The Algorithm Came for the Translators First

Isabelle Mercier was reviewing a pharmaceutical patent application at 11:47 pm on a Thursday in February when she realised her profession was ending. The 52-year-old had spent twenty-three years translating technical documents between French and English for the World Intellectual Property Organization in Geneva — the kind of precision work that had always been described as untranslatable by machines. But that night, checking a colleague's output against the original German, she noticed something. The syntax was too consistent. The terminology choices were too rapid. Her colleague had stopped translating. She had started editing.

"I knew it immediately," Mercier told me over coffee near the Palais des Nations in March. "The cadence was wrong — too smooth. When you translate, you struggle with certain passages. You leave fingerprints. This had no fingerprints." She paused. "My colleague was using Claude or GPT. And she was getting through four times as much material."

Mercier's revelation is now playing out across the global translation industry — and increasingly, across white-collar professions that once seemed secure from automation. The International Association of Professional Translators and Interpreters estimates that demand for human translation services fell 26 percent between 2023 and 2025. The Bureau of Labor Statistics, which had projected 4 percent annual growth for translators through 2032, quietly revised its forecast downward in January. The new projection: negative 12 percent growth annually through 2030.

What the Data Actually Shows

The thing is, we have been here before — or thought we had. Every few years, machine translation was supposed to eliminate human translators. Google Translate launched in 2006. Neural machine translation arrived in 2016. Each time, the profession adapted. Translators became "post-editors," reviewing machine output rather than creating from scratch. But the current wave is different, and the difference is measurable.

The mathematics are brutal. If one worker can now do what three did before, two workers become redundant — not through malice, but through arithmetic. And translation, it turns out, was merely the canary.

Maria Pearson studies labour economics at the Peterson Institute for International Economics in Washington. She has spent the past eighteen months tracking what she calls "white-collar compression" — the phenomenon of AI tools eliminating mid-skill knowledge work while leaving both routine physical labour and highly creative work relatively intact. "The pattern we're seeing," she told me, "is that AI is not replacing jobs so much as it's replacing tasks. But when you eliminate enough tasks, the job itself becomes unrecognisable — or unnecessary."

The Uncomfortable Truth About "Reskilling"

The standard response to AI displacement — from governments, from consultancies, from tech companies themselves — is reskilling. Workers whose jobs are eliminated will learn new skills and transition to new roles. The International Labour Organization's 2025 Global Employment Trends report devoted forty-three pages to reskilling programmes. The World Economic Forum's Future of Jobs Report 2025 described reskilling as "the defining challenge of our generation."

But here is what the data tells us: reskilling programmes have historically modest success rates, and the success that exists is heavily skewed toward workers who were already highly educated and already in growing sectors.

Autor, who has studied labour market disruption for three decades, does not believe AI displacement will lead to mass permanent unemployment. But he is increasingly worried about what he calls "the middle-out" — the hollowing of middle-skill, middle-income work that once provided stable paths to the middle class. "The jobs that remain," he told me, "will increasingly be either very high-end cognitive work or physical work that's hard to automate. The middle is compressing."

Where the Scientists Disagree

Not everyone agrees that the current moment is exceptional. Erik Brynjolfsson, director of Stanford's Digital Economy Lab, argues that we are in a productivity paradox similar to past technological transitions. "We saw the same pattern with electricity, with computing, with the internet," he told a Brookings Institution panel in February. "Massive displacement predictions that eventually resolved into restructured, often better, employment."

The historical parallel is comforting — and contested. Daron Acemoglu, also at MIT, points out that previous technological transitions unfolded over decades, giving labour markets time to adjust. "The steam engine took a century to fully diffuse," he noted in a recent working paper. "Generative AI is diffusing in years. The adjustment period is compressing while the displacement is accelerating."

The International Monetary Fund's World Economic Outlook from January 2026 attempted to quantify the difference. The Fund's economists estimated that AI could affect 60 percent of jobs in advanced economies — "affecting" meaning either complementing or substituting for human labour. In emerging markets, the figure was 40 percent. But critically, the Fund could not determine what proportion of "affected" jobs would be enhanced versus eliminated. The uncertainty itself was the finding.

The Policy Vacuum

What makes the current moment distinctive is not just the speed of displacement but the absence of coherent policy response. The European Union's AI Act, which took full effect in February 2025, focuses primarily on safety and fundamental rights — important concerns, but largely silent on labour market effects. The United States has no comprehensive AI legislation at the federal level. China's regulatory framework prioritises state control over technological development rather than worker protection.

Some countries are experimenting. Spain introduced a "right to human review" in February requiring that workers be informed when AI tools are used in decisions affecting their employment. Germany's largest union, IG Metall, negotiated the first AI-specific collective bargaining agreement in December 2025, limiting the deployment of automated systems in manufacturing without union consultation. South Korea's National Assembly is debating a bill that would require companies to pay into a "technological displacement fund" proportional to their AI-related productivity gains.

But these remain isolated experiments. The ILO's 2025 report called for a "new social contract for the AI era" — a phrase that has appeared in seven major policy documents since 2023 without any concrete international framework emerging.

What We Still Don't Know

Back in Geneva, Isabelle Mercier has not been laid off — not yet. She now spends her days doing what she once considered unthinkable: reviewing AI-generated translations, catching errors that machines cannot see, adding the human judgment that still matters in legal and diplomatic contexts. Her productivity has tripled. Her team has shrunk from twelve to five.

"I used to love the struggle," she told me as we walked past the Broken Chair sculpture outside the United Nations. "Finding the perfect word. Making something work between languages that shouldn't work. That's gone now. I'm a quality control inspector."

The question that haunts labour economists — the one that Autor and Acemoglu and Brynjolfsson circle around without resolving — is whether Mercier's experience is a temporary adjustment or a permanent condition. Will new forms of work emerge to absorb the displaced? Will the productivity gains from AI translate into shorter working hours, higher wages, and broader prosperity? Or will the benefits accrue primarily to capital owners while workers compete for diminishing scraps?

The honest answer is that we do not know. The economic models that worked for previous technological transitions may not apply to AI. The policy tools that eased past disruptions may prove inadequate. And the workers caught in the transition cannot wait for academic consensus.

Mercier's daughter is studying computer science at ETH Zurich. "She's learning to build the systems that will replace her mother," Mercier said with a thin smile. "I tell her: learn the machines, but don't forget what they're replacing. Someone has to remember what we lost."

The algorithms cannot tell us what that loss will cost. Only time — and the choices we make now — will reveal the answer.