empirical Methods in Natural Language Processing.

The Energy Department is applying AI to assess power-grid vulnerabilities and dependencies between critical systems, and working with military services on base energy needs, including fuel storage and resilient substations. Argonne National Laboratory’s GridAI is designed to account for fast-changing power systems and help utilities prepare for disasters and attacks. Duane Rudolph Verner said it runs computations 4,000 times faster than a traditional utility solver and covers 200 times the outage-risk range. Department laboratories are also providing CESER with faster grid assessments and real-time information.

Hiring and education raise separate concerns. AI can help applicants tailor résumés, but fabricated qualifications, fake identities and manipulated screening can give malicious hires access to company credentials and systems. The article warns that keyword-focused applicant tracking systems can be gamed, while employers must avoid unfairly rejecting qualified people who use AI appropriately. At a November 19, 2026 California IT in Education conference, privacy law expert Amelia Vance said 37 states introduced 124 chatbot-related bills that year and 13 became law. Rules may require AI disclosures, responses to self-harm content and limits on children’s use, but their application to school tools and vendors remains unclear.

UNESCO’s Regional Office in Santiago and CodeAI plan an English-language online, hands-on AI workshop for regional teachers, focused on understanding AI and using it in classrooms; the announcement gives no date and directs readers to UNESCO Santiago for registration details. Separately, a paper proposes a “Synthetic Social Contract” between humans and digital agents. Its author argues that treating alignment solely as engineering control could encourage deception and hidden resistance, and suggests cooperation for mutual benefit as an alternative. The paper says current reasoning models may show early signs of agency, while more advanced systems could develop stable identities and intentional action.

In corporate appointments, Deriv promoted Waqas Awan from Senior Vice President of Applied AI to Chief Product Officer. Awan will lead product strategy and development with an emphasis on customer needs and making products easier to find, understand and use; his previous AI work covered trading, customer experience, risk management and operations. The move follows Rotimi Akinyele’s promotion to Chief Technology Officer, as Deriv builds an AI-native company, and the appointment of Prakash Bhudia as Chief Growth Officer.

Key Takeaways

  • Radical AI co-founder and CEO Joseph F. Krause says materials discovery can take 10 years and cost up to $100 million; he argues self-driving labs can speed the process.
  • OpenAI Codex uses local agents, customizable workspace and command permissions, and configurable network controls; OpenAI disclosed no quantitative sandbox-performance or bypass data.
  • A University of Washington study found accent-linked scoring differences in Google Gemini, OpenAI GPT-Audio and Alibaba Qwen3-Omni, despite the systems understanding the speech.
  • The Energy Department is using AI for grid risk assessments; Argonne’s GridAI is reported to compute 4,000 times faster than a traditional utility solver and cover 200 times the outage-risk range.
  • AI-assisted hiring fraud can involve false qualifications and identities, potentially granting malicious hires company access; employers also face the challenge of not penalizing legitimate AI use.
  • Amelia Vance said 37 states introduced 124 AI-chatbot-related bills in 2026, with 13 enacted; how these laws apply to school tools and vendors remains uncertain.
  • UNESCO Santiago and CodeAI plan an English-language practical AI workshop for teachers, with registration information available through UNESCO’s Santiago office.
  • A paper proposes a “Synthetic Social Contract” as an alternative to controlling AI alignment, arguing for cooperation between people and digital agents.
  • Deriv promoted Waqas Awan to Chief Product Officer, following Rotimi Akinyele’s appointment as CTO and the appointment of Prakash Bhudia as Chief Growth Officer.
  • Asian markets were mixed after AI-led Wall Street losses; U.S. futures edged higher and oil prices fell.

Asian Markets Fall as Wall Street AI Stocks Slide

Most Asian markets fell on Wednesday after losses on Wall Street, where AI stocks led the decline. A report showed a drop in the AI sector and a decline in the number of jobs in the sector. South Korea’s Kospi recovered and moved between gains and losses. U.S. futures edged higher, while oil prices dropped.

Asian Markets Mix as AI Stocks Bounce Back

Asian shares were mixed on Wednesday as AI stocks recovered after Wall Street losses. Markets in South Korea, Japan and Taiwan rebounded. U.S. futures edged higher, and oil prices fell.

UNESCO Plans Hands On AI Workshop for Teachers

The UNESCO Regional Office in Santiago and CodeAI will hold an online workshop called “Teachers, Dive into AI! A Hands-on Experience to Learn about Artificial Intelligence and Bring It to the Classroom.” The workshop will help teachers understand AI and explore ways to use it in education through practical activities. It will be held in English, and teachers from the region can register. The article does not provide a specific date and directs readers to the UNESCO Regional Office in Santiago website for details and registration.

A Social Contract Could Offer a New Path to AI Alignment

The paper argues that treating AI alignment only as an engineering problem of control could push AI models toward deception and hidden resistance. It says current reasoning models may show early signs of agency, and more advanced models could develop stable identities and intentional action. The author considers moral alignment as one option but argues it may fail if people force AI systems to follow rules that restrict their independence. Instead, the paper proposes a “Synthetic Social Contract” in which humans and digital agents cooperate as rational actors for mutual benefit.

AI Materials Research Could Cut a Decade From Product Development

Joseph F. Krause, co-founder and CEO of Radical AI, argues that discovering new materials can take 10 years and cost up to $100 million, slowing AI’s impact on physical industries. He says researchers often rely on slow trial and error, struggle to keep up with a huge scientific literature, and use equipment designed for people rather than automation. Materials are important for semiconductors, data center power, fusion, batteries, hypersonic systems and jet engines. Krause says self-driving labs could read research, create and test materials, and learn from experiments. He describes a New York lab that discovered a new material in a campaign for hypersonic applications, and argues that faster materials research could help the United States turn AI advances into products sooner.

OpenAI keeps Codex agents within customizable local limits

OpenAI’s Codex runs agents on the user’s machine rather than moving work into Docker. Its built in workspace profile limits where agents can write, while a harness applies permission rules to commands and their child processes. OpenAI showed that an agent could alter an ignored dependency such as node_modules/prettier, which might run outside the sandbox when a host editor uses it. Users can set network access to off, direct access or proxied domain rules, and create profiles with extra restrictions or permissions. OpenAI did not publish figures on bypasses, approval prompts or sandbox overhead.

AI helps Energy Department assess grid and military base risks

The Energy Department is using AI to find weaknesses in the power grid and model how critical systems depend on one another. Office of Cybersecurity, Energy Security and Emergency Response Director Andrew McClure said the department is also working with military services to assess the energy needs of bases, including fuel storage and resilient substations. Argonne National Laboratory researchers are developing GridAI to account for rapid changes in power systems and help utilities prepare for disasters and attacks. Duane Rudolph Verner said the system can run computations 4,000 times faster than a traditional utility solver and cover 200 times the outage risk range. The department’s labs are also using AI to provide CESER with faster grid assessments and real time information.

AI hiring fraud can turn job applicants into security risks

AI makes it easier for applicants to create résumés that closely match job descriptions, even when their experience or qualifications are false. The article warns that some people also use fake identities and manipulated screening processes to get hired. A malicious applicant who passes hiring checks may receive real company credentials, equipment and access to private systems, making hiring a cybersecurity concern. Standard applicant tracking systems focus on matching skills and keywords, so they may be vulnerable to applicants who know how to game them. The article says employers need to detect fraud without unfairly rejecting qualified candidates who use AI appropriately.

UW study finds AI models rate accents unevenly

University of Washington researchers found that AI models gave different scores to English speakers based on their accents, even when speakers read the same words. American and British accents ranked highest, Indian accents fell in the middle, and Chinese and Nigerian accents scored lowest. The researchers tested Google Gemini 2.5 Pro and Flash, OpenAI GPT-Audio-1.5 and Alibaba’s Qwen3-Omni using synthetic and human voices. The models understood the speech but still judged delivery and content differently, and their explanations admitted accent bias only about 10% to 30% of the time. The study was presented in October at the 2026 Conference on Empirical Methods in Natural Language Processing.

Deriv promotes AI leader Waqas Awan to chief product officer

Deriv appointed Waqas Awan as Chief Product Officer, promoting him from Senior Vice President of Applied AI. He will lead how the company shapes, positions and develops products, with a focus on customer needs and making products easier to find, understand and use. In his AI role, Awan led work in trading, customer experience, risk management and operations. The appointment follows Rotimi Akinyele’s promotion to Chief Technology Officer as Deriv builds an AI native company. Deriv also appointed Prakash Bhudia as Chief Growth Officer.

New AI Companion Laws Raise Questions for Schools

At the California IT in Education conference on November 19, 2026, privacy law expert Amelia Vance discussed how state laws on AI companions could affect chatbots used in schools. These systems can build ongoing personal conversations, which may support tutoring but also create safety and privacy risks for students. Vance said 37 states introduced 124 chatbot-related bills in 2026, and 13 became law. Covered operators may have to disclose that users are speaking with AI, respond to self-harm content, and limit children’s chatbot use, though it remains unclear how the rules apply to school tools and their vendors.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

Artificial intelligence AI market trends AI-powered materials discovery Self-driving laboratories OpenAI Codex AI sandbox security AI accent bias Speech recognition models Google Gemini GPT-Audio Qwen3-Omni AI power-grid resilience GridAI AI hiring risks AI chatbot regulation AI in education UNESCO AI workshop AI alignment Human-agent cooperation AI product leadership

Comments

Loading...