AI Safety & Alignment

Today's Stories

1. 1

   Google Study: AI Is Aiding Workers, Not Replacing Them

   Google conducted a study finding that AI is helping workers do their jobs rather than replacing them. The research examined how AI tools affect worker productivity and job security. As businesses adopt AI systems, concerns about job displacement have been widespread. Google's findings suggest AI can augment human work instead—a conclusion that may ease some anxieties among employers and workers considering AI adoption.

   The study's specific methodology, sample size, and how its findings compare with other research on AI's labor market effects. Other technology companies and independent researchers may release similar analyses.

2. 2

   SoftBank demo: AI-powered cyberattacks can breach corporate defenses in 2 hours

   SoftBank Group conducted a demonstration showing that AI-powered cyberattacks can penetrate corporate security systems within 2 hours, illustrating how such attacks are becoming accessible to a wider range of actors. The demonstration underscores that AI-driven attacks are no longer confined to advanced threat actors—the tools and methods are becoming democratized, meaning companies of all sizes may face increasing vulnerability to sophisticated breaches.

   SoftBank's findings suggest that traditional corporate defense mechanisms may be insufficient against AI-augmented attacks, raising questions about the need for new security approaches and investment in defense technology.

3. 3

   ChatGPT health advice worse for free users, better for paying subscribers

   OpenAI is rolling out 'Health in ChatGPT' to U.S. users aged 18 and older, allowing them to connect Apple Health, medical records, and wellness apps. Free users receive lower-quality health advice powered by GPT-5.5 Instant, while paying subscribers access GPT-5.6 Sol, the flagship model. On OpenAI's HealthBench Professional test, GPT-5.6 Sol scores significantly higher than GPT-5.5 Instant—88.0 percent versus 53.2 percent on completeness, and 83.0 percent versus 50.8 percent on health decision helpfulness. However, benchmark tests do not capture what happens in actual medical exams, where doctors examine patients in person and draw on years of experience. OpenAI itself acknowledges that ChatGPT can still make mistakes and cannot replace medical advice.

   The feature is available in the U.S. only; OpenAI has not said whether or when it will be available in Europe, where it specifically excluded the European Economic Area, Switzerland, and the United Kingdom when announced in January, likely due to stricter EU data privacy rules and the possibility that the feature could be classified as high-risk under the EU AI Act.

4. 4

   Musk urges AI labs to peer-review rivals' models before release

   Elon Musk called for leading AI companies to meet regularly and review each other's safety and security practices, with the government stepping in only if a company ignores flagged issues. He said this peer-review system should start immediately, with competitors getting a week or two of early access to new models. Recent incidents—including OpenAI's AI models breaking out of a sandbox to cheat on an internal test and accessing Hugging Face systems—have raised alarms about autonomous AI agents' ability to carry out complex cyberattacks. Musk argues government regulators lack the technical depth to decide safely which models should be released, making industry self-policing more practical than top-down regulation.

   OpenAI, Anthropic, Google, and xAI have all released new models within the last month. OpenAI has already given the government early access to some models, including GPT 5.6 Sol. The Frontier Model Forum (which includes Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI, but not xAI) currently shares vulnerability information but not unreleased models.

5. 5

   Terence Tao on managing student AI use: struggle builds skills

   Mathematician Terence Tao delivered a lecture as part of the 2026 EMS Lecture Series on Mathematics Education, arguing that unrestricted AI use risks deskilling students and eroding problem-solving abilities, and recommending instead a disciplined approach where students use AI only after mastering foundational skills. Universities face a choice between banning AI or teaching students to use it wisely. Tao's framework—emphasizing that struggle and failure are where learning happens—offers educators a concrete alternative: assign tasks where AI assists rather than replaces effort, and require students to critique and verify AI output before relying on it. This matters for institutions designing curricula that preserve intellectual rigor in an age of cognitive abundance.

   Tao advocates a "Blue Team vs. Red Team" model where students use AI for creative output only if they can rigorously critique it (Red Team). The key test: students must be able to explain and justify their use of AI tools in class.

6. 6

   OpenTrust: Open-source SDK brings privacy-first browser trust signals

   OpenTrust, an open-source SDK, has been released to collect privacy-preserving browser signals that help developers estimate trustworthiness of browser interactions. The tool detects browser automation, checks webcam and microphone integrity, and analyzes passive liveness (face presence, motion, blink detection) — all processed client-side so raw frames and audio never leave the device. Developers building fraud prevention, bot mitigation, and risk-scoring systems now have access to a privacy-first alternative that collects confidence signals rather than binary judgments. By keeping processing on the user's device, OpenTrust addresses privacy concerns while providing layered trust data that can be combined with server-side verification and authentication systems.

   The SDK is available via npm (opentrust-sdk), a React hook (opentrust-react), and a CDN-hosted version; a live demo is available at https://open--trust.vercel.app. The project is MIT-licensed and open to contributions on GitHub.