AI is making waves across healthcare, but not all of it is positive. Researchers at UT Southwestern found that triage tools used in emergency rooms can pick up and even amplify human biases. When patient records contained stigmatizing language about mental health or substance use, some AI models lowered the priority of patients who later worsened. Nearly 10 percent of patients were incorrectly downgraded in certain cases, raising serious questions about how these tools are tested before deployment.
On the clinical side, a separate study offers a more hopeful picture. Dr. Adrián Mosquera Orgueira and his team built a machine learning model trained on data from over 5,000 myelofibrosis patients. The model proved better at identifying high-risk cases than older methods. Still, the researchers caution that real-world testing remains limited, and AI can perform well in studies without translating that success to actual patient care.
In the world of large language models, DeepSeek has updated its cost-efficient V4.1 Flash 1017 model. The model scored 45 on the Artificial Analysis Intelligence Index, outperforming GLM 5.3 and Kimi K3. It also improved on scientific reasoning and reduced hallucinations, while keeping pricing steady at $0.30 per million input tokens and $1.20 per million output tokens during peak hours.
Hardware developments are also shaping the AI space. Asus launched the Intel Arc Pro B70 workstation graphics card, built for AI inference tasks. The card packs 32 GB of ECC GDDR6 memory, 32 Xe cores, and 256 XMX AI engines, delivering 367 TOPS of INT8 performance. It supports up to four cards in a single workstation, making it easier to run large local models using frameworks like PyTorch, vLLM, and Ollama.
Synthetic polling is another area drawing scrutiny. Professor Jennifer Stromer-Galley warns that AI personas used in polling may not reflect real voter behavior. These personas follow fixed instructions and give consistent answers, while real voters change their minds. Although synthetic polling can generate thousands of responses from a standard sample of about 1,200 people, experts worry that hidden biases may drown out minority viewpoints.
Genie, a startup building personalized AI personality assistants, has raised $1 million in funding since launching in 2020. The company says its AI can understand emotions and adapt to individual users through natural language processing and machine learning. Dr. Joanna Bryson and other experts urge careful consideration of the ethical implications involved in creating AI companions tailored to specific people.
The White House has launched a 120-day review of AI risks, led by Director Jay Clayton under the Super Intelligence Force. The review will examine cybersecurity, biosecurity, and model control, with Defense Undersecretary Emil Michael handling reporting protocols. No new rules have been introduced yet, though existing government powers may be used.
In education, the University of Florida now offers over 200 AI courses spanning all 16 colleges. An AI certificate is available to undergraduates in every major, covering AI use, evaluation, and ethics. Director Hans van Oostrom says early exposure to AI learning is a critical step for students across all fields.
A new Illinois law taking effect October 1 restricts employers from using AI to screen job applicants, aiming to curb hiring bias. Yet research suggests AI tools can sometimes discriminate more against women and minorities than human decision-makers. The law also requires written explanations and employee training, but experts question whether these measures go far enough.
The University of Hong Kong and the Global Cybersecurity Alliance have opened a Centre for AI and Cybersecurity Research, led by Professor Ho Chen. The center will focus on secure AI, AI-based cyber defense, frontier technology risks, and global security standards, with plans to share research, benchmarks, and open-source tools.
Factory has partnered with Databricks to help enterprises modernize data and software systems. The collaboration connects Factory Droids with Databricks tools to streamline application and data migration. The companies are already assisting a large global enterprise in moving a complex platform handling petabytes of data across hundreds of thousands of tables.
Key Takeaways
- AI triage tools in emergency care can deprioritize high-risk patients when records contain stigmatizing language, with nearly 10 percent of patients incorrectly downgraded in some cases.
- A machine learning model trained on over 5,000 myelofibrosis patients improved high-risk identification compared with older methods, though real-world testing remains limited.
- DeepSeek V4.1 Flash 1017 scored 45 on the Artificial Analysis Intelligence Index, surpassing GLM 5.3 and Kimi K3, while maintaining pricing at $0.30 per million input tokens.
- Asus launched the Intel Arc Pro B70 workstation graphics card with 32 GB GDDR6 memory and 367 TOPS of INT8 performance, supporting up to four cards per workstation.
- AI polling personas may not reflect real voter opinion shifts because they follow fixed instructions and give consistent answers regardless of changing circumstances.
- Genie, a startup building personalized AI personality assistants, has raised $1 million in funding since 2020, though ethical concerns remain about AI companions.
- The White House has initiated a 120-day AI safety review covering cybersecurity, biosecurity, and model control, led by Director Jay Clayton.
- The University of Florida now offers over 200 AI courses across all 16 colleges, with an AI certificate open to undergraduates in every major.
- Illinois enacted a law on October 1 restricting AI in hiring, but research suggests AI tools can sometimes discriminate more than human decision-makers.
- The University of Hong Kong and Global Cybersecurity Alliance opened a Centre for AI and Cybersecurity Research focusing on secure AI and global security standards.
AI shows promise and limits in hematology care
Dr. Adrián Mosquera Orgueira and his team studied how AI can help in hematology. They built a machine learning model using data from over 5,000 myelofibrosis patients. The model improved identifying high-risk patients compared with older methods. However, AI still faces limits, including few real-world tests and the risk of performing well in studies but not in actual patient care.
AI triage tools can copy human biases in emergency care
Researchers at UT Southwestern studied how AI tools prioritize emergency patients. They analyzed records from Stanford Medicine and Beth Israel Deaconess Medical Center using three AI models. Stigmatizing language about patients sometimes caused the models to lower the priority of people who later worsened. Nearly 10 percent of patients were incorrectly downgraded in some cases.
Stigmatizing details lead AI to deprioritize high-risk patients
A study by UT Southwestern researchers found that AI triage tools can be swayed by how patient histories are written. The team used records from two major hospitals and tested three open-source AI models. Details about mental health, substance use, or frequent emergency visits sometimes made the AI deprioritize higher-risk patients. The study calls for testing each AI tool against real patient results before use.
DeepSeek V4.1 Flash 1017 model shows strong performance gains
DeepSeek updated its cost-efficient model, DeepSeek V4.1 Flash 1017, with notable improvements in reasoning and knowledge. Its score on the Artificial Analysis Intelligence Index rose to 45, ahead of GLM 5.3 and Kimi K3. The model improved on scientific reasoning and reduced hallucinations. Pricing stayed the same at $0.30 per million input tokens and $1.20 per million output tokens during peak hours.
Asus launches Intel Arc Pro B70 workstation graphics card
Asus released the UGen Intel Arc Pro B70 32 GB GDDR6 graphics card for AI inference and workstation tasks. The card uses Intel Xe2-HPG architecture with 32 Xe cores and 256 XMX AI engines, delivering 367 TOPS of INT8 performance. It has 32 GB of ECC GDDR6 memory and supports up to four cards in one workstation for running large local models. The card works with open frameworks including PyTorch, vLLM, and Ollama.
AI polling personas may not reflect real voter opinion shifts
AI personas used in synthetic polling may not change their answers like real voters do. Professor Jennifer Stromer-Galley says AI personas built on fixed instructions give consistent answers every time. Real voters change their minds when circumstances change. A good survey should include about 1,200 people, but synthetic polling can generate thousands of personas. The bigger concern is bias in AI models that may hide the views of real outliers.
Genie startup creates personalized AI personality assistant
Genie is a new startup developing an AI personality tailored to each individual. Creators Candace and Parker say the AI can understand emotions and offer advice. The technology uses natural language processing and machine learning to adapt to users. The team has raised $1 million in funding since starting in 2020. Experts like Dr. Joanna Bryson urge careful thought about the ethics of AI personalities.
White House launches 120-day AI safety review
The White House AI task force started a 120-day review of AI risks. Director Jay Clayton leads the Super Intelligence Force, which met on Oct. 15. The review will cover cybersecurity, biosecurity and model control. Defense Undersecretary Emil Michael will work on reporting protocols. The task force has not created new rules yet but may use existing government powers.
University of Florida expands AI education across all majors
The University of Florida has been teaching AI skills to students since 2020. UF now offers over 200 AI courses across all 16 colleges. A university-wide AI certificate is open to undergraduates in every major. The program focuses on understanding AI, using it, evaluating it and AI ethics. Director Hans van Oostrom says exposing students to AI learning is a key first step.
Illinois AI hiring law may not solve bias problem
A new Illinois law that took effect Oct. 1 stops employers from using AI to screen job applicants. The law aims to prevent bias in hiring. However, research suggests AI hiring tools can be just as biased as humans. A study found AI tools may discriminate more against women and minorities. The law also requires written explanations and training for employees. Experts say these rules may not be enough to stop bias.
HKU and Global Cybersecurity Alliance open AI security research centre
The University of Hong Kong and the Global Cybersecurity Alliance created a new Centre for AI and Cybersecurity Research. Professor Ho Chen will lead the centre as Acting Director. The centre will study secure AI, AI-based cyber defence, frontier technology risks and global AI security standards. It plans to share research findings, security benchmarks and open-source tools.
Factory partners with Databricks to power data-driven software
Factory announced a partnership with Databricks to help customers build or migrate software using the Databricks Data and AI Platform. The companies are helping a large global enterprise migrate a complex data platform with petabytes of data and hundreds of thousands of tables. Factory Droids and Databricks tools will connect application and data migration into one workflow. The partnership aims to make large-scale data and software modernisation less risky and more coordinated.
Sources
- AI in Modern Hematology: Insights from Clinical Applications in Myelofibrosis
- Artificial intelligence-enabled triage tools may reflect human biases
- Stigmatizing patient details can push AI to deprioritize high-risk emergency cases
- DeepSeek v4.1 Flash 1017 Places Ahead Of GLM 5.3, Kimi K3 On Artificial Analysis Intelligence Index
- Asus expands workstation AI hardware options with Intel
- Synthetic public opinion polling: AI personas don't always change their minds like real voters do
- This startup wants to give everyone an AI personality
- White House Begins 120-Day AI Safety Review Amid Growing Security Concerns
- From education to innovation: How UF is preparing students to thrive with AI
- Opinion: Illinois’ new AI hiring law has a human problem
- HKU, Global Cybersecurity Alliance Launch AI Security Centre
- Factory and Databricks: Building Software Powered by Data
Comments
Please log in to post a comment.