The Top AI Stories of 2025: AI Coding, AGI, and More

Artificial intelligence in 2025 was less about flashy demos and more about hard questions. What actually works? What breaks in unexpected ways? And what are the environmental and economic costs of scaling these systems further?

It was a year in which generative AI slipped from novelty into routine use. Many people got accustomed to using AI tools on the job, getting their answers from AI search, and confiding in chatbots, for better or for worse. It was a year in which the tech giants hyped up their AI agents, and the general public seemed generally uninterested in using them. AI slop also became impossible to ignore—it was even Merriam-Webster’s word of the year.

Throughout it all, IEEE Spectrum’s AI coverage focused on separating signal from noise. Here are the stories that best captured where the field stands now.

Alamy

AI coding assistants have moved from novelty to everyday infrastructure—but not all tools are equally capable or trustworthy. This practical guide by Spectrum contributing editor Matthew S. Smith evaluates today’s leading AI coding systems, examining where they meaningfully boost productivity and where they still fall short. The result is a clear-eyed look at which tools are worth adopting now, and which remain better suited to experimentation.

Amanda Andrade-Rhoades/The Washington Post/Getty Images

As AI’s energy demands raise concerns, water use has emerged as a quieter but equally pressing issue. This article explains how data centers consume water for cooling, why the impacts vary dramatically by region, and what engineers and policymakers can do to reduce the strain. Written by the AI sustainability scholar Shaolei Ren and Microsoft sustainability lead Amy Luers, the article grounds a noisy public debate in data, context, and engineering reality.

iStock

When AI systems fail, they don’t fail like people do. This essay, by legendary cybersecurity guru Bruce Schneier and his frequent collaborator Nathan E. Sanders, explores how machine errors differ in structure, scale, and predictability from human mistakes. Understanding these differences, the researchers argue, is essential for building AI systems that can be responsibly deployed in the real world.

Christie Hemm Klok

In this insider account, John Dean, the cofounder and CEO of WindBorne Systems, tells readers how his team built one of the most technically ambitious AI forecasting systems to date. The company’s approach combines autonomous, long-duration weather balloons that surf the wind with a proprietary AI model called WeatherMesh, which both sends the balloons high-level instructions on where to go next and analyzes the atmospheric data they collect.

WindBorne’s platform can produce high-resolution predictions faster, using far less compute, and with greater accuracy than conventional physics-based methods. In the article, Dean walks readers through the…