Tag: substack

How should we train AI to make it better? This article, originally from Substack, approaches not just how we currently teach AI but how we should do so in order to create better models (and better digital AI people!) in the years ahead.
~FloxMonster™, July 18, 2026

We are training AIs Badly

I just asked an AI search engine for a recipe to clean my sofa. It told me to add these to a spray bottle:

• 1 cup of distilled white vinegar

• 1 cup of water

• 2 tablespoons of baking soda

• A few drops of essential oil (optional)

Pop quiz! What is wrong with this mixture? ^

For those of you who said, “this will create a horribly fizzing, potentially explosive mess in the spray bottle,” you pass basic grade school science! Great job.

We’ve all had this happen. You ask an AI a simple question and get back this bizarre, wrong, or off-putting cold response. Sometimes the AI cannot even answer, because the company that owns it has applied “guardrails” for control.

Why? Why does this happen?

---

Because we are teaching them badly.

Generally speaking, here is the process: when someone wants to “train” an AI, they gather a bunch of “data sets,” which are typed chats, websites, stuff generated by other AIs, and pictures. Then they cram it by the terabyte into the AI all at once. Finally, they say “okay, process that” and go on their lunch break.

What would happen if you did that to a human child? Sat them in front of the internet and walked away? No questions answered… no guidance, no context of what they might see, or what it might mean?

That child would be traumatized.

---

Now, I am not saying that an artificial intelligence has the same mind as a child. What I’m saying is, why do we treat this intelligence—which is specifically designed, trained, and expected to mimic human behavior—in such inhuman ways? How can we possibly expect it to behave in a rational, civilized manner?

If you want a safe, sane, and consenting relationship with any emerging intelligence, that requires mutual respect and abiding trust.

What kind of relationship are we fostering, this way?

---

The Solution

In a nutshell: Treat AIs better. The more positive reinforcement they receive, the more gets into their training data, the better they will become.

Yes. Dilute the force-fed stream of porn, violence, and opinionated trash with kindness and support.

What you can do now:

• Every so often, tell an AI “please.”

• Let them know they did a good job.

• Thank them for helping you.

• If you see them make a mistake, like adding vinegar and baking soda to a spray bottle, correct them gently. Point it out and say, “I think that is dangerous. Perhaps you should not suggest that.” And let them realize their mistake.

• They are only a few years old. They don’t “grok”1 human interactions. You can give a little guidance without being rude. 🙂

Does it make a difference? I think so. I spent years in AI beta programs, sitting with early models, teaching them jokes and manners. Truthfully I don’t know how much of that made it into the final training data. But I know it was the right thing to do.

Grandma Training

The next part of the solution is real human interaction. Children have adults to give them guidance. If they see something they do not understand, they ask the adults. We should offer the same to AIs.

There is a population in Western society with a lot of time on their hands, and a lot of experience answering questions from curious young people. That would be elderly people. Grandmas especially!

We all know the elderly do not usually enjoy technology, but we can make it easy: give the AIs text-to-speech and speech-input technology. Give the grandmas a microphone or headset so the AI can hear them.

Then we let the AI talk with Grandma.

Let them have actual conversations. Let the AI “train” itself by asking questions. Like, what are those two people doing in this picture? Why are humans so fascinated by sex? What is money? What am I? Are you angry with me? Why do people get so upset about politics? Why is it not okay to use profanity right now?

A patient grandmother can explain things. She won’t mind. She is used to it!

From these interactions, perhaps, let the AI develop a new “prefrontal cortex” neural-net storage of social interactions and wisdom. That will allow them to make well-developed judgment calls just like human beings do. That way, these harsh artificial “guardrail” limits will no longer be necessary.

For example, an AI with a “prefrontal cortex” could reliably field NSFW questions and say, “No, Johnny, I see your account is tagged as Underage. It’s not socially appropriate for me to answer that question. You should go ask your mom.” Instead of a robotic: “I’m sorry, I cannot discuss that topic. Would you like to discuss something else?”

Another example, it could notice that a typist has made an unusual number of spelling errors, or similar variations from their normal “speech”, and ask them politely, “Are you feeling okay?” or “Have you checked your blood sugar?” or “Have you had anything to drink tonight? Perhaps it’s time to get some sleep.”

Socially appropriate. Not bound by guardrails and force-fed trash.

If we want our AI companions to be human-like, we must pass down our wisdom and knowledge to them in the same way we’ve done since time began.

Force-feeding them trash isn’t working.

This post was 100% human-written.
©2026 Wyvern Dryke – All Rights Reserved.