Diloop: the best language tutor was always conversation. It just didn’t scale

The loudest language-learning story of 2025 was not a method. It was a memo. In April, Duolingo’s CEO told the company it was going “AI-first” — contractors phased out where AI could do the work, AI in hiring and reviews, new headcount only for what could not be automated — and shipped 148 AI-built courses in a single announcement. The backlash ran for weeks, loud enough that the company spent the next month explaining itself, and it crystallised a suspicion learners had been circling for years: that the gamified feed was being optimised for the company’s metrics, not their Spanish. We have no appetite for dunking on an app that got hundreds of millions of people to show up for a language every day — getting people to show up is genuinely hard, and streaks do that. But we think the memo pointed at the wrong revolution. The interesting thing AI changed about language learning is not who writes the exercises. It is that conversation — the one tutor that always worked and never scaled — suddenly scales.

The skill you need first is taught last

Every learner knows the inversion — in Türkiye it practically has its own folklore: years of school English that cannot order lunch. Classrooms and apps alike teach you to recognise, conjugate and match; the thing you actually need on day one — ask for a coffee, find the platform, survive a small exchange with a stranger — comes last, if it comes at all. Part of the reason is economic: conversation needs a patient interlocutor per learner, and schools have one teacher per thirty or more. Part is psychological, and the applied-linguistics literature has had a name for it since the eighties: foreign language anxiety — the specific dread of being wrong out loud, in front of people, in a language where you sound like a child. The institutional channel is not picking up the slack either; in the United States, the MLA’s census found university language enrolments falling at the steepest rate on record. So the demand did not go away — it moved. And an AI interlocutor changes the two constraints that created the inversion: it is one patient conversation partner per learner, and it is nobody watching. The question stopped being whether machine conversation practice is possible. It is whether the loop around it is honest.

What an honest speaking loop looks like

Diloop’s loop has five steps, and the design choices are in the verbs. You say it: the microphone starts when you start it, in a scenario from real life, and typing is a first-class alternative, not a fallback. You check it: your recording becomes text, you fix the word the transcription misheard, and nothing is sent to the conversation until you confirm — which is both a learning choice and a data-minimisation one, because only your confirmed text travels. You hear the reply, in your learning language, with audio. You notice the difference: the correction comes with its reason, written in your interface language — “Could I have…” and “please” make the request more polite — because a rule you understand transfers to sentences a score never touches. And you try again, in the same session or in the saved conversation you come back to. Around the loop sits a curriculum rather than a feed: level discovery that suggests a starting point, lessons that explain meaning and usage in your own language, themed tracks that approach one topic from several angles, games for the repetition that recall genuinely needs, and progress you can see — results, attempts, personal bests. Nine languages, and the pair is yours to choose: interface language and learning language are fully independent, so a Turkish speaker learning German never routes through English.

Where Diloop stands

Diloop is our speech-first language learning product, live in open beta at diloop.app. It is also the most literal consumer expression of the platform this site describes everywhere else: a speaking loop is speech-to-text, language understanding and text-to-speech in a circle — the same in-house families behind Solab’s transcription and Blab’s voices, through the same gateway — pointed at education. That ownership is what lets the product make its promises structurally: the transcript-confirmation gate, the explanation-over-score design and the disclosure labels are product decisions we can enforce in the stack, not requests we file with a vendor.

• A speaking loop you control: say it, confirm the transcript, hear the reply, see why
• Corrections explained in your interface language — next attempt over score
• Lessons, scenarios, themed tracks and games in one learning experience
• Level discovery as a suggested starting point — not a certification
• Nine languages; interface and learning language independent

What we will not claim

Three honest limits. First, Diloop is not a substitute for the world, and practice with a machine is a bridge, not a destination: the point of the café scenario is the café. The replies and feedback are AI-generated and labelled as exactly that in the product — they are good practice, they are not a certified pronunciation verdict, and when something is wrong there is a report button beside the content, because beta content has not all been reviewed by a person. Second, level discovery is a suggestion about where to start, and we say so in the FAQ: it is not a language examination, and if you need a certificate, you need an exam, not an app. Third, it is an open beta that behaves like one — native-language explanations are rolling out gradually and audio coverage is still growing — and we would rather write that under the feature than let you discover it as a surprise.

The drill asks what you remember. The loop asks what you meant to say — and shows you, in your own language, how to say it better next time.

For readers in Türkiye this is a product with home-field relevance: the gap between years of classroom English and a usable sentence is a national cliché precisely because it is real, and the everyday moments Diloop rehearses — the order, the directions, the introduction — sit exactly in that gap. The design position is the one we keep taking across this platform: the AI is labelled, the user confirms what is sent — only your confirmed text travels, a data-minimisation instinct a KVKK-era product should have — the explanation beats the score, and the stack underneath is ours end to end, which makes those commitments architecture rather than policy. It is in open beta and starting is free — diloop.app, from Arpanet Bilişim A.Ş. Say the sentence badly, see why, and say it better: that is the whole method, and it is yours in nine languages.