"Don't You Just Upload It to ChatGPT?" The Quiet Argument Inside Every Knowledge Job

There is a particular conversation happening in gyms, kitchens, and group chats right now, and it almost always starts the same way. Someone learns what you do for a living, mentions ChatGPT, and asks some version of the question that has become a small cultural artifact of this moment: don't you just upload it?

A freelance translator based in Ottawa wrote up her version of that exchange recently, and it is worth reading not because it is unusual but because it is so ordinary. She was leaving a boxing class early to handle three translation deadlines that had stacked up over the course of an evening. A fellow gym-goer, polished and on her way in, could not understand the urgency. "It won't take long," she said. "Don't you just upload the documents to ChatGPT?"

The punchline of the piece, and it is a good one, is that the woman asking turned out to be a Director General of Human Resources and Corporate Services, currently acting in Workforce Planning. When asked whether she used AI much at work, she replied: "Oh, I can't! It's really not reliable enough." The same tool that would supposedly make a translator's evening trivial was, in her own department, too unreliable to touch.

That contradiction is the whole story, and it is showing up everywhere.

The pattern: confident about other people's jobs

The interesting thing about the "just upload it" assumption is how consistently it points outward. People tend to trust AI most for work they do not personally do. The translator distrusts it for translation but might use it to draft an email. The HR director distrusts it for HR but assumes it handles translation. The developer who knows exactly where the model hallucinates in code will happily let it summarize a legal document they would never have the expertise to check.

This is a known cognitive habit, and it has a name adjacent to it: the Gell-Mann amnesia effect, where you notice a newspaper getting your specialty wrong, then turn the page and trust it completely on everything else. AI triggers the same reflex. Competence is the thing that makes you skeptical, and most of us are only competent at one or two things.

Community sentiment reflects this split cleanly. On forums where professional translators gather, like the long-running ProZ community, the dominant tone is not panic about replacement but irritation at clients who think the work is now a button press. On the other side, plenty of project managers and procurement teams genuinely believe machine translation has closed the gap, and they are renegotiating rates accordingly. Both groups are looking at the same technology and reaching opposite conclusions, because they are measuring different things.

What the tools actually do

The translator's account is unusually fair to the technology, which is what makes it credible. She is not a refusenik. She has used AI since machine translation got good, and she describes a workflow that a lot of working professionals will recognize.

She feeds 500-page client style guides into a model so it can flag when she has broken a formatting rule. She uses it to extract specialized terminology from reference documents and build glossaries, which she notes is faster than Ctrl+F and less maddening. These are real productivity gains, and they are also nothing like "upload the document, get a finished translation."

The failures she lists are specific and familiar to anyone who has pushed these tools hard: it invents acronyms and organization names, silently skips entire sentences, ignores supplied terminology "unless repeatedly threatened," and sometimes misses the point of the source text entirely. None of these are exotic edge cases. They are the daily texture of working with large language models on anything where correctness is non-negotiable. The current generation of models, including the GPT-4 class systems and competitors, are dramatically better than the DeepL and Google Translate era she came up in, and they still do all of this.

The gap between "impressive demo" and "shippable work product" is where the entire disagreement lives. A model that is right 95 percent of the time looks magical in a demo and becomes a liability in a 500-page legal document, because the 5 percent is invisible until someone who knows the subject finds it. Verification does not scale down with capability. If anything it gets harder, because more fluent output hides its errors better.

The counter-argument worth taking seriously

It would be easy to read this as another "AI is overhyped" piece, and the honest version of the debate does not let translators off that easily. The strongest counter-point is the one the gym-goer actually made: it keeps getting better. That is true, and dismissing it is the move that ages badly.

The trajectory matters. Tasks that were genuinely impossible for machines five years ago are now routine. Literary translation, tone, register, the localization work the translator describes as her real value, all of that is exactly the kind of fuzzy, judgment-heavy work that newer models keep encroaching on. There is a reasonable case that the category of "things only a human can do here" is shrinking, even if it is not gone.

There is also an economic argument the essay raises and then, understandably, resents. If AI lets a translator finish in six hours what used to take ten, clients will ask why they are still paying for ten. Her rebuttal is sharp: you do not pay your roofer less because he uses a hammer instead of his bare hands. Professionals have always used tools, and the value is in the judgment, not the keystrokes. That is correct as a principle. It is also true that markets do not always reward principles, and that rates for commodity translation have already been falling for years. Both can be true at once.

Where the "it keeps getting better" argument gets weaker is the assumption that better automatically means autonomous. Improvement on the model side has not been matched by any equivalent improvement in our ability to trust unverified output. A roofer with a nail gun still has to be a roofer. A faster, more fluent model still produces work that a competent human has to be able to evaluate, which means you still need the competent human. The tool changes what the job feels like. It has not yet removed the person who can tell when the tool is wrong.

The tell

The detail that lingers is the HR director's own answer. She works in workforce planning, the literal job of deciding which roles an organization needs, and she will not trust AI with her own files because it is "not reliable enough." Then she assumes it can do someone else's profession in an evening.

That is not hypocrisy so much as the default human setting right now. We are all generous with our estimate of how easy other people's expertise is, and stingy about our own. The translator's instinct, to learn the tool, fold the useful parts into her workflow, and stay skeptical of the magic-button framing, is probably the only sane posture available. Treat it as a fast, confident, occasionally fabricating assistant that has to be checked by someone who knows the answer. Coach it like a toddler, as she puts it, because that is roughly the supervision overhead it currently demands.

The people most at risk are not the ones doing the work and watching the tool fail in granular ways. They are the ones two steps removed, certain it already works, signing off on output nobody qualified ever read.