AI Can Translate Words, but Not What It Means to Speak a Language

NEW YORK — Real-time translation can help someone understand another person in the moment, but it cannot replace the connection, cultural understanding and personal satisfaction that come from learning and speaking a language, said Cem Kansu, Duolingo’s Chief Product Officer.

The rapid advancement of real-time artificial intelligence translation tools has ignited a debate within the global language-learning sector over the difference between automated translation and human language proficiency. Apple has added live translation capabilities to AirPods, while Google has developed continuous translation tools for consumer hardware.

Duolingo Inc. and Rosetta Stone are among the industry incumbents defending the value of language study. Their argument is that algorithms can swap words successfully but cannot foster true “proficiency,” a multidimensional skill rooted in human-to-human connection, cultural nuance and adaptive communication.

The distinction between “fluency” and “proficiency” is also used in educational and linguistic frameworks. Todd Hughes, a linguist who trains language tutors at Rosetta Stone, defines fluency as the “flow” or speed of delivery. Proficiency is a broader measure of functional ability.

Hughes has worked in computer-assisted language learning (CALL) since writing his 1995 doctoral dissertation on Spanish vocabulary acquisition. For him, language ability is less a question of flow than of skill: “For me, it’s a question of skill and a question of ability over a question of flow,” he said.

The American Council on the Teaching of Foreign Languages (ACTFL) Proficiency Guidelines and the Common European Framework of Reference for Languages (CEFR) codify this distinction in global educational standards. They assess reading, writing, listening and speaking, including accuracy, vocabulary breadth and the ability to navigate unexpected conversational shifts.

Kansu noted that many people study languages for high-stakes purposes, including career advancement and higher education. In professional and academic settings, translation hardware may relay words accurately, but it cannot demonstrate a person’s ability to sustain a conversation, ask follow-up questions or adjust to changes in tone.

Sociolinguist Dell Hymes coined the term “communicative competence” in 1972 to describe the ability to apply grammatical rules appropriately in different social situations. The concept includes pragmatics, or the social rules of language, as well as idiomatic expressions.

An English speaker who says “don’t pull my leg” might use the Spanish expression “no me tomes el pelo,” literally meaning “don’t pull my hair.” Modern translation algorithms can swap the phrases, but they often fail to recognize whether the expression fits a formal corporate setting, a casual exchange with a peer or a conversation with a relative.

The commercial stakes are high for companies built around language education. Duolingo was founded in 2011 by Luis von Ahn and Severin Hacker, went public on the NASDAQ in 2021 and built a multi-billion-dollar business around gamified mobile learning. Rosetta Stone, founded in 1992, pioneered digital language education and was acquired by educational technology firm IXL Learning in 2021. It has long relied on structured immersion methodologies.

Both companies are operating in a market where consumer AI can instantly translate spoken conversations, potentially challenging the perceived necessity of learning a second language. Hughes described AI tools as a “huge library” that never closes, useful for retrieving material or generating exercises but incapable of practicing active, real-time social negotiation.

Short-form video on platforms such as TikTok has added another complication. “Polyglot” creators can attract millions of views by rapidly switching between languages to order food or surprise native speakers. Their curated clips may rely on rote memorization of scripted greetings rather than the sustained ability needed to discuss abstract concepts, negotiate business terms or manage unexpected changes in dialogue.

Hughes’ own language experience reflects the uneven abilities that can exist across different modalities. He grew up hearing French from his grandmother, speaks Spanish at home with his Costa Rican husband, teaches German, and studies Nordic languages and Japanese.

He describes his Spanish as near-native. He can read a Danish novel with ease, yet his spoken Danish and listening comprehension remain challenging in a room of native speakers. A single label of “fluent” rarely captures a person’s actual linguistic limits.

The ultimate limitation of generative AI in language acquisition, according to experts, is its lack of genuine reciprocal interaction. True language acquisition requires a feedback loop involving non-verbal cues, including nodding and facial expressions, as well as adjustments to immediate social feedback.

“When was the last time that you had a really, really, really interesting conversation with a computer?” Hughes asked, emphasizing that language remains fundamentally a tool for human-to-human connection.