Aug 5, 2026, 12:54 PMIn short I'd say AI reviews fail to capture the essence of what reviews are about.
They are not about preemptively catching every possible bug at whatever cost.
Reviews are an inherently human process that involves exchanging information and opinions to find better solutions, compromises and overall to build better software and become better developers.
Aug 5, 2026, 4:51 PM@pac85 imho, contributors should never see the direct output of sashikobot, it should solely be visible to reviewers/maintainers, and let them sort out the bullshit from the useful. oh wait, then this grift to the Linux Foundation's corporate overlords would actually have negative consequences for the almost entirely corpo-backed maintainers... hrm, I wonder why they don't do it this way?
@pac85
This hits on a couple of the ongoing issues with trying to use AI for productivity in general. You can't trust it, and even when it's right it's not necessarily being useful, so it just creates all this text you have to go through... Which ends up being more time spent on messing with the random text generator than actually building understanding within the team and pushing progress on the project.
Aug 5, 2026, 6:25 PM
@atax1a @Raccoon @pac85 I think what you're describing is a result of two key indicators: (a) someone wants to "code" without actually taking the time to LEARN what they're going to do (I blame the "everyone needs to code!!!1" movement) and (b) "nerd" (not a negative thing) culture being appropriated by the mainstream. I'm sure that a third and forth item should be present; (c), which is the instant gratification of information being readily accessible, vs. previous methods (i.e. books) and (d) non stop hype cycles on whatever du jour "technology" MUST be implemented and the negativity projected unto others for using tools that _work_, i.e. PERL, PHP, HTML, etc. I've got nothing but a middle finger for those assholes. Anyways, rant over! Thank you, drive through!
An AI review will _always_ miss intent.
It cannot infer intent even when it _reads_ like it's inferred intent.
But it sure will generate some words for ya. <3
@pac85 It sounds like a worse version of PCI compliance, where an it technician is certifying that a website is secure enough to take credit cards. The process starts by the site being scanned and then a report being made of everything that could on any way conceivably be used to hack the site.
The one time I did it, I remember having to explain with a straight face that no, it's not possible to hack a site through a referenced image.
@pac85 really insightful posts, thanks for taking the time to write it up in so much detail.
Back in the day when I worked on static analysis tools, which had a fair number of false positives, everybody’s first reaction was to put them into the build process and crank the warnings up to the highest level … which had the same kind of counterproductive effects you’re describing here. The more successful deployments made it easy for devs to run the tools on their own before checking in (if they wanted to) and focused the build checking on high-accuracy messages and high-severity bugs where it was okay to wade through some false positives because the real ones were so consequential. Sometimes less really is more!