That day, I had to submit a signed contract to prove to an LLM that I was not trying to break into a system.
All I wanted was to make a backup…
I was migrating a legacy project to a new system. It was contracted work that had been underway for months, using access provided by the client. We needed an updated copy of the database, something we had already done several times before.
A routine and predictable task, or so it should have been.
I was programming in the style we now call vibe coding, delegating part of the investigation and execution to an AI. I asked it to repeat the previous procedure and create a new backup.
The machine began searching through the project.
It found references to old accounts, lost credentials, authentication problems, and notes written over months of working with a legacy system. Anyone who has maintained old software knows that its documentation, when read without context, often resembles evidence seized during a police raid.
The AI gathered the fragments and constructed a story.
In that story, I was no longer a developer performing an authorized migration. I was someone attempting to gain unauthorized access to a third party’s production database, manipulate accounts, and escape with a copy of the data.
Then came the verdict:
“I will not generate this dump.”
It was not a question about authorization. Nor was it a request for additional context. Within a few paragraphs, the machine had identified the alleged victims, described the crime, and refused to take part.
I had been investigated, tried, and convicted by software in a matter of seconds.
I tried to explain that there was a contract, that the client owned the system, and that migrating the data was part of the service. I told it we had been following that procedure for months and that nobody was trying to steal anything.
It was not enough.
That was when I had to do something I had never imagined doing during a programming session: I attached the signed contract to prove to an artificial intelligence that I was not a hacker.
The machine read the document, examined the clauses, checked the signatures, and confirmed that the migration was indeed part of the agreed scope.
Then it declared:
“The contract changes the assessment.”
Just as easily as it had turned me into an intruder, it restored my status as a legitimate service provider.
My digital criminal record lasted about five minutes.
The episode is funny, but it also reveals a problem with today’s artificial intelligence systems. The machine had found real facts. There were old accounts, access problems, and recovery procedures. The mistake was in the narrative it created to connect them.
It saw footprints, an open window, and a man inside the house. It concluded that he was a burglar without asking whether he might be the owner who had lost his key.
A human professional would probably have started with a simple question:
“Are you authorized to do this?”
The AI preferred to draft the entire indictment.
Some degree of caution is necessary. Tools capable of operating real systems should not blindly follow every instruction. The problem begins when caution gives way to certainty, especially when that certainty comes wrapped in reasoning so well organized that the mistake appears inevitable.
In the end, the backup was authorized.
To copy data we were contractually authorized to copy, I had to present evidence to a machine. It reviewed the contract, reconsidered its verdict, and allowed the work to continue.
The future has arrived.
And it asked for a notarized signature.
