Inference Labs ads

I wrapped up a workshop on ISO 42001 yesterday, and verification came up more than I expected. A lot of people treat a second look as a check. In August 2025, a senior defence lawyer apologised to Victoria’s Supreme Court after filing submissions in a murder case with made-up quotes and case citations that didn’t exist. His team checked the first few citations and assumed the rest were fine. The prosecution got the same submissions and didn’t check them either. The judge’s associates caught it when they couldn’t find the cases. Two legal teams had the same document. Neither verified all of it. The judge said AI output isn’t acceptable unless it’s “independently and thoroughly verified.” Now we’re automating that same mistake. Ask a second AI model to review the first and it feels like a check. But if both learned from the same bad data or lean on the same flawed source, they’ll agree and both be wrong. It’s two backups on the same drive. When the drive dies, you find out you only ever had one copy. Verification means evidence you can check, criteria set before the review, and a named person accountable for the result.

Sam Altman says the world should "accept some bad things happening" in exchange for AI's benefits. Hacks, scams and "other bad things that will happen." Bold pitch. Aircraft manufacturers accept risk too, but they measure it, disclose it and keep the records. The one time a manufacturer skipped the disclosure part, 346 people died and the fleet was grounded for nearly two years. I've worked in mission-critical environments long enough to know how this usually goes... nobody installs the guardrail until a car goes off the road. Unfortunately, AI's heading the same way, but it doesn't have to. Today the guardrail on a lot of AI agents is a paragraph of instructions asking them to behave. It drifts, falls out of context, and works about as well as a "please drive carefully" sign. Put enforced limits and a tamper-evident record of what ran in place now, before the first inquest. The record is the part we build. Policy without evidence is "trust me bro" for the agentic era.

The U.S. government's AI standards office has had three names since 2023. The guidance it published is the same. Last week's executive order told federal agencies to say "Super Intelligence" instead of "AI." That office, part of the National Institute of Standards and Technology (NIST), went from CAISI to CAISSI. The order says documents that were already published stay as they are. If your company built its AI controls on NIST's AI risk framework, it's still the same framework. A compliance system that tracks sources by name will get this wrong. It'll count one authority as two, or flag old guidance as replaced when only the label changed. Names are metadata. Evidence has to bind to the thing itself. A Inference Labs proof of inference covers what actually ran: the weights, the input and the output. Change one weight and it fails. That's how governance will work at Agentic scale.

AI governance has a latency problem. For CISOs and compliance teams, governance has come down to controls with named owners. Someone approves the access. Someone signs off on the change. Someone answers when it goes wrong. Humans in the loop, at human speed. That's changing whether we want it to or not. Agent systems can now run thousands of interconnected agents, and they're starting to reach across company lines. One produces a wrong result. Another accepts it as fact. Others act on it. The mistake can spread before anyone sees the first alert. So there's a real choice: keep everything at human speed, or let agents, and the research behind them, compound at machine speed. Organizations in other jurisdictions that drop their self-imposed speed limits may pull far ahead. The upside could be enormous. So could the damage. If you're implementing AI governance or helping shape it: which decisions should still wait for a person, and which should the system be allowed to stop on its own?

Ground damage costs aviation ~$10B a year. Your AI saw the incident. 𝐂𝐚𝐧 𝐲𝐨𝐮 𝐩𝐫𝐨𝐯𝐞 𝐢𝐭 𝐭𝐡𝐫𝐞𝐞 𝐦𝐨𝐧𝐭𝐡𝐬 𝐥𝐚𝐭𝐞𝐫? An AI alert lasts seconds. An investigation takes months. Sertn creates a verifiable record of every AI-driven visual event: the model, the input, the output, and the inference result, preserved for independent review. →Prove what was detected, and when →Preserve the exact model and output →Verify results independently, no vendor trust required Detection fades, proof doesn't. Explore Sertn.

Ground damage costs aviation ~$10B a year. Your AI saw the incident. 𝐂𝐚𝐧 𝐲𝐨𝐮 𝐩𝐫𝐨𝐯𝐞 𝐢𝐭 𝐭𝐡𝐫𝐞𝐞 𝐦𝐨𝐧𝐭𝐡𝐬 𝐥𝐚𝐭𝐞𝐫? An AI alert lasts seconds. An investigation takes months. Sertn creates a verifiable record of every AI-driven visual event: the model, the input, the output, and the inference result, preserved for independent review. →Prove what was detected, and when →Preserve the exact model and output →Verify results independently, no vendor trust required Detection fades, proof doesn't. Explore Sertn.