Why Verification Beats AI Generation
By Alfonso Gonzalez Bartolessis
Every few months a new tool promises to make translation free. The pitch is always the same: the machine writes, you save money, the problem is solved. And every time, the same thing happens. The text arrives faster than ever and someone still has to decide whether it is right.
That gap between producing text and knowing whether the text is any good is the subject of The AI Imperative, a presentation by Jourik Ciesielski, CTO of Elan Languages, delivered for our Master in AI and Innovation for Localization. Our new audio summary takes the argument apart and follows it to its conclusion: generation has become a commodity, and verification is the business.
It matters to anyone who works with words, code or content, not only to translators. The same shift is happening in law, in software and in publishing. Localization simply got there first, which makes it the most useful place to watch.
Generation became a commodity. The value moved to the people who can tell what is worth keeping. A deep dive into Jourik Ciesielski's presentation The AI Impera
▶ Listen to the summaryThe gold rush paradox
Ciesielski opens with a thought experiment. Imagine a machine that prints endless gold coins. Everyone invests in bigger and faster machines, and gold becomes so abundant that it is worth the same as gravel. In that economy the winners are not the people operating the machines. They are the appraisers, the ones who can look at a mountain of shiny metal and say which pieces are worth keeping.
That is where we are with artificial intelligence. We are not short of content. We are short of people able to judge it.
The valuations do not match the results
The presentation contrasts two signals from the same market. On one side, an August 2026 report says Anthropic's IPO valuation depends on a revenue forecast of 190 to 200 billion dollars by 2028, and funding rounds raise 500 million dollars to reach an 11 billion valuation at series D. On the other, Duolingo's stock fell 64 percent in six months after announcing an AI first strategy.
The market overvalues the tool and underestimates the friction of applying it. It is as if everyone were buying expensive mining equipment while the miners go bankrupt in the field. A technology that changes everything does not fit inside a business model built on the assumption that creating content is slow and costly. When creation becomes instant, the old model is not improved, it is over.
Generation is a commodity, verification is the business
The economics are blunt. When supply is infinite, price falls to zero. Andrew Chen's argument, quoted in the presentation, is that we can now generate unlimited mathematical proofs but have few mathematicians to verify them, unlimited code but few senior programmers to review its architecture, and unlimited legal filings but a fixed number of courts and judges. Human attention is static. It does not scale.
The conclusion is the line that holds the whole talk together: when generation costs nothing, value moves to knowing what is right.
Verification is judgement, not proofreading
The obvious fear is a demotion to professional proofreader, sitting at a desk saying yes or no to machine output all day. The presentation reframes it. An AI can draft a thousand variations of a merger agreement in ten seconds, but choosing the right one takes a master lawyer who understands the client, the vulnerabilities and the regulatory climate. The premium moves from typing to judgement.
Think of the AI as an eager intern arriving with a thousand ideas. You are the executive who knows which one survives contact with reality. That is not a smaller role. It is a role with more responsibility and, properly played, more leverage.
Localization saw it first
The technology behind modern AI was not built for chatbots. It was built for translation. Attention Is All You Need, the paper by researchers at Google and the University of Toronto, tested English to German and English to French. Attention resolves ambiguity: in the sentence about the animal that did not cross the street because it was too tired, or too wide, the model shifts the pronoun according to the final adjective.
Google Neural Machine Translation arrived in 2016 and DeepL in 2017. The industry had a decade of head start. Then, as one slide puts it, it wasted the advantage. Instead of selling verification, cultural adaptation and strategic advice, much of the market sold post-editing at lower rates and treated a paradigm shift as a slightly faster typewriter.
Why the machine still needs you
Two limits keep translation from being a solved problem. The first is that models are a mirror of society. Trained on human content, they reproduce the prejudices present in that data, so a profession statistically associated with one gender defaults to that bias. The second is context. Text is handled as tokens, numerical fragments linked by probability, without any lived experience of what a word does in a room.
That is the root of the so called hallucinations. A translation fails because the model has the syntax but not the situation. It knows the words and not the music.
What this means for language professionals
If you spend most of your week producing raw text, you are competing in a market where the price is falling to zero. If you spend it deciding what is right, adapting it to a real audience and defending that decision, you are selling the one thing that does not scale: judgement.
That is also the argument for training. The translators and project managers who will be in demand are the ones who can evaluate machine output, explain why it fails and redesign the workflow around it. Our Master in AI and Innovation for Localization exists for exactly that transition.
Listen to the audio summary
The full summary runs for 20 minutes in English and is available in English, Spanish, French, German, Italian and Portuguese, with subtitles and a branded PDF. You can find it in the TranslaStars Audio catalogue.
Written by Alfonso Gonzalez Bartolessis