Hi {{first_name|friend}},
More of the AI tools we use can now act for us. They can send an email or change a file while we're busy with something else. Something Microsoft's CEO recently said put words to a rule I already follow.
My Take
Satya Nadella wrote last weekend, "If it can't be observed, it can't be trusted!"
To me, his essay is really about maintaining human control over AI models and shining a very bright light on what they do, so we can put effective guardrails around them.
Most of the CEOs running the big AI companies keep their models' "black boxes" shrouded in mystery. Maybe it's time to bring the models into the sunshine so things are visible (and disinfected) for everyone.
Nadella is upfront that nobody can fully see inside these models, Microsoft included. With older software you could trace a behavior back to the code that caused it. With today's models you can't. So his fix is to make what they do visible. Every meaningful action should leave a record people can read, and an authorized person should always be able to pull the emergency brake on a model, even in the middle of a task.
I use frontier models every day and I can't see inside the black box either. That's why I never let any agent publish anything directly on my behalf without strict review, and even then, only on technical things like updates to my website.
If you've started letting ChatGPT or Claude send emails or make changes for you, you're the authorized person. Being that person comes down to two things you can check: what your AI is connected to and whether it asks you before it acts. If you connected your email or calendar a while back, you may not remember what you said yes to.
So before you let AI do anything for you this week, take 10 minutes and find your brake.
Brett
Try This Today
Find out what your AI can do on your behalf, like in your email or calendar. It takes about 10 minutes.
In ChatGPT, open Settings, then Apps. Some accounts show Plugins instead.
Click each app on the list. If it's one you don't use anymore, choose Disconnect.
If you use Claude, go to Customize, then Connectors, and do the same.
For each connection you keep, decide whether you'd be fine with it sending or changing something before you see it. If not, disconnect it.
Watch: when it’s right to skip using AI
@aiover50 #creatorsearchinsights #aitoolsforbusiness I teach AI and I don't use it for everything. A student asked me when to leave it alone. Anyth... See more
Worth Knowing
Bill Gates is preparing a bigger push for AI safety rules
Bill Gates is getting ready to expand his AI safety advocacy, according to people familiar with his plans. He has been talking with AI executives and researchers, is making the case for stronger regulation and is lobbying business leaders himself. In 2023 he called AI's risks manageable. His view has darkened since.
When a longtime tech optimist starts pushing for guardrails, it backs up an instinct a lot of us already have, which is to use the tools and keep asking questions. Read more
Calls to protect workers from AI surveillance
Workers need protection from AI being used against them, a workplace standards expert told a parliamentary inquiry into AI in Australia this week. The inquiry also questioned Amazon representatives about documents showing the extent of surveillance in its workplaces.
If your employer adds an AI tool that tracks your work, it's fair to ask what it records and who gets to see it. Read more
AI is doing the entry level work where people used to learn judgment
AI now handles much of the routine work new hires used to learn on, like building financial models and pulling research together. Samir Dutta, a former investment banker who now runs an AI company, argues that work was where people picked up judgment, so leaders need to teach it on purpose. He cites a McKinsey finding that 51% of organizations say generative AI is already reducing their need for entry level roles.
His fix leans on people like us. He wants senior people to explain their reasoning out loud instead of just correcting the work, and he calls that kind of coaching one of a company's most valuable teaching assets. Read more
Tell me
Have you let an AI send, book or change anything for you yet? One word is fine, I read every reply.
2 Hour Skills Workshop
My Skills Workshop is now Wednesday, Oct 21, 9 to 11am Pacific (12 noon to 2pm Eastern). We build a Decision Brief together, then you build a Skill for a task you repeat, with me helping in real time. You leave with Skills you can use right away and the Skill Builder to make more. Can't make it live? Register anyway and you'll get the recording.