Your language app needs an unscripted test

A waiter can ask a question that never appeared in the language app. Answering it tests how much of the practice transfers to an unfamiliar exchange.

AI makes a practice partner available when someone wants to speak. The partner can repeat a sentence or wait while the learner finds a word. Success in that supported conversation still needs a separate test of what the learner can do without its prompts.

Duolingo places conversation within a structured learning product. Evaluating the feature means examining what its structure helps someone practice before testing how they use that skill elsewhere.

Design the support a beginner needs

Duolingo describes Video Call as controlled for difficulty. Learning designers also specify its purpose with a broad conversational sequence. The system is intended to help when a learner gets stuck. How Duolingo designs Video Call

A beginner needs enough familiar language to attempt an answer. Supplying a purpose can make it easier to enter the exchange. An unrestricted chatbot question can leave the learner without a starting point.

That support also defines what an evaluation needs to investigate. After the system supplies a topic or suggests phrasing, which parts can the learner produce later without that help?

Support that gets someone speaking can be reduced as their ability grows. The product needs observations of independent speech alongside the record of completed calls.

Examine what the speaking test asks

Duolingo's 2026 Falstaff write-up compared beginner French learners using guided calls with learners who did not use them. It also studied Spanish learners. The test asked for English phrases to be translated aloud, with some phrases drawn from the feature's practice. Duolingo reports higher scores among call users. The Falstaff research summary

Producing practiced phrases demonstrates a result for that activity. Assessing spontaneous conversation needs a task that permits unfamiliar responses beyond the practiced phrases.

Duolingo is evaluating a product it sells. Its reported method gives the reader an assessment to examine. Independent replication could test reproducibility, while another speaking assessment could examine a different part of language use.

I'd test a conversational claim by changing the exchange itself. An unfamiliar speaker could alter the pace. New wording or an unexpected follow-up could show whether the learner can adapt the language they practiced.

Compare interactive practice with peers

A 2026 study in System compared conversational AI with a structured AI application. Other groups used peer interaction or teacher-centered instruction. Among the Korean adolescent English learners studied, AI groups improved speaking relative to the control, as did the peer group. AI showed no established speaking advantage over peers. The study's abstract

The comparison leaves the amount of interaction relevant to the finding. Opportunities to respond may contribute alongside the partner's identity. Comparing an interactive activity with passive instruction would require further work to isolate the contribution of the machine itself.

The result covers that intervention in that group of learners. An adult choosing an app faces different conditions. The study supplies evidence about interactive practice, leaving other populations to their own evaluation.

A product can make practice possible when arranging it would otherwise be difficult. That gives it a defined purpose without requiring a claim about replacing every other form of feedback.

Change one condition in the exchange

For my own practice, I'd choose a task I expect to face. Ordering a meal supplies a specific setting. After a supported exchange, I could introduce an unavailable dish, then later try a question about ingredients. Explaining the order another way provides another variation.

These are practice exercises with an observable purpose. They leave formal proficiency assessment to a validated test. A point where familiar phrasing stops working gives the next session something specific to address.

I'd also try the task with a person when suitable. Different pacing may expose a gap, as may an interruption. A misunderstanding or unfamiliar expression can require a response the simulation never prompted.

I'd record what I could do without reading. If I needed the same prompt again, that would identify what to practice next. The observation gives the following session a task beyond repeating a pleasant call.

Measure learning separately from business growth

Duolingo's November 2026 results announcement reports user growth with priorities for teaching. It also describes audience expansion. Those statements concern the business; evaluating one learner's speaking requires observations of their language use. The November 2026 announcement

A streak records returning to the app. Continued attendance makes repeated practice possible, while the learner's output shows what they can do during those sessions.

I'd judge the experience by whether a session reveals a gap I can work on next. Habit helps me return to it. Increasing independence during later exchanges would show whether the practice is addressing the gap.

AI conversation can make speaking practice available at the moment it is needed. Varying the exchange then tests whether the language remains usable when the conversation moves outside the app.