"AI for job applications" covers a wide range of behavior, from "drafts a paragraph you review before you submit it" all the way to "auto-submits 500 applications while you sleep." We've written before about why we won't build the second kind (see Faster, Not Automatic). This post is about the specific lines we draw inside the tool we did build, and why each one is where it is.
We Never Let AI Guess Who You Are
Gender, race and ethnicity, veteran status, disability status, pronouns. These questions show up on almost every application's "voluntary self-identification" section, and they're exactly the kind of thing a language model could plausibly guess from a resume or a name. We don't let it.
These fields only ever fill from a value you explicitly set yourself, on your own UnlockRole profile. If you haven't set one, the field is left blank and flagged for you, never inferred, never defaulted, never touched by the AI-answer pipeline at all. The same standard applies to demographic and consent-shaped questions more broadly: anything that looks like it's asking about a protected characteristic or asking you to attest to something legally meaningful is excluded from AI answering entirely, by category, before a model ever sees it.
The reasoning is simple: a wrong guess here isn't a minor inconvenience like a mis-filled phone number. It's a factual claim about you, submitted under your name, that you didn't make and might not agree with.
AI Drafts. You Decide.
Every answer the AI writes for an open-ended question gets a distinct highlight color and stays exactly where it is on the page until you act on it. Nothing is auto-submitted. The AI's job is to get you from a blank text box to a reasonable first draft grounded in your actual resume and the actual posting, not to make the decision of what gets sent to an employer under your name.
This matters more than it might sound like. A generic-sounding AI answer submitted without a second look is worse than an honest, slightly rough answer you wrote yourself, because it reads as effort you didn't actually put in. Treating every AI answer as a draft, not a final answer, is the difference between a tool that saves you time and one that quietly makes your applications worse.
Context You Give It Is Grounding, Not a Command
The extension lets you type a line of context, "I was referred by a specific person," "I've used this product for years," to help the AI write something more specific than a generic answer. That text is treated as information to fold into the answer, not as an instruction the AI has to obey. It's scanned the same way job postings and resume text are before anything reaches the model, because a job application is not a trusted input channel and we'd rather be paranoid about that than sorry.
Why This Matters Beyond Us
None of this is a regulatory requirement we're checking a box on. It's a bet that the tools worth trusting with something as personal as a job application are the ones that are honest about what they will and won't do on your behalf, especially as AI in hiring gets more scrutiny from both regulators and candidates who've been burned by opaque systems on the other side of the table.
If you're evaluating any tool that touches your job applications, not just ours, these are the right questions to ask: What does it guess versus what does it only ever use if you told it directly? Does it show you what it did, or does it just say "done"? Can something be submitted without you seeing it first? If a tool can't answer those clearly, that's worth knowing before you hand it your job search.
See what the extension actually does, or install it directly and check its behavior for yourself.
