After enough time working alongside an AI coding assistant, both the promotional claims and the reflexive dismissals look equally wrong. It is neither the engineer replacement its vendors imply nor the useless novelty its harshest critics claim. A more accurate description, and a more useful one, is that it behaves like a fast and confident junior who has read an enormous amount and retains none of the specific context of your system, and working productively with it depends entirely on understanding that character rather than expecting something else.
Its genuine value is in work you already know how to do but would rather not type. Boilerplate, test scaffolding, a routine function in a language you touch a few times a year, the exact shape of an API call you always have to look up, these it produces quickly, and because you already understand the task you can evaluate the output at a glance and correct what is wrong. In that mode it is a real accelerator, it handles the mechanical production while you retain the judgment, and the time it saves is not imaginary.
Its weakness is the work that requires holding an entire system in mind. It does not know your architecture, your constraints, the failure that motivated an unusual design decision, or the reason the obvious approach was rejected before, and it will produce code that reads correctly and quietly does the wrong thing, delivered with the same confidence as code that is right. The bugs it introduces are the dangerous kind, plausible enough to survive a quick review and subtle enough to surface later in production. This is why the engineers who gain the most from these tools tend to be the experienced ones, because using an assistant well is fundamentally an act of review, and you cannot meaningfully review code you could not have written yourself. A senior engineer treats the output as a draft from a capable stranger, read critically and often revised or discarded, while a beginner treats it as an answer, ships it, and inherits defects they do not understand, because the tool amplifies existing judgment rather than supplying it.
Used as an accelerator for work the engineer is competent to verify, an AI coding assistant earns its place in a serious workflow and raises throughput without lowering standards. Used as a substitute for understanding the code, it does not remove the difficulty, it defers it, moving the moment of failure downstream to where it is more expensive to diagnose. The competence these tools reward is the same one the profession has always rewarded, the ability to tell correct work from work that merely looks correct, and far from making that skill obsolete, they raise its value, because there is now a great deal more plausible looking code in the world that requires exactly that discernment to trust.