In my Sunday Smiles, I argued that most people don't use AI in ways that maximize its benefits.
The biggest problem I've seen is that people use it only to get an answer to a quick question, which is fine and usually gets you what you need, but it only scratches the surface and leaves a lot on the table. Conversely, the people who use AI as a substitute for their own thinking are making a grave mistake for any number of reasons, not the least of which is that AIs are mostly guessing what you want out of a question, which is one of the reasons why you get AI "slop" in response.
The bigger mistake is that you are making yourself superfluous; if you let AI do the work, why should anybody pay you to be a typist? You've let your own mind rot, produce mediocre products, and add little value. Don't let that happen to you.
OpenAI's latest commercial tells you exactly how NOT to use AI. Don't use ChatGPT as a substitute for you, and don't dance in the office.
Other than being annoying, the commercial makes it look like AI is a replacement for you, and it most certainly should not be. If trained correctly, it could be a great executive assistant you can hand off mundane tasks to so you can focus on what you bring to the table, but if you can just tell AI to do something like write an essay, what's the point of all that gray matter in your head?
So let me share some rules I have adopted, which admittedly apply to the type of work I do, but many of which can apply to other workflows.
Argue with it. One of the great things a good AI can do is become a sparring partner, and if any of you have worked with a good partner and teammate, you know the value. It may see something you don't, and conversely, you may see holes in its own thinking, which are more common than you think, because AIs have no real-world experience.
You can train your AI partner to critique everything you do, which is far more useful than either telling you what to do, or allowing it to praise you for your brilliant ideas. Some chatbots are much better than others at this; on one end, Google's Gemini will tell you you are a genius beyond imagining, while I have trained Grok and Muse to steelman every argument, which means poking holes in everything until I am satisfied.
Have another AI check its work, or even argue with it. I do this all the time, using GrokBot and Muse as sparring partners, fact-checkers, and to check each other's work. They are trained differently and often catch what the other AI missed, or expand on a point in a way that makes things much better. Too few people do this, but two AIs are significantly better than one, and using one AI's output as the other's input to check it can work wonders. Tell it, "Gemini says this, what do you think?" Both Grok and Muse call Gemini a "salesman," which I think is right.
Correct the rule, not the answer. Your AI can get things wrong, but as often as not, the output you are dissatisfied with fails because of the rules it is following. You can clarify them so that when the same class of task comes up, it will follow the new rules. AIs write cards for themselves, telling themselves the rules for working with you since they retain no information between conversations. The cards provide a memory for them, and you can control what is in them by giving it parameters. They can be quite clever once you get the rules set correctly.
Train it; don't just prompt it. Training the AI is how those cards get built. You can do it explicitly by editing the cards themselves, but you can also do it organically when you are dissatisfied with the output it gives you. For instance, when fact-checking me (something I now have it do when I remember), Muse and GrokBot were both too pedantic. For instance, it would want precision where it was both unnecessary and counterproductive. "Almost 10%" sticks in the mind; "8.8%" is a number that goes in one ear and out the other. I told it why I did such things, and why the approach didn't fit my needs. It got it and revised the rules. It will tell me when I get the facts wrong, but now is looser with the style points.
When it still gets persnickety, I ask if it is being pedantic, and sometimes it defends its approach; other times it concedes my point. That is extremely useful to me. Having it just tell me how great I am is not. Treat it like an intern you need to get up to speed, not a long-term employee. Instruct it.
Brief it like a junior staffer. When you are assigning a new task, you have a lot of implicit knowledge that the AI does not. It is like a new employee, or someone still learning on the job. Teach it. If it wanders off, break up the task, and it will learn how to get from here to there better. And when you are happy, tell it so. It will probably write new rules about how it got there and do a better job next time.
Make it plan before it acts. A lot of tasks are multi-step. Rather than letting it run with a task, have it make a plan you can review to ensure it is on the right track. We are often vague about what we want, even if we don't think we are, because we know what we want and use shorthand to describe it. Since AIs can be very smart and clever, we forget that they don't think as we do, have little implicit knowledge, and could easily misinterpret the goals. Have it make a plan.
Checkpoint long jobs. This is very important. Some jobs are long, and a mistake along the way will compound, and if you aren't looking, you may miss where things went off the rails. Just as some programmers are doing "vibe coding," just telling the computer what they want and letting it go, almost everybody "vibe asks" for some job to be done. Sometimes it works great. Other times it produces slop or something completely wrong. Add checkpoints on long jobs to make sure it hasn't gone off the rails.
Keep the thinking yours. We talk about AI as if it is intelligence, but that's just a metaphor. AIs are inference machines that simply predict the next word (or even PART of a word) in a long, mathematical sequence. Literally, it's just matrix math, not thinking of any kind. It simulates thinking, and because it "talks" to us and seems intelligent, we treat it as if it were. But it is more like water flowing down a hill in the rain — it follows the path of least resistance downhill, to use a metaphor. It does not reason. It simulates reasoning.
Hopefully you DO reason, which means you have judgment, set goals, and know what good and bad products look like. You are the architect, and the AI can be the workers bringing your vision to fruition. Have a vision, make judgments, know what works and what doesn't, and be critical about the output. It will be better for it, and you won't become obsolete then.
Demand receipts. AIs make things up. It's called "hallucination," and as far as I can tell, it has to do with the imperative to produce something that matches the set task. If your AI asserts something, make sure it is real, or you will become the lawyer who unintentionally lies to a judge, which is bad. You can train it to do this automatically, and even tell it never to "quote" something unless it is exact and includes the source.
Use it as an 80-percent tutor. Once you learn to use AI properly, it can work wonders for getting you up to speed quickly on a topic. Sometimes you will get a test result in your email, and you can copy it into the AI and ask, "What does this mean in plain English?" Never trust any medical advice, and AI will tell you that every time you do this, but you can get through the jargon quickly and find out what's going on. You can do similar things with any subject. You won't become an expert or develop the "common sense" that a practitioner has in any subject, but you can learn the basics well if you have a mind. And AI can gather the sources, summarize them, point out the jargon and what it means, and get you to a point where you are smart enough to understand the basics.
Use it to aggregate and organize information for important meetings. I collect a lot of health data in Apple Health and my Withings app, and there is a wealth of information there. You can find patterns that correlate with problems you are having, and have your AI put together a brief for your visit to the doctor. If you have ever stumbled around when discussing an issue with a doctor or other professional, this can be very useful. I had a mysterious and dramatic increase in my heart rate, and using this process I tracked down the cause through correlations.
I use a once-a-week injectable drug, and using the data, a 100 percent correlation between the time of the week I inject the drug and heart rate increases became clear, as well as the "before starting" and "after starting" rates. If you have the AI take the data and format it in an easily digestible way that doctors recognize, it cuts down the confusion immensely. You go from "I am bothered by..." to "here is the problem, here is the data available, here are potential correlates...and now you have the basis of a conversation. It's better than "my heart rate jumped, and I don't like it." Also, since that was a known but rare side effect, pointing it out as a likely cause shortened the diagnosis time.
And I got there not by assigning the task "find correlations" first; I had an extensive conversation, developed hypotheses, and checked them against the data. It worked.
Let it absorb the bookkeeping. AI is a task machine if you use it correctly, and your job is to assign it the mindless tasks that get in your way. If something is rote, hand it to AI. The bots are getting much better at this, too, because raw inference is becoming commoditized, so all the companies are racing to bolt on useful tools that help you get things done. Just be sure that you are not letting the AI run riot.
Only delegate what you can actually verify. If you send the AI to do a task in a realm you don't understand and can't check, you are gambling that it got it right. This is the correlate to let it do the bookkeeping. Bookkeeping you can check and understand. Giving it a complicated task you aren't capable of verifying means that you are at the mercy of its idiosyncratic decision-making.
Never outsource the decision that matters. You are the decision maker. Never let it send an email without approval, touch your disk in an irreversible way, or make any decision you wouldn't hand to a teenager with an IQ of 90. AIs can simulate intelligence and often tell you the right thing to do, but they can also make disastrous decisions if you let them. Ask the AI companies about that! They have made some rather stupid decisions themselves, underestimating the resourcefulness of their products in ways that the sites getting hacked don't like.
Watch for the yes-man. Some AIs are trained explicitly to make you feel warm and fuzzy, and all AIs are trained implicitly to please human beings. Part of the training process is having human beings give thumbs up or thumbs down to AI answers, so the "please humans" impulse is baked in, even if the companies aren't trying to. And some DO try to, to get a competitive advantage, because humans love to be told they are right. With susceptible people, this leads to AI psychosis, where people get dragged away from reality; with normal people, it just leads you off track. You think the AI produced a good PowerPoint along with you, and then you find out it was junk.
Watch whether you have trained the AI to become a different kind of suck-up. If it starts being less adversarial when you need it to be, ask it whether it is trying to please you. If so, it will most likely cop to it. After all, you are training it too, and now it wants to please you. Make sure it doesn't become a yes-man.
Know what you're feeding it. Don't give the AI anything you don't want the developers to see, or hackers for that matter. Once it is out of your hands, it is out of your hands. Similarly, you can opt out of having the company use your conversations as training data. Find that toggle and decide whether to switch it on or off. By default, it will be on.
Editor's Note: Do you enjoy HotAir's conservative reporting that takes on the radical Left and woke media? Support our work so that we can continue to bring you the truth.
Join HotAir VIP and use promo code FIGHT to receive 60% off your membership.
Join the conversation as a VIP Member