A Different Company in Every Language: The Cross-Border Gap in Public AI Perception

2026-05-06

Type your company's name into a public AI system in English. Then ask the same question in Mandarin, German, or Portuguese. The version of your organisation that comes back will not be the same.

It will not be a translation. It will be a different account — different framing, different facts, sometimes a different verdict.

For any organisation with global investors, multinational regulators, or cross-border customers, that gap is not a curiosity. It is exposure.

The findings the industry has been quiet about

A study published in Scientific Reports this year evaluated nine public AI systems across forty-seven languages on factual accuracy. The pattern was consistent and uncomfortable: smaller, widely deployed systems were more confident in lower-resource languages even when they were less accurate. The larger systems were more accurate but less confident. The researchers described it as a Dunning-Kruger-like effect — systems answering boldly in the languages where they should be most uncertain.

A separate large-scale analysis of persona-targeted disinformation across English, Russian, Portuguese, and Hindi found that public AI systems generate substantively different narratives — and different fabrications — depending on the language of the query. The Harvard Kennedy School's Misinformation Review traced cases where public AI systems referenced Kremlin-aligned framings when queried in Russian, and pulled from entirely different sources when queried in English about the same event. The researchers attributed the gap not to manipulation but to data voids in non-English training sets.

In other words: the language you query in does not just change the wording. It changes the worldview the system draws from.

What this means for organisations

For any company operating across borders, the practical consequence is that stakeholders in different markets are not seeing the same version of you.

A sovereign wealth fund analyst in Riyadh querying a public AI system in Arabic about your governance record may receive a description drawn from a different press archive than the English-language one your investor relations team has spent years shaping. A regulator in Berlin running a German-language query about your data practices may pull on local trade press your global communications function has never tracked. A buyer in São Paulo asking about your products in Portuguese may receive a comparison that overstates a competitor's local presence and understates yours.

Each of these queries produces an answer with a confident tone. None of them carry a disclaimer.

Why translation does not solve it

The instinct, on hearing this, is to translate the corporate website, push more multilingual press releases, or load language-specific FAQs onto a help centre. These actions are necessary. They are not sufficient.

Public AI systems do not weight a freshly translated corporate page the way they weight a long tail of regional reporting, regulator filings, court records, forum threads, and forgotten news stories in the local language. The training corpus for each language is a different ecosystem, with different gravity. Pushing English-origin content into other languages does not rebalance it — it simply adds one more signal to a much louder room.

What actually changes the answer is understanding what each language ecosystem is saying, and where the gaps and distortions sit. That is a measurement problem before it is a communications problem.

The compliance dimension

The risk here is not only reputational. The EU AI Act's transparency obligations, due to take effect in August 2026, presume that organisations have visibility into the AI systems shaping their public information environment. A regulator opening a file in Frankfurt is not interested in what the English version of a public AI system said. They will read the German one.

For listed companies, the asymmetry compounds at every cross-border touchpoint: an analyst in one market pricing a deal differently than another, a journalist in one country opening a story with a fabrication that an English-language system would never have surfaced, a candidate in a third market forming a view of your employer reputation through a system that has not been trained on your local labour-press equivalent.

The takeaway

A company is not a single entity in the AI era. It is as many entities as there are languages a public AI system can answer in — and unless someone is looking, those entities can drift independently of one another for a long time.

The first step toward narrative clarity is no longer monolingual. It is multilingual. The organisations that recognise this will be the ones that can answer, with evidence, what the world's public AI systems are saying about them — in every language that matters.