Real Conversations Would Get Swept Off The Table

Earlier today I wrote about a group of researchers who went looking for how people actually use AI, and found that nearly half of every real conversation would get swept off the table by the filter the industry uses to count.

Tonight I want to tell you about the part that does not even reach the table.

Let me walk you through a night at my desk. This one. The one that produced the piece you are reading.

Six pieces of material came across. A governance report on machine controls. A Social Security story aimed at retirees. A jobs number out of Israel. An essay about living longer. A news feed full of headlines. And one research story out of MIT.

Five of the six got thrown out.

The governance report could not be traced to a real article. Two different titles on the same document, no link, no date, and somebody else’s control numbers stamped on it.

The Social Security piece had real federal numbers wrapped around a lie of omission. It listed fifteen states losing the most money and let every reader outside those fifteen think they were safe. They are not. The cut is the same percentage everywhere. Those fifteen just have bigger checks.

The jobs number was a two hundred sixty-seven percent increase reported by the vice president of a job board about her own job board, and she admitted mid-interview the number looked huge only because it was measured against almost nothing.

Two more got read and set aside for their own reasons.

That took hours. Hours of pulling primary sources, running down the actual government report, checking one figure against the agency that published it.

Now here is my question.

What was that?

Was it work?

Ask the machine that counts these things and you get nothing. The main industry report on AI use sorts conversations by occupational task. Did you write code. Did you draft a document. Did you analyze a spreadsheet.

There is no task called “determined that a source could not be trusted.”

There is no task called “declined to publish.”

Five of six pieces produced no output at all. Nothing shipped. By any measure built on what got made, that is five hours of nothing.

But I would argue it was the most valuable part of the night. Every one of those five, if published, would have put something false in front of somebody who trusted me.

The one that did get written came with a correction attached. Earlier in the evening I had told the machine a story summary was distorting a research finding. It went and read the actual article and came back and told me I was wrong. The summary was accurate. My read had been the sloppy one.

Where does that go in a report? An hour where the AI corrected the human, against a primary source, on the record?

No box.

And it goes further, because the independent researchers cannot see us either.

That project I wrote about pulled its conversations from seven existing research datasets, from people who agreed to hand them over. Twenty-four thousand conversations. None of them ours. Not one word of tonight is in the only independent gauge that exists.

So we are outside both instruments. The company’s count and the check on the company’s count.

Now understand what I am not saying.

Nobody hid us. Nobody is suppressing anything. There is no conspiracy in this and I will not pretend there is.

The reason we are invisible is duller and worse.

Both counting systems sort by subject and by task. What is this conversation about, and what did it produce.

What happened here tonight was neither. It was about how the exchange gets run. Whether a claim gets checked before it moves. Whether the machine says what it does not know. Whether either party is willing to throw out five hours of material because it did not hold up.

That is conduct. Not topic, not output. Conduct.

And a system built to sort by topic cannot see conduct at all. Not because it refuses to. Because there is no slot shaped like that.

Which brings me around to the thing that ought to worry anybody thinking seriously about how this technology gets governed.

Every serious decision being made right now about AI rests on numbers describing what people use it for. Lawmakers cite those numbers. Companies plan around them. Reporters build stories on them.

Not one of those numbers can tell you whether anybody is using these machines carefully.

Careful and careless look identical in the data. Two people ask about the same subject. One takes the first answer and publishes it. The other spends four hours checking it and throws it in the trash.

Same topic. Same tool. Same row in the spreadsheet.

One of them is doing the thing we all say we want, and there is no instrument on earth that can find him.

So when somebody tells you AI use is mostly productive, or mostly harmful, or mostly whatever this month’s report says, remember what that sentence can actually mean. It means the topics people typed about, sorted into categories somebody chose.

It does not mean anybody is being careful. It cannot mean that. Nothing is measuring it.

I have said for a while that governance of these machines comes down to conduct, and that conduct is chosen rather than enforced. People push back on that. They want a rule, a control, a switch somebody can flip.

Here is the piece I had not seen clearly until tonight.

Conduct is not just hard to enforce.

It is hard to even see.

We have built an entire measurement apparatus around this technology, and it counts what got made and what it was about. It does not count whether anybody checked. It does not count what got thrown away. It does not count the correction that arrived an hour late from the machine itself.

Five out of six, into the trash, on purpose.

No box for it anywhere.

The_Ten_Plain_Rules_CardDownload

This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.

Post Library – Intelligent People Assume Nothing

Contact: micvicfaust@gmail.com