the consistent gaslighting from ai

like most people in the industry I would say I've been thinking a lot about ai lately. I'll sheepishly admit that for a while I was definitively on the hater train. no matter what model you used it just seemed to produce subpar work. but around the beginning of 2026 things progressed and as a result I've been forced to sing a different song. these tools are actually useful now and you don't have to be some prompt expert to get real value out of them.

that being said I have noticed a troubling pattern that I don't have a good solution for. the models are confident. really really confident. their language is exact and if you don't know the things you're working on very well it's very easy to accept something they give you as true when the reality is the statement was at best true when the model was trained and maybe it was never true at all.

some personal examples from recent months:

a pr review ai at my company insisted that the version of golang I had specified in a container build did not exist. it was quite insistent that the latest version was 1.23.x at a time when 1.25.y releases had hit double digits for the y. it believed this was the case even though the ci checks that built the container had succeeded. despite replying to its comments with links to the project release page I was unable to convince it that the version was correct and I had to merge the pr over its objections.

in another pr review the ai agent stated that my dependabot additions referenced non-existent ecosystems and github action versions. the ai was sure that I had meant docker instead of the docker-compose I had added. it told me the latest version of actions/checkout was 4 when it was really 6. again the ci success and supporting documentation links failed to change its mind.

and it's not just ai pr review where I see this. a coding ai agent recently told me that it had to use awk for integer division in a shell script because bash did not support arithmetic operations natively. at least in that case when I pointed it towards the man page it was able to course correct. but if I didn't know it was wrong it would have convinced me it was right with its authoritative tone.

I guess it all boils down to "trust but verify" with these things. but how many of the current ai users know enough about the things they're vibe coding to really do that?