AI Risks
Power concentration, governance gaps, societal disruption. The risks of AI are not science fiction — they are already visible in data, in policy failures, and in the decisions being made right now without democratic oversight.
Dario Amodei says the AI industry, including his own company, needs to deliberately slow down model development until safety catches up. Sam Altman and Elon Musk agreed within hours. Yann LeCun called it fake.
An Anthropic researcher used Claude Code to find a heap buffer overflow in the Linux kernel that had gone undetected for 23 years. Five vulnerabilities are already patched. Hundreds more are waiting. The bottleneck has shifted from detection to human validation.
Statistical backdoors in open-source AI model weights are invisible to code review. They live in the weights — not the code. Thousands of servers in critical infrastructure are exposed. No CVE system exists for them. No audit standard exists yet.
AI diagnostic systems fail on 66% of critical conditions in synthetic tests. Rare diseases fall outside training data. Hospitals are deploying these systems without knowing where the gap exists.
AI code assistants leak credentials at 3.2% of commits — 2.4x the human baseline. The CVE curve is accelerating. The tool that promises to secure your code is systematically exposing it.
The Pentagon is deploying $13.4 billion in autonomous weapons. China and Russia are building their own. The UN resolution is non-binding. No country has agreed to limit autonomous lethal systems. The rules are being written by the deployments themselves.
Voice cloning needs 3 seconds of audio. Courts have no forensic standard to detect it. Cases involving contested synthetic evidence are already inside the legal system — they just don't appear as 'deepfake cases' in the records.
A Canadian expert on AI policy testified before Parliament describing AI agents as a national security emergency. The loss of control incidents are already documented. The window to act, he argues, is closing fast.
AI data centers are consuming electricity at a scale that is reshaping national grids. The costs are being distributed across everyone — including the majority who never use AI at all.
Data centers will consume 70% of all high-end memory chips in 2026. The costs land on PC buyers, smartphone users, and anyone still trying to access affordable consumer electronics.
Meta is building the world's largest AI data center in rural Louisiana. The electricity bill, and the gas plants, go to everyone else.
Bernie Sanders delivered the most direct Senate speech on AI and work in a generation. The data is largely real. The framing is sometimes strained. Both things matter.
