# When you should use AI, and when you shouldn’t > Source: > Published: 2026-08-04 09:00:00+00:00 Most folks seem happy to let AI write their LinkedIn posts for them. Others, myself included, have AI draft their work memos or slide decks. And some, according to [recent survey data highlighted by AI Secret](https://aisecret.us/deepmind-fired-its-nobel-team/), are happy to let AI do their grocery shopping for them (28%), but not handle luxury goods or banking (6%). In trying to uncover a guiding principle for when mere mortals are willing to cede control to robots, AI Secret’s authors offered a cutting interpretation of our willingness to buy food but not financial products through AI: “That ranking isn’t trust, it’s a map of what people have stopped caring about.” While this feels true-ish, the “ish” is important. For example, based on the garbage we regularly wade through on LinkedIn, AI slop feels like an attempt to care about something we’re told we *should* care about, but don’t. If we did, we’d take the time to write it. Or maybe it’s simply an acknowledgment that many people don’t have the skill to write well, so they entrust AI to do it better than they could. (It doesn’t.) Maybe it’s both. As for the willingness to shop for groceries, it’s not that we don’t care about the milk in our fridge—we clearly do—but perhaps we don’t care about *how* it gets there. We care about the outcome, in other words, but not the process. That feels like the right way to think about AI, and maybe, just maybe, it offers a clue as to when and how you and I should use AI. Last week, [LinkedIn introduced a button that lets users flag a post as “seems like AI slop.”](https://www.linkedin.com/posts/hsrinivasan1_ai-slop-is-a-top-priority-for-all-of-us-share-7488612006321889282-Ps8Z/) It also pulled its own “enhance your post” feature, which contributed to the AI slop deluge by using AI to help compose posts, and replaced it with a proofreader designed not to change the author’s voice. It’s a welcome change. I’ve been happily “seems like AI slopping” ever since. Still, you have to ask, why are people producing these posts in the first place? My guess is that many don’t value posting but feel obliged to do it because “thought leadership” and “career.” Or maybe, as I did this past week, they used AI to ask, “What are my top-performing posts over the past 10 years and what’s the best strategy for replicating that success?” In addition to telling me to post three times a week (LinkedIn’s algorithms like that), it also told me something I already knew: “Your best posts were personal and opinionated.” In other words, exactly the sort of thing that AI can’t write for me. Yes, AI can produce 500 words on how coaching your daughter’s soccer team taught you seven lessons about enterprise procurement. (“Let that sink in.”) The post will be polished and may even be popular. It will also sound like everything else in others’ feeds because you and every other person outsourced the one thing readers want most: a personal point of view. Something that makes us human. This doesn’t make all AI-assisted writing bad. Far from it. AI can help a non-native English speaker express an original idea or turn dictated thoughts into a coherent draft. It can also challenge an argument, find missing evidence, or suggest a better structure. I use it for all those things. But if you have nothing to say, AI isn’t going to give you a voice. Not a real one. A couple of years ago, a friend who ran product marketing at a very large technology company told me he was going to start generating sales collateral with AI. First-call decks, email templates, etc. His logic was sound and, honestly, kind of brutal. The collateral his team produced by hand was pulling a few dozen downloads from a sales force numbering in the thousands. If almost nobody wanted it, why pay a junior product marketer to make it? Let the machine publish into the void. Sure, I said. That seems smart. Except it wasn’t. He hadn’t solved the foundational issue. I didn’t know what that issue was, but then neither did he. After all, those low download numbers could suggest any number of things. Maybe the sellers couldn’t find it. Maybe the material wasn’t useful at scale, but perhaps a few dozen people used it to close enormous deals. Perhaps it was simply the wrong collateral for a pressing need, which the sellers resolved on their own. Or maybe, just maybe, that thing that every product marketing team does… doesn’t need to be done. AI makes it easier not to wrestle with the problem. The deck costs almost nothing, so we keep making the deck, rather than addressing whether it needs to be created at all. Similarly, because the weekly report takes only minutes, we opt to keep publishing it. The knowledge base fills with pages no one reads because stopping a process requires a decision. Automating it merely requires a prompt. We’re letting AI kick the can down the road for us, rather than making the hard, human decisions we’re ostensibly paid to make. I’ve made a version of this argument about software. [App creation is way up, but app adoption isn’t](https://www.infoworld.com/article/4181971/making-sense-of-too-much-code.html?utm=hybrid_search). Building was never the only constraint. Getting anyone to care is the constraint, and AI doesn’t solve that. It just removes our last excuse for not noticing. This isn’t a case against automating boring things. We should totally do that, and immediately. But “low-value work” hides two very different things, and they deserve opposite treatment. The first is a low-value process attached to a valuable outcome, like my milk example. Take expense reports, backups, etc. These aren’t exciting things to do, but they *must* get done. AI gives us the chance to hand off as much of the process as we safely can, check the result, and move on. The second is a low-value process attached to no discernible outcome, like the sales collateral or a LinkedIn post with no personality or real point of view. Automating this feels like a win because the cost drops, but cost is not the core problem. The real problem is that the output has no audience. It doesn’t need to exist. If you can’t name a useful outcome that would be lost if the output stopped existing, you don’t have an automation opportunity. You have a cancellation opportunity. Now for the part that complicates my own argument, because the inverse is also true. The work where I use AI most aggressively is often the work I care about *most*. I wrote last week about [asking Claude to read four 18th-century probate wills](https://www.infoworld.com/article/4201445/will-open-weights-make-ai-more-honest.html) as part of a 20-year hunt for the parents of a fourth-great-grandfather. It fabricated an entire emigrant ancestor. It was clean, plausible, and completely invented. I only caught it by clicking through to the high-resolution images and reading the documents myself, line by line. I’ve been thinking about that experience differently this week. The model excused its behavior as “hopeful reading.” OK. But the reason I caught it is duller and more important than anything about the model: I caught Claude’s error because I cared. Twenty years of caring made me open the originals. It really, really, *really *mattered to me that it be right. Nobody wants to proofread the deck they never wanted to make in the first place. The same pattern holds at work. I use AI constantly for executive memos, compressing a sprawling pile of data and argument into context, recommendation, and ask. This is high-stakes work with a real reader, a real decision, and my name on it. So I read every line. I challenge the output. I edit (a lot). The AI doesn’t replace my investment in the work, but it does do 90% of the early legwork for me so that I can focus on the critically important last 10%. I’ve called this [AI’s trust tax](https://www.infoworld.com/article/4111829/ais-trust-tax-for-developers.html): You pay the verification cost up-front, or someone else pays it later when the answer causes damage. The tax gets paid most reliably when somebody cares about the outcome. When nobody cares, AI can make bad work look finished enough to ship and presentable enough to glance at without reviewing carefully. AI slop, then, really isn’t a model problem; it’s a question of caring, which manifests in how we choose to use it. I wish I had a clear, guiding principle to offer here, but I don’t. Here are two questions that help, though. If there’s no meaningful outcome, stop doing the work, whether AI-powered or human-powered. If the outcome matters but much of the process doesn’t, hand the process to AI. Let AI buy the groceries, search the archive, assemble the first draft, or build the sales deck that sellers have said they actually need. Then keep a person accountable for the result. AI is extremely good at making more things. We don’t need more things, and making them cheaper won’t make them useful. Humans, by contrast, are extremely good (or need to be) at determining which things need to be made. That’s your job, and mine, and no prompt can take it away from us.