Multilingual AI interviews: Cantonese, Mandarin, and code-switching

A multilingual AI interview is one where the candidate speaks the language they think in and is scored on the same rubric as everyone else. Lantern interviews in 40+ languages and dialects natively rather than in translation, with Cantonese and Mandarin as first-class languages, and follows a candidate who switches between them mid-interview.

We built Lantern in Hong Kong, which is why this is a product decision rather than a roadmap item. In this market an interview that starts in one language and finishes in another isn't an edge case. It's Tuesday.

Why does Hong Kong break most interview tooling?

Because most tooling assumes one interview means one language, chosen in advance. Hong Kong hiring routinely involves Cantonese as the working language, Mandarin for regional counterparts, and English for technical and legal vocabulary — often inside a single answer.

A candidate describing a project will explain the situation in Cantonese, name the systems in English because that's what they're called, and switch to Mandarin when the story moves to a mainland counterpart. Nothing is wrong with that answer. It's how professionals here actually talk. Tooling that forces a single-language selection at the start makes the candidate perform a translation of their own experience while being assessed on how fluently they do it.

What does "first-class Cantonese" actually mean?

It means the interview is conducted in Cantonese rather than conducted in English and translated. The questions are asked in Cantonese, the answer is understood in Cantonese, and the follow-up is generated from the Cantonese answer.

The difference shows up in the follow-ups. A translation pipeline flattens the specific thing a candidate said into a generic version of it, and a generic input produces a generic probe. When the model works in the language directly, the follow-up can pick up the actual phrase the candidate used. Regional dialects are supported for the same reason: a candidate shouldn't have to standardise their speech to be understood by the thing assessing them.

What happens when a candidate switches language mid-interview?

Lantern follows. A candidate can open in Cantonese, drop into English for the technical section, and come back — no restart, no mode switch, no penalty.

This is the feature we get asked to prove most often, and it's a fair thing to test in a demo. Code-switching isn't a fallback for candidates with weak language skills; it's usually a marker of the opposite, and a system that treats it as a failure state is measuring the wrong thing.

Does interviewing in your own language cost you points?

It shouldn't, and that's an explicit design constraint rather than a hope. Scoring runs on one rubric audited for language drift, so nobody scores lower for interviewing in the language they think in.

This is worth being precise about, because it's the most common and most reasonable objection we hear. The risk in any multilingual assessment is that the same answer earns a different score depending on the language it was given in — usually because the evaluation was implicitly calibrated on one language and everything else is measured against that baseline. Treating language drift as something to audit for, continuously, is the only version of this that holds up. It sits alongside the broader bias monitoring described in how we approach the EU AI Act.

What changes operationally when language stops being a constraint?

The scheduling maths changes first. In most multilingual hiring, the bottleneck isn't interviewer capacity in general — it's capacity in a specific language, held by two or three people who also have their own jobs.

• No queuing behind the bilingual interviewer. Every candidate gets the first conversation at the same speed, regardless of which language they need.
• No language-based triage. Teams stop informally deprioritising candidates who'd need a harder interview to arrange, which is a fairness problem that rarely gets named out loud.
• Night and shift coverage. Frontline and hospitality candidates interview at 11pm after a shift, on their phone, in their own language.
• Regional reqs open properly. A role hiring across Hong Kong, the mainland and Southeast Asia runs one process rather than three.

Gammon Construction reached shortlist 40% faster with Lantern, screening thousands of applicants without adding headcount. Hyatt saw 70% faster time-to-fill and 7x ROI on high-volume hospitality roles — the kind of pipeline where scheduling and language coverage decide who finishes the process at all.

Common questions

Do we have to choose the candidate's language in advance?

No. The candidate speaks, and the interview goes where they take it. Choosing in advance is what causes the problem this feature exists to solve.

How do reviewers read a transcript in a language they don't speak?

The scored write-up and the evidence links reach the hiring team in the language the team works in, while the transcript stays in the language it was spoken in. The evidence remains inspectable in the original, which matters if anyone ever needs to check the assessment against what was actually said.

Is scoring quality the same across all 40+ languages?

One rubric is applied to every candidate and audited for language drift, and Cantonese and Mandarin are first-class because of where the product was built. We'd rather you tested this on your own pipeline than took a blanket claim from us — bring real candidates in your real language mix.

If your pipeline runs in three languages and your interview process runs in one, that's the demo to ask for.