His Danish dilemma illustrates the difference between sounding fluent and actually becoming proficient. Technology can help someone learn vocabulary, practice skills, or move smoothly through a translated exchange. But proficiency is about what someone can actually do across reading, writing, listening, and speaking—and communicating proficiently still requires navigating another person.

Hughes, who trains language tutors at Rosetta Stone, knows that gap all too well. He speaks Spanish at home with his Costa Rican husband, teaches German, grew up hearing French from his grandmother, and studies Nordic languages for fun. He recently added Japanese to the rotation ahead of a trip to Japan. 

Yet his abilities shift from one language and one skill to another. For Hughes, fluency describes flow—how smoothly and quickly someone speaks. Proficiency is a broader measure of ability. It asks whether someone can understand what is being said, find the vocabulary they need, and communicate accurately when a conversation moves somewhere unexpected.

“For me, it’s a question of skill and a question of ability over a question of flow,” Hughes told Fortune.

Social media confuses fluency and proficiency

Native speakers pause, stammer, and lose their train of thought constantly, making smooth delivery an imperfect way to judge how well someone knows a language. When someone says they are fluent in six or 16 languages, Hughes has a simple follow-up: “What can you do?”

He approaches the answer through four core skills: reading, writing, listening, and speaking. A person can be strong in some and shaky in others, which is why a single label rarely captures much.

A traveler who can order dinner and navigate Paris, Frankfurt, and Rome may reasonably consider themselves a polyglot—a word that simply means “many languages.” A linguist might ask harder questions about vocabulary, pronunciation, accuracy, and whether that same person could discuss an abstract idea in each language.

Social media has made that distinction even more difficult to see. TikTok is filled with polyglots bouncing between languages while ordering food or surprising strangers. The videos are entertaining, often impressive, and perfectly built for an algorithm that does not have time to sit through a verb-conjugation lesson. But a polished clip can show someone firing off a perfect greeting without revealing how they might manage 10 minutes later, once the conversation has wandered off script.

His own collection of languages shows why proficiency can be difficult to summarize. Hughes describes his Spanish as near-native and feels as comfortable speaking it as English. He can understand and read French extremely well, but makes mistakes while speaking. Danish presents almost the opposite problem. Reading comes easily, while listening and pronunciation remain a challenge.

A library that cannot nod back

Technology can help people work on those different skills. AI can generate practice exercises, correct writing, and create reading material at a learner’s level. With Apple bringing Live Translation to AirPods, the company says the feature can help someone understand a conversation in real time.

Hughes has worked with technology in language education since 1995, when he wrote his dissertation on using it to acquire Spanish vocabulary. He’s wary about translation tools and their ability to capture the social and cultural cues surrounding the words.

In his view, AI is like a “huge library” that never closes. It can produce a beginner-level story on demand and can remove some of the friction involved in finding material appropriate for a learner’s level.

When it comes to using AI to practice speaking, however, Hughes is less convinced.

“I’m not there yet,” he said.

People learn languages from other people

Similarly, Duolingo, whose famously persistent green owl has spent years reminding users to finish their lessons, is also not preparing to concede defeat to a pair of translating headphones.

“Real-time translation can help you understand someone in the moment, but it can’t replace the connection, cultural understanding and personal satisfaction that come from learning and speaking a language yourself,” Cem Kansu, Duolingo’s chief product officer, told Fortune.

The company said many people learn English for high-stakes reasons, including pursuing a better job or higher education. For those learners, the problem is not simply that a translation may be unreliable. Even an accurate translation does not demonstrate proficiency. It can relay the words in one exchange, while proficiency means being able to continue the conversation, ask follow-up questions, and respond when the wording or tone changes.

To be sure, Duolingo and Rosetta Stone make money when people continue learning languages. But Hughes pointed out that words go beyond their dictionary definitions and carry more social baggage that must be learned from interacting with people.

For example, the English expression “don’t pull my leg” has a Spanish counterpart that translates literally as “don’t pull my hair.” An app might know to swap one idiom for the other, but that does not mean it understands the tone behind it. It may not know when an expression sounds playful or rude, or whether someone knows their Costa Rican mother-in-law well enough to use it.

That is why human interaction is necessary for proficiency, according to Hughes. In a live exchange, a learner has to understand what another person said, choose a response, watch how it lands, and adjust when the conversation changes. The listening, speaking, cultural knowledge, and social cues all happen at once. A tool can suggest the next sentence, but it cannot do that work for the learner.

Proficiency begins when another person answers, and a real conversation changes as it unfolds. People watch for nods and other signs of active listening, follow what the other person is saying, and adjust their response. Those social cues are part of communicating in any language, and Hughes does not think AI has a good grasp of them yet.

“When was the last time that you had a really, really, really interesting conversation with a computer?” he asked.