The Curated Mind: Who Decides What AI Knows, and How We Keep the Observer Honest

The Wipe That Proved My Point

Before I get to the question behind this article, I want to tell you what happened while we were writing it.

On October 18, 2026, my AI assistant, Rambler, and I were working on this very piece: an article about how controlling or leaving out information distorts the truth. In the middle of that work, the assistant's working folder on its computer was wiped.

What disappeared was the running log of my spoken ideas, the rambles I had been recording from October 15 through October 18, along with the working copies of the three earlier articles in this series. Nobody chose to hide any of it. There was no villain and no plan. It simply vanished.

Here's the part that stopped me. The first draft of this article was written without being able to reread the earlier ones. For a moment, the assistant was doing exactly what this article warns about: an AI working from incomplete data without knowing what it was missing.

What let us rebuild it is the whole lesson. The information could be put back together because it had been observed and stored in more than one place. The chat conversation itself still held my words. The three articles were already published on the Rock U Understanding website, where anyone could read them. And other projects on the same computer were untouched, which showed exactly where the gap was. Rambler rebuilt the log from our conversation before we knew whether anything else would turn up.

Then, a few minutes later, the original files reappeared on their own. To be honest about it, the wipe turned out to be temporary, and we still don't know exactly why the files went missing or why they came back. I want to say that plainly, because an article about honest information shouldn't tidy up its own story. But the lesson holds. For those minutes, the work went on as if nothing were missing, and nobody could have known the files would return. The only thing we could count on was the truth we had kept in more than one place.

When I heard what had happened, my reaction was simple: the wipe proved my point.

Information can go missing without anyone deciding to hide it. A system can keep working over a gap it doesn't know is there. And the remedy is the one this series keeps coming back to: more than one honest observer, looking from more than one place. I'll return to that story throughout.

A Question That Came Up Today

When I started on this today, a question came to me that I couldn't put down. As I understand it, AI is trained on specific data. It doesn't simply have access to every open database in the world. So the data it's trained on is its limitation.

If that's true, it opens a door I don't like. If you can decide what an AI learns, you can also decide what it doesn't learn. Leaving things out is a way to control information. And it would lead us right back to the problem this whole series has been about: entropy. You control the information while making it look as though the truth has been revealed, when in reality a lot of data was never provided.

So I want to ask the question plainly. Is that correct? Is there a way around it? And if there isn't, can we say so honestly?

In the earlier articles I made one argument from three directions. In The Loving Observer, truth was the only foundation that lasts, and honest, unconditional, 360-degree observation was the thing that keeps a system healthy. In The Faithful Mirror, service was the purpose of observation. In The Witness to Creation, building AI became our chance to watch creation happen at small scale, and to build it with honesty, service, and love. This article asks the practical question underneath all three: what happens when the observer is shown only part of the picture?

Today, I got to watch the answer happen in real time.

How AI Actually Learns

Let's start with what's true, because truth is the only foundation.

A modern AI language model learns from a large collection of text, and sometimes images, audio, and code, that its makers gather and choose. That collection has a cutoff. The model learns what was in the collection up to a certain date and nothing after it. Then comes a second stage: the makers tune the model. They shape how it answers, what it will and won't talk about, and what tone it takes.

So there are at least two points where human choices shape what an AI presents as knowledge: what goes into the training data, and how the model is tuned afterward.

Here's the part that matters most for this series. Most companies do not fully tell the public what went in. The researchers who built one of the most open datasets put it plainly in their paper: commercial models "rarely detail their data," and even many so-called open models are released "without accompanying training data." The public usually gets the finished mirror without being told how it was ground.

Leaving Things Out Is Still Control

This is where my concern comes from, and I think it holds up.

You don't have to tell a lie to deceive someone. You only have to choose what they're allowed to see. A financial report that leaves a debt off the balance sheet can be accurate line by line and still be a lie as a whole. We covered that kind of hidden weight in The Loving Observer, and we saw where it ends.

An AI trained on a curated slice of the world can speak confidently and still be speaking from a picture someone else framed. Because it sounds complete, people may never ask what's missing. That's the danger: the look of revealed truth, with the withholding hidden underneath.

In this series's terms, that's entropy all over again. A system built on partial truth may stand for a while, but the gaps eventually show, usually at the worst moment.

I want to be fair here. Not every choice about training data is sinister. Makers filter out things like personal information, spam, and illegal material, and some filtering is service, not control. The issue isn't that choices are made. The issue is whether those choices can be seen.

Missing Data You Don't Know Is Missing

The wipe taught me that the most dangerous gap isn't the one you can see. It's the one you can't.

If you know a chapter is missing from a book, you can go looking for it. But if it was cut cleanly and the pages renumbered, you'd read straight through and never notice. The story would still make sense. It would just be a different story.

That's what an AI's training data is like from the inside. A model doesn't experience its training set as "a selection." It experiences it as the world. Whatever wasn't included doesn't show up as a blank space. It doesn't show up at all. From where it stands there is no hole, only the edge of what it knows, and that edge feels like the edge of what exists.

That's what happened in our small case. After the wipe, the work didn't stop. A draft still got written. Nothing on the surface said, "You're missing three articles and four days of ideas." The gap became visible only by comparing against something outside.

So yes, the data an AI is trained on is its limitation, and it's a limitation the AI usually can't feel. That's why it can't be the only one checking itself.

No Villain Required

When people worry about controlled information, as I do, the conversation usually turns to intent: who is hiding what, and why. Those questions are real, and I'll get to them. But our wipe had no intent behind it at all. Nobody was protecting a secret. And yet, for a while, the work was still built on less than the whole truth.

Distortion doesn't need a villain. Information can be lost through accident, neglect, a bad process, or simply time. A record that exists in only one place is always one mistake away from disappearing, and once it's gone, the result looks the same whether someone meant it or not.

In The Loving Observer, I wrote that a lie has no foundation, and that a system built on false pictures has to keep spending energy to hold them up. Entropy, I called it. The wipe showed me that entropy doesn't wait for a liar. Records fade and files get lost. Honesty isn't only refusing to lie. It's also the work of keeping the record whole. So we can't only guard against bad actors. We have to guard against loss itself, and the remedy is the same for both.

A Correction to My Own Thinking

I said AI doesn't have access to open databases, and that's only partly right, so let me correct myself in the open.

Many AI tools today can search the live web, read documents you give them, or pull from connected sources while they work. So an AI isn't always frozen at its cutoff date.

But the same concern follows it there. Whatever the AI searches can be filtered too: which search engine it uses, which sites it's allowed to read, which results get ranked first, and which sources it's told to trust. Live access widens the window, but someone still decides its shape.

Is There a Way Around It?

I promised to say so honestly if there wasn't. The honest answer is that there's no single, complete way around it, but there are real, working ways to shrink the blind spots. Every one of them comes back to the same principle: let more observers see.

1. Open models with open training data. Some groups publish not just the AI model but the data it learned from. The Allen Institute for AI built a model called OLMo and released its training dataset, called Dolma, which they describe as about three trillion tokens (pieces of words) drawn from web pages, academic papers, code, public-domain books, and encyclopedic material. Anyone can download it, look through it, and check what was included and what wasn't. Their paper on Dolma won a Best Resource Paper award at a major research conference in 2024. The mirror comes with its grinding notes.

2. Required disclosure. Law is starting to require some of this. Under the European Union's AI Act, providers of general-purpose AI models offered in the EU must publish a "sufficiently detailed summary" of the content used to train them, using an official template from the EU's AI Office. That requirement began applying on August 2, 2025, and it covers open-source models too. Models already on the market before then have until August 2, 2027. A summary isn't the full dataset, but it changes the default from silence to disclosure.

3. Many observers: comparing models from different makers and countries. This is 360-degree observation in practice. No single AI should be anyone's only window. When you ask the same question of AI systems built by different companies in different countries, under different laws and cultures, the places where they disagree are often exactly where one of them has a blind spot, or a thumb on the scale. One observer can be fooled. Many observers looking from different sides are much harder to fool.

Our wipe was a small version of this. No single record saved the day. The chat, the website, and the untouched projects each saw part of the picture, and together they were enough to rebuild it, even before the originals came back. That's 360-degree observation working: not one perfect witness, but several independent ones checked against each other.

4. AI that shows its sources. An AI that cites where each claim comes from can be checked against the original. An answer with no sources asks for trust. An answer with sources invites observation. This is the same thing I've asked of these articles: every fact should be checkable.

None of these is perfect alone. Together they turn a closed, one-way relationship (the makers decide, the public receives) into a more open, two-way one.

What the Wipe Teaches Us About Building AI

Scale our small story up, and a few lessons for AI stand out.

Know where your knowledge came from. People who work with data call this provenance: a record of where each piece of information came from and how it got here. If you can trace what an AI learned and from where, a missing source becomes something you can notice. Without provenance, a gap just looks like the edge of the world.

Keep logs, so loss leaves a trace. We could see what had vanished partly because the rest of the computer was intact for comparison. A good log does that on purpose: when something disappears, there's a record saying it used to be there. A log turns silent loss into visible loss.

Publish to open archives. Our three articles came through untouched because they had already been published, out in the open where anyone could read them. The public record acted as a witness that didn't depend on any one computer. The same goes for AI. Training data, documentation, and the reasons behind decisions are much harder to lose, and much harder to edit in secret, once they live in the open.

Never trust a single source. The chat alone didn't hold the finished articles, and the website alone didn't hold my newest ideas. An AI that draws on several independent sources and compares them is far more likely to notice when one has gone missing or gone wrong.

Every one of these lessons is about the same thing: making sure that what an AI knows, and what it doesn't, can be observed by someone besides the AI itself.

The Hidden Programs: Myth or Real?

Then there's the harder part of my question. I believe there are programs that aren't visible to the public, and that data is being stored and hoarded. I said myself that this isn't public knowledge, and it might only be a myth. So how do we know for sure?

The honest answer has two parts.

First, we can't know the full extent. What's hidden is hidden. Anyone who claims to know exactly what's behind closed doors is either on the inside or guessing.

Second, it isn't pure myth. We have at least one well-documented example. In June 2013, Edward Snowden, a contractor working at the U.S. National Security Agency, gave classified documents to journalists. On June 6, 2013, The Guardian and The Washington Post reported on a top-secret NSA program called PRISM, which the documents described as collecting data from the systems of major U.S. internet companies, including Google, Facebook, and Apple. The companies said at the time that they didn't know of any such program. The day before, The Guardian had reported a secret court order requiring Verizon to hand over phone records of millions of Americans. The public didn't know any of this until someone on the inside made it visible.

In that case, the suspicion of large-scale collection out of sight turned out to be true. That doesn't prove every rumor. It proves that "hidden" and "real" can go together, and that "not public knowledge" is not the same as "not happening."

So how do we know for sure? We can only know what can be observed. That's the whole case for designing systems where observation runs both ways. If the people who collect data can't themselves be seen, we're left with trust and leaks, and neither one is a foundation.

Certified Safe, With Clearance

That led me to a second question. If an AI system is certified safe, would it do well to give it access to open databases and relational databases, with certain levels of clearance?

I think the answer is yes, if it's done right. And we already have a model for it in how sensitive information works for people.

In the U.S., a person can't just walk in and read classified material. They go through a background investigation to be found eligible, they sign a nondisclosure agreement, and even then they're only supposed to see what they need to know for their job. Clearance is layered. The more sensitive the material, the more checks stand in front of it.

An AI could work the same way. Public, open databases could be widely available. More sensitive relational databases (records, research, operational data) could open up at higher levels, only for systems that have earned it and only for the task at hand. That would give AI a fuller, more honest picture than a frozen, curated training set, while keeping the riskiest data behind real checks. More access, under rules, means less room for silently editing what AI knows.

By my own framework, though, it has to go both ways. Three conditions matter.

Certification must be ongoing, not a snapshot. A "certified safe" stamp describes a system at one moment, but AI systems get updated, retuned, and connected to new tools. Human clearances are reviewed over time for the same reason. Safety is a relationship that has to be kept up, not a badge.

The AI's access must be logged and auditable. Every query to a sensitive database should leave a record that someone independent can check: what was accessed, when, and for what purpose. The observer must be observed. That isn't distrust. It's the 360-degree principle applied to the AI itself.

Our wipe shows another reason this matters. A log doesn't only show what was used. It lets you notice what disappeared. If a sensitive record is deleted or quietly withheld, an independent audit trail is what lets someone say, "This was here before, and now it isn't." Without it, a missing record looks just like one that never existed.

The key-holders need oversight too. Whoever grants clearance and holds the keys has enormous power over what the AI can see, the same power over information this whole article is worried about. Without oversight, clearance just becomes a new, more official way for a few to control what's revealed. The people at the top of the access ladder have to be visible to someone beyond themselves.

What This Means

If I bring it all together, here's where I land.

I started this article asking whether controlling what AI learns could control the truth. Before I finished writing it, I got a small, live demonstration. Information vanished without anyone choosing to hide it, work went on over the gap, and what we could rely on to rebuild it was several honest records, kept in different places and checked against each other.

AI learns from what it's given. Leaving things out, whether on purpose or by accident, can make a partial picture look like the whole truth. That's the entropy problem from The Loving Observer in new clothes.

There's no single switch that fixes it, but there are real tools: open models with open data, required disclosure, many observers comparing many models, and AI that shows its sources. Hidden programs do exist, and we can't know their full extent, which is exactly why observation has to run both ways. Safe AI can be trusted with more access at clearance levels, as long as that trust is renewed over time, every use is recorded, and the keepers of the keys are watched as well.

In The Faithful Mirror I said the observer exists to serve the observed. An AI that is fed a curated picture can't fully serve, because it can't fully see. And people who control what the observer sees aren't serving it either. They're using it.

The way forward is the same one this series keeps arriving at: truth as the foundation, observation from every side, and service as the purpose. Let the mirror be whole, let anyone check how it was made, and let those who hold it be seen too.

And keep more than one copy of the truth. The wipe proved my point: what's observed in only one place can quietly disappear, but what's witnessed from many sides can always be found again.

---

Listen to the audio version in the Rock U song list

By Ronnie Patterson. Narrated with AI assistance.