HN Debrief

"Don't You Just Upload It to ChatGPT?"

  • AI
  • Work
  • Language
  • Software Engineering
  • Economics

The post is a translator’s answer to a now-common question: if ChatGPT can translate, why pay a human? The author’s point is not that AI is useless. It is that translation is one of those fields where non-experts routinely mistake fluent output for correct output. A machine can produce something readable very quickly, but that is different from preserving tone, intent, rhythm, cultural references, and consistency across a full work. That gap stays invisible to the people least equipped to judge it, which is exactly why they are confident AI is enough.

Treat AI translation and AI coding the same way you treat any cheap high-speed draft tool. Use it where failure is visible and low cost, but keep humans on the hook anywhere nuance, regulation, brand voice, or long-term maintenance matter.

Discussion mood

Mostly sympathetic to the author and skeptical of claims that fluent AI output equals expert work. The mood was also resigned. Many expect buyers to replace good human work with cheaper "good enough" output anyway, especially in low-stakes or invisible quality domains.

Key insights

  1. 01

    Verification cost gets pushed onto experts

    What looks like cheap automation often just moves the hard work downstream. AI can produce a script, a lesson plan, or a translation that passes a surface test, then a senior engineer, domain expert, or editor has to absorb the risk of reviewing a large blob under time pressure. That is worse than ordinary assistance because the person doing the verification, not the person who generated it, gets paged when the hidden flaw finally matters.

    Price AI-assisted work around review liability, not generation speed. If your team is accepting large AI drafts, tighten ownership and review standards before those drafts become production obligations.

      Attribution:
    • OptionOfT #1
    • baby_souffle #1
    • acyou #1
    • d_runs_far #1
  2. 02

    The real gain is compressing grunt work

    The strongest pro-AI case was not that models are better than experts. It was that they erase boring, backlogged work that skilled people never had time to finish. Fixing hundreds of TypeScript strict-mode errors, cranking out thin CRUD routes, and scaffolding internal tools are all tasks where speed matters more than elegance, and people reported real productivity gains when they stayed in the review loop.

    Aim AI at stale backlogs, repetitive migrations, and disposable tooling first. That is where the speedup is easiest to capture without pretending draft generation solved the whole problem.

      Attribution:
    • r3trohack3r #1
    • greiskul #1
    • black3r #1
    • kouteiheika #1
  3. 03

    Benchmarking translation on famous books is contaminated

    Using a canonical work like The Three Musketeers to test AI translation tells you almost nothing about general capability because the model likely saw the original, multiple translations, or parallel texts during training. Several people noted that this can collapse the task into recall or style mimicry rather than fresh translation. The more revealing test is recent or obscure material where neither source nor translation is already in the corpus.

    Do not make product or hiring decisions from showcase prompts on famous texts. Test with fresh, domain-specific material that the model could not have memorized and that your users actually care about.

      Attribution:
    • geon #1
    • svara #1
    • j_w #1
    • Swizec #1
  4. 04

    People trust AI most outside their expertise

    A recurring pattern was that users dismiss AI in their own domain and praise it in everyone else’s. That is not a paradox. It is what happens when confidence rises as the ability to detect mistakes falls. The result is widespread overtrust in medicine, law, software, and translation by people who only see whether the output sounds plausible.

    Assume enthusiasm about AI quality is partly a measurement problem. If a workflow depends on non-experts approving AI output, add a second check by someone who can actually spot subtle failure modes.

      Attribution:
    • s_tec #1
    • xp84 #1
    • CGMthrowaway #1
    • dmitrygr #1
  5. 05

    Reliability can improve through redundancy

    One useful counterpoint was that unreliable components do not automatically make an unusable system. Engineering often stacks weak parts into stronger systems through margins, comparison, and fallback logic. Applied to AI, that means multiple model runs, cross-checking outputs, and explicit acceptance criteria can turn a flaky generator into a serviceable subsystem, but only when the failure modes are understood well enough to catch.

    If you want to operationalize AI, invest in harnesses, evals, and redundancy instead of pretending one model call is a finished product. Reliability work belongs in the system design, not in optimism about a single answer.

      Attribution:
    • Aurornis #1
    • tgma #1
    • pianopatrick #1
  6. 06

    Good-enough automation can poison the knowledge base

    Several comments zoomed out from labor displacement to information quality. If novices increasingly use AI to generate text in domains they do not understand, the web fills with smooth but shallow explanations. That degrades the material future models train on and makes it harder for readers to find expert-level guidance, especially once companies decide the cheaper mediocre version is acceptable.

    Protect your highest-value knowledge assets from slop contamination. If you publish technical or domain content, make sure it stays expert-authored and clearly differentiated, because discoverability will get worse as generic AI text floods the channel.

      Attribution:
    • perrygeo #1
    • chrsw #1
    • layer8 #1
  7. 07

    Paid translation includes accountability, not just output

    A human translator is not only selling words on a page. They are certifying that the rendering is faithful enough to stand behind in business, legal, or reputational settings. That distinction matters because many buyers are not purchasing raw comprehension. They are purchasing someone who can answer for the result if it is challenged later.

    When you sell or buy AI-assisted professional services, define who is accountable for correctness. That accountability layer is where much of the remaining human value sits, and it should be contracted and priced explicitly.

      Attribution:
    • juancn #1
    • JackFr #1
    • carlosjobim #1

Against the grain

  1. 01

    The capability curve is still moving fast

    The skeptical framing may already be behind the frontier. One comment pointed to simple-prompt results on hard mathematics and argued there is no special barrier that reserves intent, nuance, or cultural reasoning for humans forever. The stronger version of this claim is not that current translation is perfect. It is that many arguments for permanent human advantage quietly assume progress will stop at today's model weaknesses.

    Do not build long-term strategy on today's defects alone. Re-test the newest models against your hardest tasks on a fixed schedule, because the threshold for what counts as automatable is moving.

      Attribution:
    • atleastoptimal #1
    • jaggederest #1
    • TZubiri #1
  2. 02

    Market demand may not support premium quality

    Even commenters who agreed human translation is better said that many clients simply do not need professional-grade work. If the commissioner wants a quick article summary, tomorrow's slide deck, or a rough commercial translation, a perfunctory output may clear the bar. In that world, being better is not enough unless the buyer actually bears the cost of subtle mistakes.

    Segment work by the customer's real risk, not by your craft standard. Premium human service will hold where errors are expensive or visible, but commodity work will keep collapsing toward the cheapest acceptable output.

      Attribution:
    • Marsymars #1
    • throw310822 #1
    • robertnowell #1
    • xboxnolifes #1
  3. 03

    Translation may be one of AI’s strongest domains

    A few commenters flatly rejected the article's premise for ordinary translation work. Their claim was that language transfer is exactly what transformer models are unusually good at, and that even small local models already handle forms, administrative documents, and straightforward prose well enough that a human loop often adds little. Literary work was treated as the exception, not the rule.

    If translation is part of your business, separate literary or high-stakes work from routine operational text. The latter is where AI substitution pressure is likely to arrive first and hardest.

      Attribution:
    • tiborsaas #1
    • Ancapistani #1
    • majdalsado #1
    • dyauspitr #1

In plain english

AI ↩
Artificial intelligence, software that can generate or analyze text, images, code, or other outputs.
ChatGPT ↩
A widely used conversational artificial intelligence product from OpenAI that generates text responses from prompts.
CRUD ↩
Create, read, update, delete, the basic operations used in many business applications and database-backed software.
Gell-Mann amnesia effect ↩
The tendency to spot errors in one domain you know well, then still trust the same source in domains you know less about.
transformer ↩
A neural network architecture introduced for language tasks that became the foundation of most modern large language models.
TypeScript ↩
A programming language that adds static type checking to JavaScript.

Reference links

Background concepts

  • Gell-Mann Amnesia effect
    Used to frame why people trust AI most in fields where they cannot judge its mistakes
  • Ninety–ninety rule
    Referenced to argue AI speeds up the easy first part of software work while the last part stays hard

Social posts and examples

Translation quality examples

Related author post