DuoChat — Duolingo Case Study

Home

Aaron

✌️

Prado

Projects

Home

Aaron

✌️

Prado

Projects

DuoChat

DuoChat is a new feature designed for Duolingo that makes language learning practical and engaging through short role-play chats and a built-in phrasebook, helping learners practice phrases they can actually use in everyday situations.

App audit

Wireframing & Prototyping

User Flows & Journeys

User Research & Personas

Interaction Design

Role

Designer

Researcher

Duration

~1 Month

Tools

Figma/Figjam

ChatGPT

Notion

TL;DR

Identified the gap between Duolingo's lesson-based learning and real-world conversational confidence

Audited Duolingo's existing flows and design system to design two features that extend - not replace - what's already there

Designed DuoChat, a scenario-based conversational coach, and a Phrasebook widget for quick phrase recall

Mapped the Human vs AI tradeoff and landed on a hybrid opt-in model to preserve user trust

Validated the concept with peers — core idea landed, DuoCon 2025 confirmed the gap was real

Six iPhone mockups showing DuoChat feature screens including lesson section overview, role-play conversation interface, today's review screen, notebook with saved phrases, and the phrasebook home screen widget

The Problem

Duolingo's quirky sentences are intentional. Their research shows strange and memorable phrases beat realistic every time. But knowing that a bear drinks milk doesn't prepare you to order coffee in a café.

The gap is well documented. Research in second language acquisition shows that recognising vocabulary and producing it in conversation rely on entirely different cognitive processes, which explains why learners can finish lessons feeling confident, then freeze the moment a real person speaks to them. This isn't a motivation problem. It's a practice gap.

Duolingo builds vocabulary well but focuses on recognition tasks like multiple choice and translation
An independent study found that Duolingo users showed significantly lower performance on speaking tests compared to their written scores
Without low-stakes conversational practice, learners lose confidence - especially those without access to native speakers

This isn't a motivation problem. It's a practice gap.

Duolingo builds vocabulary well - but focuses on recognition tasks like multiple choice and translation

An independent study found that Duolingo users showed significantly lower performance on speaking tests compared to their written scores Strivecloud

Without low-stakes conversational practice, learners lose confidence — especially those without access to native speakers

DuoChat is designed to close that gap.

Research

This project started with my own frustration as a Duolingo learner, then grew through conversations with other learners - travellers, newcomers, and long-time users.

The pattern was consistent. People with 100-day streaks and solid vocabulary scores still froze in real conversations. The app had done its job, they had built recognition. What it hadn't built was the confidence to produce language under pressure.

This project started with my own frustration as a Duolingo learner, then grew through conversations with other learners - travellers, newcomers, and long-time users.

What I did

Talked with learners, traveller's and newcomers about where the app let them down in real life

Scanned Reddit threads - hundreds of posts surfacing the same frustration: great for vocab, useless for conversation

Audited Duolingo's existing flows and design system to understand what already existed before designing anything new

Mapped the Human vs AI tradeoff in a mind map to decide how DuoChat should handle conversational practice

Talked with learners, traveller's and newcomers about where the app let them down in real life
Scanned Reddit threads - hundreds of posts surfacing the same frustration: great for vocab, useless for conversation
Audited Duolingo's existing flows and design system to understand what already existed before designing anything new
Mapped the Human vs AI tradeoff in a mind map to decide how DuoChat should handle conversational practice

Talked with learners, traveller's and newcomers about where the app let them down in real life
Scanned Reddit threads - hundreds of posts surfacing the same frustration: great for vocab, useless for conversation

Audited Duolingo's existing flows and design system to understand what already existed before designing anything new

Mapped the Human vs AI tradeoff in a mind map to decide how DuoChat should handle conversational practice

User quotes

“I wish I could have practiced sentences I’d use traveling.”

“I don’t use it anymore… it didn’t really help me learn.”

“It’s great for vocab, but not for most real scenarios.”

Low-fidelity wireframes for DuoChat showing early explorations of the conversation interface, scenario selection, and phrasebook screens

Low fidelity wireframes created from early sketches for pre-testing

Design Decisions

Design within the system, not on top of it

Before designing anything, I audited the app. Not just the visual layer — the component patterns, the tone, how Duolingo introduces new features to existing users. That constraint came first because it had to. A feature that looks like it came from somewhere else doesn't get used.

Working from Duolingo's unofficial design system meant every screen, toggle and label already felt familiar. The goal was for a user to encounter DuoChat and assume it had always been there.

Before designing anything, I audited the app. Not just the visual layer - the component patterns, the tone, how Duolingo introduces new features to existing users. That constraint came first because it had to. A feature that looks like it came from somewhere else doesn't get used.

Working from Duolingo's unofficial design system meant every screen, toggle and label already felt familiar.

The goal was for a user to encounter DuoChat and assume it had always been there.

Hex #CE82FF

Hex #2B70C9

Hex #FF4B4B

Hex #4B4B4B

Hex #f7f7f7

Hex #58CC02

Feather

Noto Serif

Hex #CE82FF

Hex #2B70C9

Hex #FF4B4B

Hex #4B4B4B

Hex #f7f7f7

Hex #58CC02

Feather

Noto Serif

The colour and font used for Duochat.

SWOT analysis evaluating AI-powered conversational practice for DuoChat. Strengths: real-time practice, personalised feedback at scale, and alignment with Duolingo's AI roadmap. Weaknesses: simulated conversation can feel artificial, variable user trust, and risk of reducing personal achievement. Opportunities: bridging the gap between lesson completion and real-world confidence, competitive positioning, and future extensions including smartwatch and classroom modes. Threats: potential to replace genuine learning, negative user sentiment around AI undermining personal effort, and reputational risk if AI ethics are handled poorly.

SWOT analysis card showing AI strengths for DuoChat: real-time conversational practice, personalised feedback at scale, enhanced Super subscription value, alignment with Duolingo's AI roadmap, and cross-language support from day one

SWOT analysis card showing AI weaknesses: simulated conversation feeling artificial, variable user trust in AI, over-automation reducing personal achievement, long-term trust damage from poor AI quality, and need for continuous evaluation

SWOT analysis card showing threats: AI replacing genuine language learning, reduced user agency, trend-chasing perception, negative sentiment around undermining personal achievement, and reputational risk from poor AI ethics handling

2. Let the user decide

Conversational AI in a learning context raises a real question: does simulated conversation help, or does it undermine the feeling of genuine progress? Rather than assuming one answer, I mapped the tradeoffs.

Scripted scenarios feel safe but limit flexibility. Full AI role-play is powerful but alienates users who don't trust it.

The hybrid model lets the learner choose — basic scripted practice always available, AI role-play opt-in and clearly labelled. Trust preserved because the user stays in control.

Scripted scenarios feel safe but limit flexibility. Full AI role-play is powerful but alienates users who don't trust it.

The hybrid model lets the learner choose — basic scripted practice always available, AI role-play opt-in and clearly labelled. Trust preserved because the user stays in control.

3. Feel like Duolingo, not a bolt-on

The original plan was a standalone DuoChat section - its own screens, its own navigation paths, its own user journey.

A separate destination inside the app.

Auditing the existing flows killed that idea quickly. A new section meant more navigation steps just to find the feature, more friction for first-time users, and a real risk of disrupting the habits of people who already know exactly where everything lives. A feature nobody can find is the same as not shipping it.

The pivot was to nest everything inside patterns that already existed.

Four precise changes, nothing more:

DuoChat surfaced in the Section Overview of individual lessons - appearing exactly when a learner has finished building vocabulary and is ready to use it

Notebook added to the Practice screen alongside Speak and Listen - phrase recall as a natural part of the review loop

DuoChat toggle added to Super benefits — discoverable for curious users without being forced on anyone

Navigation bar shortcut added

The best product changes are the ones existing users barely notice - until they can't imagine the app without them. For the people who never touch DuoChat, the app they know still works exactly the way it always did.

The original plan was a standalone DuoChat section - its own screens, its own navigation paths, its own user journey.

A separate destination inside the app.

The pivot was to nest everything inside patterns that already existed.

Four precise changes, nothing more:

DuoChat surfaced in the Section Overview of individual lessons - appearing exactly when a learner has finished building vocabulary and is ready to use it
Notebook added to the Practice screen alongside Speak and Listen - phrase recall as a natural part of the review loop
DuoChat toggle added to Super benefits — discoverable for curious users without being forced on anyone
Navigation bar shortcut added

Side-by-side comparison of the old and redesigned Duolingo Super screen, showing the addition of DuoChat AI as a toggleable benefit alongside updated navigation and practice sections

Before and after screens and navigation bar

4. Practice shouldn't stop when the lesson does

One pattern from research: learners mentally save a phrase during a lesson, then lose it by the time they actually need it. The gap between learning and using is often just a few hours.

The Phrasebook widget closes that gap. Phrases pinned from DuoChat or lessons sit on the home screen — one tap away in the moment they're needed.

It extends Duolingo's value beyond the app session, into the real situations learners are actually preparing for.

Feedback & Validation

Critical

Some dismissed “chatbot” as a lazy add

Constructive

Requests for clearer user flows and real-use examples (GIFs/videos).

Positive

Idea is plausible and useful, people liked the phrasebook + practice concept.

Critical

Some dismissed “chatbot” as a lazy add

Constructive

Requests for clearer user flows and real-use examples (GIFs/videos).

Positive

Idea is plausible and useful, people liked the phrasebook + practice concept.

How I interpreted this

The core concept was validated, but the presentation needed to emphasise action over displaying product images. The mixed reactions appeared to arise from the presentation style and inherent biases (e.g anti Ai or sentiments) , rather than the actual feasibility of the features.

Reflection

This project taught me that designing for an existing product is a different discipline to designing from scratch. The constraint isn't just visual — it's behavioural. Every decision had to account for the habits of people who already use the app daily and weren't asking for anything new.

The most useful shift was reframing the question. Not "what should DuoChat be?" but "where does it already belong?" That single question changed the structure of the whole feature.

A peer-reviewed study of 385 Duolingo learners found that AI conversational practice significantly increased real-world speaking confidence after just one month, which suggests the gap DuoChat is designed to close is both real and closable. That's worth testing properly.

Next Steps

Run moderated usability tests focused on onboarding, phrase saving,To determine the retention and return rates.
Explore a smartwatch prototype.
Explore student modes as future extensions.

Shortly after this project, Duolingo shipped their version.

The gap was real.

Now we find out if learners use it, or go back to the streak.

Decorative illustration accompanying the contact call to action

Let's Work

Together

Let's Work

Together

Let's Work

Together