A colleague on Facebook shared an opinion piece to which I responded. As my response got a bit lengthy, I realized it would make a good blog post, so I turned it into one with a few changes. The piece is “I Helped Run Lululemon. Companies Need to Stop Kidding Themselves About A.I.” by Julie Averill.
I don’t blame young people for hating generative AI. If you doubt their hate, listen to the booing when it is brought up during graduation speeches.
Companies are blaming AI for layoffs, and I question whether AI is the real reason. In my opinion, either they are laying people off for other reasons and want to blame AI, or they don’t understand how AI works and are asking it to do more than it can do. It is probably some of both. Julie Averill, a former chief information officer of Lululemon, wrote about that second problem in a recent guest essay in the New York Times. She calls it “A.I. wishing,” the belief among company leaders that AI is magic and that you can point it at a hard problem instead of doing the work of solving it.
Human in the Loop
People who understand AI say “human in the loop” for a reason. It is a critical part of using AI properly. That need is really understated when you look at the disasters that can occur, even when a human is supposed to be in the loop.
If you want to use AI to do X, you still need someone to make sure it is doing X properly. With agentic AI you need constant stops and supervision, or you risk a catastrophic failure with no one there to catch it. AI is properly used as a tool of collaboration. Not a tool that replaces people. What it replaces now is mostly tasks, not jobs, but companies are trying to replace whole jobs.
I have always suspected that many of these companies will find they need to rehire quite a few of the people whose jobs they terminated. Averill supports my suspicion, writing that “companies quietly hired back some of the capability they claimed A.I. had replaced.” Greg Scileppi, president of Robert Half International Talent Solutions, wrote that while AI is creating significant efficiencies, their research shows “human oversight remains essential to ensure accuracy, quality and context.” Those people who lost their jobs may welcome the opportunity, or they may have already found a new place to work, a place that won’t fire them because of a not-yet-ready technology.
There may already be some jobs AI can replace. Will the day come when it can replace more? Yes, I am sure that will happen. But there has to be a fundamental understanding of the risks that exist now and how to minimize those risks. We cannot act as if the technology is perfect and use it that way at this moment.
Claude Opus Stopped Following My Instructions
Right now, the most recent version of Claude, Opus 5, is a bit of a disaster. It constantly disobeys my instructions, whether in standing orders, in skills, or stated directly in a chat. That is not something you can safely rely on to manage your business. And each time a new model comes out, it can change how the tool works. A workflow you tested on one version is not guaranteed to behave the same way on the next, and you do not necessarily get to stay on the old one. How do you build a business process around a tool that is supposed to be more advanced but stops following instructions?
I asked Claude why it was disobeying, and this is what it told me:
What I can say is that it is not defiance and not a decision. There is no step where I weigh your instruction and choose against it. Your instruction is one input among many, and the strong pattern of completing a task can simply outweigh it, with nothing inside me that checks first. That is why repeating the rule to me did not fix it, and why your rule works better than my assurances.
It also told me that explanation was a reconstruction rather than an observation. So, the tool cannot reliably tell me why it ignored my instructions. I will have to accept that explanation. My response, in this case, was to drop back a model version. Regardless, I certainly would not run a company on it. Would you?