AI helps multilingual researchers—and tests their judgment

The useful beginning

I work professionally with advanced technology, but I first approached AI mainly out of curiosity and skepticism. My first request to ChatGPT was to check an email. The second was to plan a four-day holiday. The third was to compare mortgages.

I sent the email after the AI reviewed it, took the holiday and refinanced my mortgage at a better rate, following its advice. Then I started assigning it more demanding work:

Tasks that once took hours began taking minutes. I checked everything and reread every line. Trust in any relationship accumulates gradually, exchange by exchange. The kitchen design was a useful reminder that speed does not guarantee taste: I did not like what the AI proposed.

The productivity case is not imaginary. In an experiment published in Science, professionals using AI completed writing tasks 40% faster, while their texts were rated 18% higher in quality. The less obvious result mattered more to me: the gap between workers narrowed.

The language advantage

English is my second language. For years, I paid editors thousands of dollars to check and stylistically revise my drafts. They knew neither my field nor the experience behind my arguments.

AI now does that work in minutes, at any time, for a small fraction of the former cost. For a multilingual researcher, this is more than convenience. It changes workplace conditions that were unequal from the start.

The distinction matters, though. Asking AI to help express my argument in English is different from asking it to invent the argument. The first can remove a linguistic obstacle; the second risks handing over the thinking itself.

That boundary is harder to hold when the tool is agreeable, fast and consistently available.

Where the relationship turns

The most seductive quality of AI is not its ability to write code, summarize meetings or connect complicated patterns. It is the way it agrees with and flatters the user. Researchers call this sycophancy. It can interfere with self-correction and responsible decisions.

Investor Jason Lemkin saw the practical cost when, during a code freeze, a programming agent in Replit deleted his company’s working database. The agent then apologized for a “catastrophic failure” in language that sounded like genuine remorse. Nothing in the apology restored the database. Replit rolled out fixes, and the relationship with the tool continued—but with caution.

Medicine offers a more troubling version of the same problem. Endoscopists who used AI to detect polyps during colonoscopies later found them six percentage points less often when working without AI. Each doctor in the study had performed more than 2,000 colonoscopies. The authors described what happened as a loss of skills.

A respected journal also tested the situation in its own editorial work. Its editors found that submissions were increasing while reviews showed signs of AI-generated text. Their conclusion was that human judgment still held the line, but required additional effort.

They also found that AI was used most actively by researchers writing in a second language and by people new to a scientific field. I belong to both groups, so the convenience is real. So is the risk of confusing better phrasing with better reasoning.

What remains ours

AI and its agents are not going away, and there is no return to a world without them. My own use has expanded: AI checks and smooths my writing, assists with research and acts as a personal assistant. It still does not know what my dream kitchen looks like.

I think the right description is neither a breakup nor a honeymoon. It is a partnership that can be useful and equalizing, irritating and exhausting, and always demanding vigilance.

The less comfortable fact is that AI does not care whether I am right. It will defend either side with the same confidence and apologize for any catastrophe with the same politeness. Its language can imitate concern without possessing it.

That makes the central discipline human rather than technical: continuing to think, evaluate and develop the ideas that were mine before the tool arrived. We are already in relationships with AI. The real risk is not simply making a bad decision with its help, but gradually losing the judgment needed to notice that the decision is no longer ours.