Working Up

by Dwayne Phillips

I would hope that journalists would understand AI and software better than this. I am disappointed.

Yet another story about AI written by the ignorant for the masses:

AI models that lie and cheat appear to be growing in number with reports of deceptive scheming surging in the last six months, a study into the technology has found.

Ah, yet another story about a programmer who didn’t adequately test the software they wrote. Software “scheming” and “deceiving?” And let’s not forget the “lying.” Good gosh, software that lies about its intentions. Software that has intentions of its own. Good grief.

The software did something the programmer didn’t intend. The board of inquiry, (we really don’t have boards of inquiry to grill programmers, but it sounds neat) asked the programmer questions. “I didn’t write that software to do that!” replies the programmer. Examination of the source code reveals, uh, well, yes, the instructions are there. The programmer needed a DWIM button (Do What I Mean). The programmer didn’t have a DWIM button. The computer performed as instructed.

“Let’s review the test procedure and test report,” says the board of inquiry.

“Well,” says the programmer while crouching lower in his chair in the middle of the room surrounded by the aged, wise, and vicious inquirers (again, we don’t really do this, but it makes for drama in a movie), “I sort of didn’t have a procedure and report, but I did test the software.”

AI models come from programs written by people. People test the programs a lot, a little, or not at all. There are methods of doing these things to prevent these headlines. Let’s do better.

And hey, you journalists and editors out there, do better.