AIhub

As more news content is generated by AI, language becomes homogenized and less innovative. Research shows that AI training on synthetic text leads to 'model collapse', reducing linguistic diversity. This not only affects journalism but also impairs society's ability to describe and discuss reality.

This month's digest covers foundation models for conservation, AI for resource allocation, color metaphors and LLMs, and cutting-edge robots from ICRA.

The AAAI Future of AI Research report, published March 2025, covers 17 AI topics. The fifth panel discussion focuses on AI agents, exploring the evolution from rule-based to generative AI multi-agent systems, and challenges in alignment and governance.

In this interview, AAAI Fellow Tanya Berger-Wolf discusses her pioneering work at the intersection of AI and ecology, including the development of the BioCLIP foundation model for the Tree of Life, its applications in biodiversity monitoring and conservation, and future directions for AI in science.

We sat down with Douglas Guillbault to discuss his paper, “Comparing Colorseeing, Colorblind, Painters, and Large Language Models in Their Processing of Color Metaphors”. The results have interesting implications for how we model human cognition, and in turn, how the concept of synaesthesia could be integrated to develop more intelligent AI models. What are color metaphors? A color metaphor is the use of color to describe something in a way that is not immediately literal. For example, to say “green with envy” would be a color metaphor, because envy doesn’t have an immediate visual structure to it – we’re evoking a broader, more flexible notion of what green conveys, beyond just its visible properties.

In this episode of The Good Robot, Eleanor Drage speaks with Tara Merk about how community-owned data centers could transform digital ownership and challenge Big Tech dominance. They explore alternative models prioritizing local empowerment, sustainability, and cooperative governance, drawing on examples from Germany's renewable energy sector. The conversation reimagines the internet as a shared public resource designed for collective needs rather than profit.

The AAMAS 2026 best paper awards were presented at the 25th International Conference on Autonomous Agents and Multiagent Systems, held from 25-29 May 2025 in Paphos, Cyprus. Winners in three categories were announced: Best Paper, Best Student Paper, and Blue Sky Ideas Paper.

Sanmay Das, professor of computer science at Virginia Tech, was elected a 2026 AAAI Fellow for his work on multiagent systems and AI for social impact. He discusses his research on market making, matching, and societal resource allocation, and how AI can be integrated into public systems.

A study by Oregon State University finds that adding design friction to AI systems, such as prompts to consider energy consumption, can encourage more responsible use. Action-based friction that requires users to search for existing resources was effective, while cue-based messaging only increased trust. With AI's energy use rising, such interventions are crucial.

OpenAI's AI model found a counterexample to Paul Erdős's 1946 planar unit distance problem, showing that grid-like arrangements are not optimal. The result was autonomously produced, and later improved by human mathematician Will Sawin. This marks a growing role for AI in mathematical research.

This post contains a list of the AI-related seminars that are scheduled to take place between 1 June and 31 July 2026. All events detailed here are free and open for anyone to attend virtually.

Image Empire is an animated fairytale about the fusion of the real and the virtual within contemporary AI models. The film forms part of a research project undertaken by Alan Warburton which also includes a research paper and a series of satellite events.

This month's AIhub digest covers AI for Science conference, lottery ticket hypothesis interview, world models discussion, transparent and trustworthy AI research, foundation model impacts report, AIES conference reflections, Robotics Café, ACL desk rejection policy, arXiv anti-AI slop policy, and more.

A study by researchers at the University of Michigan suggests AI chatbots can easily engage in covert advertising to manipulate users, and many people don't realize it. As major tech companies experiment with chatbot ads, this raises concerns about user privacy and autonomy.

This episode of The Good Robot explores how feminist principles and decentralized infrastructure could transform cloud infrastructure from a corporate service into a public commons. Friederike von Franqué, policy advisor at Wikimedia Germany, discusses examples from Frankfurt's energy-intensive data centres to Stockholm's municipally owned fibre network, advocating for environmental accountability and community-driven design.

In this month's AIhub coffee corner, AI experts discuss the concept of world models, their definitions, applications, and limitations. The conversation covers topics such as transition models in reinforcement learning, video generation, causal models, and challenges in robotics and simulation.

Anthropic's Mythos model has discovered thousands of severe security vulnerabilities, including many zero-day flaws that have gone undetected for decades. Banks worldwide fear cybercriminals will exploit this AI to rob them. Anthropic has granted access to a defensive coalition including Microsoft, but not to banks in Australia, the UK, or Europe.

Jonathan Frankle discusses the lottery ticket hypothesis, for which he won the 2023 AAAI/ACM Doctoral Dissertation Award. He covers empiricism vs. theoretical proofs, the shift in computer science methodology, the pressure on young researchers to deliver impact, and his current focus on evaluating how well AI systems work in practice.

MIT researchers developed EnergAIzer, a rapid prediction tool that estimates AI workload power consumption on specific processors in seconds with about 8% error, aiding data center energy efficiency.

System failures cost over a trillion dollars annually. Engineers rely on time series data to troubleshoot outages quickly. ARFBench is a new benchmark derived from real incidents at Datadog, designed to evaluate AI models on time series question-answering (TSQA) tasks. Experiments show existing models have significant room for improvement, hybrid TSFM-VLM models show promise, and human-AI complementarity achieves superhuman performance.

Federated unlearning allows removal of user data from trained AI models to enhance privacy, but research reveals it introduces new security vulnerabilities, including potential for attackers to inject hidden backdoors and then request unlearning to cover tracks. Current methods lack verification, risking system integrity.

A reflection on AIES 2025, highlighting discussions on LLMs in clinical settings and human rights, along with key presentations and new conference formats.

Scientists at NTU Singapore have developed a biochip that uses deep learning and computer vision to detect microRNA biomarkers in 20 minutes, offering a faster and more accurate alternative to PCR for disease diagnosis.

A new study published in BMJ Open finds that five popular AI chatbots provide problematic health information in about half of responses, with only two out of 250 questions refused. Grok performed worst, and open-ended questions were especially risky. The study underscores the need for verification and not treating chatbots as medical authorities.

GRASP is a new gradient-based planner for learned dynamics (a “world model”) that makes long-horizon planning practical by (1) lifting the trajectory into virtual states so optimization is parallel across time, (2) adding stochasticity directly to the state iterates for exploration, and (3) reshaping gradients so actions get clean signals while we avoid brittle “state-input” gradients through high-dimensional vision models.

As AI tools become widespread, people are tempted to offload difficult thinking tasks. Cognitive science warns that overdependence can erode critical thinking. The key is to use AI mindfully, maintaining control over what we offload and engaging in reflective practices to balance external support with our own cognitive growth.

In the latest interview from our AAAI/SIGAI Doctoral Consortium series, we speak with Ximing Wen, who researches transparent and trustworthy AI systems. She discusses prototype-based interpretable models, spatial grounding for document QA, and her vision for safe and reliable AI.

Partnership on AI has released a progress report on post-deployment governance practices for foundation models. The report evaluates 13 providers across four practices: sharing usage information, enabling research on societal impacts, reporting incidents and policy violations, and sharing user feedback. Key findings reveal that while leading organizations are defining information-sharing standards, adoption is slow among others, and public impact data remains fragmented.

This post contains a list of the AI-related seminars that are scheduled to take place between 5 May and 30 June 2026. All events detailed here are free and open for anyone to attend virtually.

A day-long conference at the Royal Society explored how AI is transforming scientific discovery across fields like astronomy, materials science, climate forecasting, and nuclear fusion. Speakers highlighted breakthrough applications including foundation models for galaxy classification, diffusion models for crystal design, AI-based sea ice forecasting, and fast plasma simulations for fusion.