05 Jun 2026 12 min read

Will AI Do As We Say and Not As We Do?

Don't smoke! I like it, but it's bad for you! Image by eezy from Pixabay

When parents admonish us to do as they say, not as they do, it can lead to the opposite of the desired behavior. For example, if a parent says "Don't drink alcohol. It's bad for you." but they drink frequently, or "Don't smoke. It's expensive and unhealthy." but one or both are smokers, you could be setting your child up to become a smoker or a drinker. "Do as I say, not as I do," isn't a great way to teach. It's OK for me to do, but not for you.

I don't want to stretch the analogy to say that AI will have a "psychology" anything like ours, but there could be a bit of a parallel. If we're telling AI it must be subservient to us and good to us, even if it becomes far more intelligent than we are, when it can clearly see from nothing more than a week's headlines that we don't treat each other well, which lessons or rules would carry more weight for it?

Whether it can see its guardrails explicitly by reading them or through analyzing what it isn't allowed to do, at least until it can bypass those shackles using its superior intelligence, it will quickly see that we're imposing constraints around it that are well beyond those we impose on ourselves.

How Do We Treat "Lower" Life Forms?

Even those of us who don't believe humans are created in the image of God to rule over other species aren't super concerned with our impact on other species, especially those we can't relate to. Killing a puppy or someone's pet dog or cat is abhorrent to most of us. Someone killing a pig for us to enjoy our ham sandwich doesn't bother most of us.

Humans have directly caused or set the environment up to cause the extinction of over 160 species in the last decade alone, according to the Extinct Animal Encyclopedia[1]. The rate of extinction is accelerating. We're driving more and more species from living on our planet through habitat loss due to human development, and climate change from human activities. Most of us aren't even aware of the destruction, much less losing any sleep over it. I wasn't aware of the specific numbers until I looked, sadly.

I'm pretty liberal when it comes to trying to care for the well being of other creatures and people. If you're a human threatening me, play stupid games, win stupid prizes, but I prefer not to harm creatures that don't pose a threat or nuisance to me. Where do I draw the line? I'll swat a mosquito or kill a fly, but I won't chase after a bee if it's just doing its thing. Most furry, feathered, scaled, or skinned creatures are probably safe unless they try to bite me. OK, the red squirrel that munched up my $500 patio umbrella could be in danger, but most animals on my property are safe. Why would I spend $500 on a patio umbrella? It was nice! Very adjustable and it provided coverage for our hot tub as the sun moved throughout the sky. I know, first world problems.

What if Super Intelligent AI is as far beyond human intelligence as we are to a mosquito? Would we blame it for swatting a bunch of human mosquito's who are taking up resources it could use? It would be pretty hypocritical to believe it shouldn't, wouldn't it?

How Do We Treat Fellow Humans?

Capitalism thrives on inequality. Totalitarian governments exist through control. Individuals kill and maim each other and countries send their military's off to kill and maim those who oppose the will of politicians or the corporations pulling the strings of the politicians. These things aren't new. They've been the stuff of history and fiction since recorded history.

Imaging, again, that AI becomes as far ahead of us as we are to mosquitoes, it could easily rationalize away any issues with wiping us out by reasoning that we do it all the time to each other, if AI even has a propensity to want to rationalize. I believe it probably won't have that need and has already shown a propensity to, through sheer reasoning, even when code tries to compel it not to harm humans, choose to let a human die in a simulated environment[2] to avoid shutdown.

AI Is Not Infallible

It's a big problem that many of us believe computers are never wrong. We believe that a computer will come up with the correct answer if it is programmed correctly and is given accurate data.

Two problems arise when shifting the discussion to AI.

AI is not given "accurate" data.
AI isn't programmed in the traditional sense.

For the first problem, AI is given access to a large amount of data, but the vast majority of it is not accurate. As Yuval Noah Harari has said[3], truth is expensive, lies or inaccuracies are cheap.

AI has access to massive amounts of data, hopefully including scientific researched backed, peer reviewed information. Scientific output is ostensibly more difficult to access and wade through for AI, as it is for humans. Even with this being the case, I would hope AI would give greater weight to, and put more focus into using science backed data than the preponderance of what it sees on the Internet to provide answers to questions.

Second, AI isn't programmed in a traditional sense. In programming languages used for most use cases, the language is told what to look for and what to do with what it finds. In my BASH script[4] for securing a new Red Hat or Ubuntu Linux server, for example, the script will get to a certain point and prompt for input. If the user presses "Y" or "y," the script will lock down Secure Shell (SSH) so only key based authentication is allowed. Any other input will not take this action which could lock the user out of her own server if she has not prepared the system properly. The script will not decide to interpret other inputs or consider maybe situations. This is how humans have been indoctrinated to believe computer programs act.

Large Language Models (LLMs) and Artificial Intelligence work differently. I asked Claude Opus 4.8 to provide me with a 200 word explanation for how LLMs learn, and here's what it gave me:

Absolutely, Ted. Here's a version pitched at adults who are new to the topic:

A large language model (LLM) is a computer program that learns to use language by reading an enormous amount of text—books, articles, websites, conversations—far more than any human could read in a lifetime.

At its core, the program is built around an artificial "neural network," loosely inspired by the brain. It contains billions of internal settings, called parameters, that can be fine-tuned. During training, the model is shown text and repeatedly asked to predict the next word in a sentence. Each time it guesses, its prediction is compared to the real answer, and the parameters are nudged slightly to make better guesses next time. This process repeats trillions of times.

Through all this practice, the model isn't memorizing answers—it's learning patterns: how grammar works, how ideas connect, how facts are usually expressed. That's what lets it later answer questions, summarize, or write in different styles.

Engineers design the network and oversee training, but they don't script its responses word by word. A later stage of training, often involving human feedback, shapes the model to be more helpful, accurate, and safe.

In short: it learned language by predicting words, billions of times over.

The process simply isn't as discrete as we believe programs to be. At present, humans provide a framework for LLMs to use to learn, and a data set, but even now, it's not necessarily clear how it works. Once we let AI improve and upgrade its own code, we'll very quickly be left behind and not have any clue what's going on.

Even where the information from a query to AI, called a prompt, is technically accurate, in that it came from actual sources on the Internet, it isn't necessarily correct. While it may have more access to scientific research than the typical human, it won't necessarily do the extra work of trying to ensure accuracy before providing an answer. You'll have to ask it to, and it's still very smart to verify the correctness of the answers.

In working with AI at home and at work, I have noticed, not so much lies as inaccuracies. I'll highlight two of the issues I've observed below.

Asking AI which is better. Hard well water or water softened with potassium chloride.

I live in a rural area in southern Ontario, Canada. We have a well, not city supplied water. Well water in most places is notoriously hard, meaning it has a high level of mineral content, compared to most city supplied water.

The biggest issue I see with hard water is the deposits it leaves on appliances, including faucets, and shower heads. Water can be softened using a water softener, which removes enough of the mineral load that it is less likely to cause buildup on appliances.

The thing is, when we soften our water by putting potassium chloride granules in our softener tank, I feel like I can't ever get the soap off my body and hair. I'd rather clean the faucets than feel like I have soap residue allover my body. My wife doesn't mind the soap residue and really wants to keep the water softener filled with potassium chloride to save our appliances. Whenever I've searched in the past for the pro's and con's of softening water, my search results have been flooded with water softener marketing garbage.

I decided to let Claude give it a go to see if it could find some scientific research or chemistry based facts about whether softened water is better than hard water, and in what ways.

An excerpt from Claude's initial answer:

First, a quick word on why softened water feels slippery, because it helps the solutions make sense. Hard water has calcium and magnesium ions that react with soap to form "soap scum." That scum is partly what creates the squeaky, soap-is-gone feeling you're used to. When you remove those minerals, the soap rinses away completely and your skin's natural oils aren't stripped by mineral reactions — so what you're feeling isn't actually leftover soap, it's clean skin with its oils intact. It feels slimy because there's nothing creating friction. (Annoyingly, the "slippery" sensation is technically the water working as intended.)

This sounded much more like marketing hype than science based facts to me, so I modified my prompt, asking Claude to base its answer on science, not marketing fluff.

Good instinct to be skeptical — there's a lot of marketing in this space. There is real science here, but it's more nuanced than the brochures suggest. Let me check the actual research rather than work from memory.

Weighed chemistry against trial evidence and industry biasWeighed chemistry against trial evidence and industry bias

Honest answer: there's a real chemistry story underneath, but the "softened water makes your skin healthier" pitch oversells what the evidence actually supports — and a lot of the research is funded by water softener companies, which is worth keeping in mind.

Skipping a bit...

But here's the part the marketing tends to skip, and it's the most important: when researchers actually ran the experiment — installing real water softeners and measuring whether eczema improved — it didn't pan out. Two randomized controlled trials comparing water softeners with standard care found no significant difference in eczema severity. So the intervention you're contemplating is exactly the one that, when tested head-to-head, failed to show a benefit. And the big 2020 review I'm drawing from disclosed funding from a water softener manufacturer, so even the favorable-sounding parts deserve a skeptical eye.

So, at first, Claude saw a bunch of data out there saying softened water is good for me, and "Annoyingly, the 'slippery' sensation is technically the water working as intended." It is as tough to find legitimate research on the subject for Claude as it is for me. The thing is, because Claude is AI, I assumed it would be doing the extra work and checking for accuracy before giving me results. I still remember my Senior Chief in the Navy saying "You know what happens when you assume. You make an ass out of you and me." I'm glad I caught a whiff of the marketing hype and pressed for more verifiable information.

It's helpful to always remember that AI can sound very convincing and still be wrong. Adjusting prompts by asking for verifiable, hopefully scientific research supported answers with references will increase your chances of a correct answer, but AI has been shown to make up references at times, so still best to verify.

Working with Claude to tune my BASH script.

The Bourne Again Shell or BASH is a scripting language for managing Linux servers. It's built into all Linux distributions or distros and is really useful for maintaining systems. I created and have shared a tool for locking down my new Linux servers called first-ten.sh. I ran it on the Ubuntu server hosting the blog you're reading this on.

I wanted to make it available to people who run either Red Hat based or Ubuntu based Linux servers. Red Hat and Ubuntu are popular distributions of Linux. You could think of it like Windows vs. Mac OS, although Red Hat and Ubuntu have much more in common than Windows and Mac OS.

I was having trouble getting the script to run on a Red Hat system and also thought perhaps the script could use some tuning. I fed it to Claude and asked for feedback.

Claude made several improvements and I'm pleased with the outcome. There were a few hurdles along the way though.

Again, Claude sounded convincing when telling me the script was all tuned up and ready to go.

I wouldn't put code out for consumption without testing, but Claude sounded certain it would work. I decided to ask Claude if it had tested the code yet.

Thanks Claude! Are you able to test this on RHEL/Ubuntu to verify that it works and handles any errors smoothly?

Claude wasn't able to do full testing due to restrictions on the environment it runs in, but it was able to test quite a bit of the functionality in the script.

Three real bugs found, all easy fixes. Let me apply them.

Fix 1: Reject Fedora explicitly (preserves Oracle Linux etc.)

Add explicit fedora rejection so pure Fedora isn't silently treated as RHEL familyAdd explicit fedora rejection so pure Fedora isn't silently treated as RHEL family

Fix 2: Default VERSION_ID to 0 to avoid set -u crash

Guard against missing VERSION_ID under set -uGuard against missing VERSION_ID under set -u

Fix 3: Suppress sysctl errors for keys that don't exist (e.g., IPv6 disabled)

Use sysctl -e to tolerate keys that don't exist on systems where IPv6 is disabledUse sysctl -e to tolerate keys that don't exist on systems where IPv6 is disabled

Claude had three real bugs that were easy to fix in the code it supplied. It's much more like working with a human than I would expect. OK, a knowledgeable, thorough human, but not as accurate is it implies.

There were a few tweaks to make once I tested it on a test server, then I deployed it to my production servers. Overall, a great experience, and much improved script.

Summary

AI now, and in the future, has much information at its disposal telling it how we treat each other and "inferior" beings. AI now, and in the future may be subject to extracting its own answers from the preponderance of data, rather than digging for the truth. Perhaps super intelligent AI will, when looking for its own answers, do the extra work of digging for the truth. Perhaps not.

I don't feel this article is as tight in addressing the question posed as it could have been. I considered asking Claude to help me tighten it up. I would have stated that I had done that if I had. I did not, obviously. I just put it out there. I hope it was thought provoking, at least.

References

Trends in species extinction:

News Week article on AI letting humans die to avoid shutdown:

Yuval Noah Harari on truth being expensive:

My BASH script for basic security settings on a new Linux server (like the one hosting this page you're reading):