where are we heading?

There was a time when a photo was just a photo. Today a photo is something that may or may not be real. It all comes down to how honest the photographer is, because you cannot tell with the naked eye what is true. The same goes for data, only worse. Data is transactions sitting behind layers of code, and usually it takes a human to read what is on row 40001 that throws off EBIT.

select randomness, sum(belopp) as total
from very_important_databasetable
where datum >= '2026-01-01'
group by human
order by total desc;

Now add a factor that generates the code, and that does not validate it before it ships, that is not monitored but gives a thumbs up once it has finagled its way to done, without ever working with the data itself. It only touched the metadata.

And still it moves faster than ever. That is where it chafes. The pressure to deliver shrinks the time to read row 40001. Validation became a bottleneck, and bottlenecks get removed.

But validation was never overhead. It was the work itself. The person who reads the row, who recognizes that the number is too round to be true, who knows that EBIT does not behave like that in March. It is not a check after the work. It is the work.

When I let something generate the code, approve itself and give a thumbs up, I have not saved on validation. I have only moved it. Either to someone further down the chain, a controller, a board, a client. Or nowhere. And nowhere is the expensive option.

Because the error in the data is invisible. Unlike the manipulated photo, there is no one looking who grows suspicious. It moves quietly through the layers. Row 40001 becomes a line in a report becomes a decision. No one lied. No one checked either.

for ten points, we are heading toward...

A place where a thumbs up is enough. Where "it ran" gets confused with "it is correct". We built tools that are convincing long before they are reliable, and we did it at a pace where no one had time to ask about the difference.

I do not think the answer is to slow down, or to keep ai away from the data. That train has left. The answer is to stop treating validation as something you do if you have time. To build the suspicion back in, deliberately, as a step and not an afterthought.

Those who work with data going forward will not be paid to produce. The machine can do that, probably faster and better as time goes on. llms are cheaper metadata consultants, more focused, and they work around the clock at a fraction of the cost. Those of us expected to be made redundant will, in the future, be paid for the questions (and the answers to them): who says this, how do they know, and what does row 40001 actually say. And maybe a little of that unpredictable creativity it takes to work across such different demands in different domains.