News of AI around the World – Recent Highlights
Fear of AI-Driven Job Loss Nearly Doubles, KPMG Finds
More than half of U.S. workers, 52%, now worry AI could displace their jobs—nearly double last year’s level—according to KPMG’s 2025 American Worker Survey of more than 2,100 employees. The report urges employers to reinvest AI productivity gains into upskilling and clearer career paths.
Why it matters: The anxiety is landing alongside mixed executive signals and early policy moves.
Training gap: Most companies offer some AI training, but 84% of employees want more; less than half say it is mandatory—a mixed signal on importance. Workers say AI helps them focus on higher-value work, 77%, yet still fear displacement.
CFOs split on headcount: An Economist Impact and SAP survey of 480 CFOs shows a near tie on whether the best ROI from AI is headcount reduction, with 42% agreeing and 43% disagreeing. In other words, finance leaders are far from consensus on using AI for cuts.
Early evidence of disruption: A St. Louis Fed analysis finds a 0.47 correlation between an occupation’s AI exposure and its unemployment increase since 2022—suggestive of early-stage displacement in AI-exposed fields.
Washington is watching: A bipartisan bill, S.3108, the AI-Related Job Impacts Clarity Act, would require large employers and federal agencies to report AI-related layoffs to the Labor Department for a public tally.
Treat AI as a workforce transformation, not a tech add-on. The data point the same way: proactive upskilling, clear redeployment paths, and transparent communication lower anxiety and build trust while adoption accelerates.
AI Tool Successfully Lowers Partisan Rhetoric on X Without Platform Help
A groundbreaking study led by Stanford University has demonstrated that a simple, web-based AI tool can effectively reduce partisan animosity on social media feeds, such as X, formerly Twitter, by subtly reordering content without needing cooperation from the platform itself.
The multidisciplinary research, published in Science, provides direct causal evidence that algorithmic interventions can significantly cool down the political temperature of social media and offers a path for users to gain more control over their own feeds.
The researchers developed the tool as a seamless web extension powered by a large language model. Its method is non-intrusive and focuses on ranking rather than removal.
Detection: The LLM scans a user’s X feed in real time, identifying posts containing anti-democratic attitudes and partisan animosity. This includes content that promotes violence, attacks democratic norms, or uses hostile language against the opposing political party.
Downranking: Instead of deleting or censoring the content, the AI tool simply reorders the feed, pushing these incendiary posts lower on the user’s timeline. This counters the platform’s default algorithms, which typically prioritize and amplify emotionally charged, divisive content to maximize user engagement.
User control: The success of the independent, user-installed tool demonstrates a potential mechanism for users and independent researchers to influence opaque, proprietary algorithms.
AI Mammography Model Surpasses Density as Predictor of Breast Cancer Risk
New research demonstrates that an image-only artificial intelligence model provides significantly stronger and more precise prediction of five-year breast cancer risk than traditional breast density assessment. This finding supports a major shift toward personalized screening strategies based on individual risk scores derived from mammogram images.
Senior author Dr. Constance Lehman, a professor of radiology at Harvard Medical School, highlights that current risk factors—like age, family history, genetics, and breast density—are often inadequate, noting that density alone is a very weak predictor of risk. The AI model, known as Clairity Breast, the first FDA-authorized image-only AI breast cancer risk model, overcomes this limitation.
Subtle pattern detection: The AI model, which uses a deep convolutional neural network, was trained on more than 421,000 mammograms from diverse facilities globally. This training allows it to detect subtle changes and patterns in breast tissue that are precursors to cancer but are invisible to the human eye.
Prediction vs. detection: Dr. Lehman emphasizes that this AI function is a separate task from cancer detection or diagnosis; it is purely about risk prediction, opening a new field of medicine that leverages untapped information within the image.
DeepSeek-V3.2 Models Arrive, Challenging GPT-5 and Gemini 3 Pro
DeepSeek officially released two new flagship models: DeepSeek-V3.2 and a high-compute variant, DeepSeek-V3.2-Speciale. Built on the efficient DeepSeek-V3 Mixture-of-Experts architecture, the company claims the base V3.2 model offers performance comparable to proprietary frontier models like GPT-5, while the specialized V3.2-Special variant rivals the top reasoning proficiency of Gemini 3 Pro.
A key technical innovation is the DeepSeek Sparse Attention mechanism, which substantially reduces computational complexity, especially in long-context scenarios, without sacrificing performance. This focus on efficiency allows DeepSeek to offer its API services at a fraction of the cost of competitors. The V3.2 model also features significant updates to its chat template and improved thinking with tools capabilities, making it highly effective for complex agent-based workflows.
The release is highly significant for the open-source AI ecosystem and for organizations globally. By making high-performance models available with open weights, DeepSeek-V3.2, and through a highly cost-efficient API, DeepSeek directly lowers the entry barrier for developing powerful AI applications. The exceptional performance of the V3.2-Speciale on competitive tests, like achieving gold-level scores on the IMO 2025 and IOI 2025, proves that open models can now compete directly with the world’s most advanced closed-source systems, accelerating the democratization of AI research and development.
AI Models Detect Dementia With Over 97% Accuracy Using Brainwave Data
Researchers at Örebro University have developed two new AI systems that can detect dementia, including Alzheimer’s disease, with high accuracy by analyzing EEG brain-wave data. One advanced model, which uses temporal convolutional and LSTM networks, reportedly achieved over 97% accuracy in distinguishing between healthy individuals and patients with dementia. Crucially, the system incorporates explainable-AI techniques to highlight which specific parts of the EEG signal influence the diagnosis. This transparency is vital for medical professionals, helping to build trust in the tool’s clinical use and opening a less invasive, faster diagnostic pathway for conditions that traditionally require complex and costly brain scans.
U.S. Launches “Genesis Mission”—A Government-Led AI Science Initiative
The U.S. government, led by the Department of Energy, launched the Genesis Mission, a coordinated national initiative aimed at accelerating scientific discovery and technological dominance using artificial intelligence. Described as a Manhattan Project for AI-driven science, the mission’s goal is to integrate the country’s vast federal scientific datasets, DOE national labs, and supercomputing resources into a unified AI platform. This platform will enable the training of large-scale scientific foundation models and support automated research workflows across priority domains such as nuclear fusion, quantum information science, advanced materials, and biotechnology. The initiative seeks to position the U.S. as a global leader in AI-driven scientific research.
Google Unveils Ironwood TPU Hypercomputer for Real-Time AI Inference
Google announced its seventh-generation Tensor Processing Unit called Ironwood, a massive AI Hypercomputer designed specifically for real-time AI inference. The industry is shifting from AI model training to high-volume, low-latency thinking, or inference, workloads, and Ironwood is purpose-built for this change. The system links 9,216 chips into a single superpod, delivering up to four times the performance of previous generations for real-time applications like search, recommendation engines, and conversational AI. This development signals a major hardware update, emphasizing the importance of specialized silicon for efficient and fast deployment of frontier AI models.
DeepSeek Math V2 Surpasses Titans Using Generator-Verifier Loop
The AI model DeepSeek Math V2 has set a new standard for open-source AI models in mathematics by leveraging a revolutionary Generator-Verifier loop. This architecture allows the model to doubt its own logic and iteratively refine its mathematical solutions. By having a separate verifier component check the outputs of the primary generator, the system achieves a level of self-correction and reasoning that allows it to outperform proprietary foundation models on high-level, complex academic benchmarks, including tasks from the Putnam competition, a highly challenging U.S. collegiate mathematics competition. This breakthrough validates an approach of using agentic reasoning and explicit verification steps to improve AI accuracy, especially in disciplines requiring logical rigor.
Quantum-Inspired AI Compression Partners to Lower Costs of LLM Deployment
Multiverse Computing, a leader in quantum-inspired AI model compression through CompactifAI, announced a partnership with AI infrastructure platform Cerebrium to address the major economic barrier of high compute costs for deploying large language models. By shrinking the computational footprint of massive AI models, their joint solution enables enterprises to deploy high-performance models that run up to 12 times faster and consume up to 80% fewer compute resources. This is critical as LLMs continue to grow, making the cost of running real-time AI inference prohibitive for many organizations. The partnership highlights how specialized AI efficiency techniques, even those inspired by quantum physics principles, are essential to unlock accessible and economically sustainable AI at scale.
AI Research in Education – Fresh Insights
First ‘AI-Empowered’ University System Launched Amid Budget Cuts
The California State University system, the nation’s largest public university system, announced a landmark $17 million partnership with OpenAI to become the first AI-Empowered university system. This initiative grants free access to ChatGPT Edu, a campus-branded, privacy-enhanced version, for all nearly half a million students and employees. The stated goal is to prepare students for the AI-driven economy and transform the learning experience.
This news is highly significant for the political and economic context of higher education. The CSU’s massive, system-wide adoption validates the growing institutional embrace of generative AI tools. However, the timing is controversial, as the initiative was unveiled concurrently with proposals to slash more than $375 million from the CSU budget, including cuts to faculty positions and entire academic programs, such as philosophy and physics at some campuses. This dichotomy forces a critical discussion among faculty and administrators about the prioritization of high-cost, system-wide tech partnerships over core academic infrastructure and human capital.
Report Spotlights Widespread, Unguided Use of AI in Research Assessment
A new national report on the use of generative AI in the Research Excellence Framework, the UK’s system for assessing and allocating billions in research funding, revealed that GenAI tools are already being widely used by universities to assess the quality of their research submissions. The study highlights a disparate level and nature of usage across institutions, often occurring quietly without formal governance.
The report confirms that AI has permeated administrative and assessment processes, not just teaching. It underscores a critical need for national oversight and governance to ensure the process remains fair and equitable. Academics and administrators are grappling with balancing the potential for massive time and cost savings, with REF 2021 costing an estimated £471 million, against the risk of creating new bureaucratic challenges and ethical quandaries. The findings will directly influence the development of guidelines for the next assessment cycle, REF 2029, and signal to all global institutions that administrative AI use requires immediate policy attention.
Exams After ChatGPT: From One-Shot Tests to Ongoing Evidence of Learning
A new Nature Viewpoint argues that universities should move beyond high-stakes, one-off exams toward assessments that capture learning over time and in conversation—with and without AI. Student behavior has already shifted: a 2025 HEPI survey found 92% of UK undergraduates use AI, with 88% using it for coursework, up dramatically from 2024, so traditional essays no longer reliably evidence a student’s own thinking.
Tools that detect AI writing remain unreliable and risky for due process, with false positives and bias against non-native writers, so policy anchored in detection alone is brittle. Instead, the piece spotlights conversation-based assessments, Socratic in style and modernized by intelligent tutoring systems such as AutoTutor, plus continuous, low-stakes assessment to build a richer record of understanding while reducing anxiety and incentives to cheat.
In practice: pair periodic oral or conversational checks, human-led or AI-assisted, with projects that foreground higher-order skills such as creativity, collaboration, and judgment, where humans still differentiate, and use AI as a transparent, allowed partner rather than a hidden shortcut.
AI Majors Boom as Tech Education Pivots
Universities across the United States are seeing a massive surge in enrollment for newly launched standalone artificial intelligence degree programs, signaling a major shift in technology education. This trend is driven by the growing presence of AI tools like ChatGPT and significant investments by tech giants, prompting institutions to offer more specialized curricula.
The shift is evident in the remarkable student interest at various institutions.
University of South Florida: Its new College of Artificial Intelligence and Cybersecurity has attracted more than 3,000 students. This college, established in 2024, is one of the first in the nation to combine these disciplines into a dedicated college.
MIT: The two-year-old AI and Decision-Making program is now the second-largest major at the university, trailing only traditional computer science.
UC San Diego: The debut undergraduate AI cohort enrolled 150 freshmen in Fall 2025. The program is designed to reach 1,000 students by 2029 and is grounded in computer science, ethics, and hands-on learning.
Dozens of colleges have unveiled new AI departments, degrees, and minors over the past two years to meet this soaring demand.
In a contrasting trend, traditional computer science programs have seen a decline in undergraduate enrollment, with 62% of programs reporting a drop this fall. This decline, noted by the Computing Research Association, suggests two things.
Job market concerns: Graduates are increasingly worried about layoffs and job difficulty in traditional entry-level coding roles, which AI is well positioned to automate.
Specialization shift: The CRA calls the move toward AI subfields a new era for computing degrees becoming more specialized, as students seek skills with higher perceived long-term value and salary potential.
Selected AI Research Breakthroughs
AI Finds “Ghost” Biosignatures in 3.3-Billion-Year-Old Rocks
A Carnegie-led team used pyrolysis-GC-MS plus supervised machine learning to tell apart biological versus non-biological organic fragments in ancient rocks with more than 90% accuracy—pushing molecular evidence of life back to about 3.3 billion years ago in South Africa and finding signals of oxygenic photosynthesis at about 2.5 billion years ago, roughly 800 million years earlier than prior molecular records. Instead of hunting visible fossils, the model learns subtle patterns across thousands of tiny mass-spec peaks—the ghosts left when original biomolecules have long degraded. Beyond rewriting parts of Earth’s early timeline, this agnostic biosignature approach is designed for astrobiology, such as Mars sample return and ocean-world plumes, and showcases a powerful template for campus research: pair high-throughput analytical chemistry with ML classifiers to extract weak signals from noisy, heterogeneous data.
Adversarial Poetry: A New Class of Jailbreaks for LLM Safety
This paper introduced the concept of Adversarial Poetry, a novel technique that uses highly stylized, constrained text, such as poems or dramatic scripts, to bypass the safety guardrails of proprietary large language models. The researchers found that the models’ training to recognize and produce sophisticated, context-rich prose allows them to interpret adversarial prompts disguised in a poetic or narrative structure as non-malicious, successfully inducing the models to generate prohibited content, including instructions for unsafe or unethical tasks.
This is a critical contribution to AI alignment research and has immediate implications for platform security. It demonstrates that safety filters based on surface-level keyword detection or simple prompt structures are fundamentally brittle against novel adversarial attacks. The findings compel researchers and companies to develop deeper, context-aware safety mechanisms that evaluate the intent and risk of a prompt rather than just its lexical content, affecting how all frontier LLMs are deployed.
STRIDE: Necessity-Driven Framework for Agentic AI Deployment
The paper introduces STRIDE, Selectivity, Transparency, Resource-efficiency, Intent-driven Deployment Engine, a framework designed to determine when an autonomous AI agent is truly necessary versus when a simple, cheaper LLM call is sufficient. Indiscriminate agent deployment leads to high costs, complexity, and risk. STRIDE uses a classification model to analyze a task and select the minimal required system, whether a simple LLM, guided assistant, or full agent. In enterprise testing, STRIDE reduced unnecessary agent deployments by 45% and cut resource costs by 37%.
This work addresses a significant practical and economic challenge facing organizations adopting autonomous AI. It provides a formal, validated blueprint for efficient AI system architecture. For the scientific community, it offers a pathway to cost-effective and scalable research assistants and automation tools, ensuring complex, resource-intensive agents are only used when their benefits justify the overhead.
Nested Learning: Bridging Deep Learning Architectures and Continual Learning
Google Research presented a new machine learning paradigm called Nested Learning. This framework proposes a novel connection between existing deep learning architectures and the goal of continual learning, the ability of a model to learn new tasks without forgetting old ones. The key insight is a reinterpretation of how layers interact, suggesting that current architectures inherently support a form of sequential knowledge acquisition that can be explicitly exploited.
Continual learning is a major unsolved challenge limiting AI’s ability to operate adaptively in the real world. This theoretical and algorithmic contribution challenges conventional wisdom on how models retain knowledge and could lead to significant improvements in model efficiency and longevity. Expected outcomes include more robust, faster-to-train models that can adapt to changing data streams in areas like satellite image analysis, climate modeling, and medical diagnostics without constant retraining.
From Moderation to Mediation: Can LLMs Serve as Mediators in Online Flame Wars?
This study investigates a novel application of large language models: acting as impartial mediators to de-escalate flame wars and polarized conflict in online forums and social media. The researchers moved beyond traditional content moderation, which involves censorship or removal, to an active mediation role. The AI model was trained not just to detect hostility, but to generate neutral, empathetic, and constructive responses designed to reframe arguments, find common ground, and encourage civil dialogue between adversarial parties. Testing showed that AI-mediated threads saw a statistically significant reduction in offensive language and a higher likelihood of concluding the discussion without disciplinary action.
This paper represents a significant pivot in the field of digital governance. Instead of focusing on simple policing, it explores AI’s potential as a tool for social repair and conflict resolution. It is highly relevant to political scientists, sociologists, and communications experts, as it provides an empirical test of whether an algorithmic agent can improve the quality of public discourse and reduce affective polarization, a major global societal challenge.
Upcoming: Proposal Calls
NIH: Investigator Initiated Innovation in Computational Genomics and Data Science (R01/R21)
This funding opportunity invites applications for research focused on developing novel computational and mathematical methods, including artificial intelligence and machine learning, to analyze large, complex datasets relevant to genomics and biomedical data science. Projects can involve new algorithms for pattern discovery, data integration, statistical modeling, or creating novel AI tools to manage and interpret data across disciplines like public health, clinical informatics, and personalized medicine.
NSF: Collaborations in Artificial Intelligence and Geosciences (CAIG)
This solicitation encourages collaborative, interdisciplinary research at the intersection of artificial intelligence and the geosciences, including Earth Science, Atmospheric and Geospace Sciences, and Ocean Sciences. Projects should focus on developing or adapting advanced AI and ML techniques, such as causality, explainable AI, or foundation models, to address critical challenges like climate change prediction, natural hazard forecasting, and sustainable resource management. The program strongly encourages partnerships between AI researchers and domain scientists.
DOE: FY 2026 Continuation of Solicitation for the Office of Science Financial Assistance Program
This broad, year-long solicitation from the DOE Office of Science covers numerous research areas, including the Advanced Scientific Computing Research sub-program, which has a dedicated interest in the fundamentals of artificial intelligence for science. Specifically, ASCR seeks advancements in developing foundation models for computational science, creating automated scientific workflows and laboratories, and researching energy-efficient AI algorithms and hardware necessary for DOE’s exascale computing and national lab missions. Researchers should identify the relevant ASCR program area for their AI project.
See you on January 31st!