Coral & AI

Our goal is to use LLMs and related tools (just "AI" from here on out) in tactful ways that we actually believe add value to language learning, whether to individual learners or, in the future, groups of learners and teachers.

The Jagged Revolution

AI is revolutionary, but the hype and speed of adoption have muddied the waters about what it's actually good for and how to use it. This is troublesome because AI has "jagged edges" and horrendously fails at some things, while being incredible at others. Further, our lack of skill and discretion when using it can amplify these edges and have led to the delirium of the last 2 years.

It certainly seems useful in many ways, but we truly just don't know about the real results in many applications. Speed vs velocity is real! This is especially true with language learning and it's the reason why there are so many "AI chat apps" that have popped up recently. It's easier than ever to make something using AI tools and this new shiny new frontier is tempting. We might not know if it works, but we know that it sells. "FLUENT IN 3 MONTHS" Instagram ads are making money off of high upfront costs and users who burnout, without really knowing if adding AI is even helpful!

Can AI chat actually simulate conversations in a way that helps people learn? I personally feel a bit divested in such interfaces, but I also feel that way in a classroom doing roleplay. Can dynamic vocabulary or grammar lessons actually make content "stick" with learners? Can fully virtual environments powered by AI keep learners improving over the long term?

The technology is a few years old and that's obviously not enough time for anyone to have these answers beyond the marketing-speak of "revolutionized language learning". Research has shown promising results¹, but they are narrow and it's just early.

We're optimistic, but we don't know either! We use these tools ourselves and some feel good, while others don't.

Strengths & Weaknesses

We view AI as a cheap way to "get reps" and exposure to level-appropriate content that is relevant to you. It is not a replacement for a human-centric learning community/experience. You have no peers nor a teacher. This dilution of community is, to me, AI's biggest weakness. What's the point of learning languages if not to be with people? Our vision of AI is as a companion tool to help enrich other learning, which should include real people.

Here's specifically what we think AI can do well and not-so-well.

The good:

• extending, enriching and varying vocab that you already know or care about
• highlighting well-known, easily understood grammar points, in narrow contexts like messages
• putting you in a low-pressure convo where you can practice
• letting you hear relevant vocab in a target language
• doing all the above on-demand, tailored to you

The bad:

• dynamic conversations always seem to fall apart or feel flat. It is, after all, just making it up
• truly guiding a learner is a mixture of objective measurement and emotional support, not just providing relevant content
• there's a lot of variance in output and "garbage in, garbage out" means the precise setup always matters
• AI answers are confident, even when they're wrong

The models and surrounding tech change fast, so our goal is to take reasonable steps to keep up, while experimenting to learn what actually works.

How Coral Uses AI

The core of our workflow tries to help you build vocabulary consistently by using a variety of methods. We want it to be vocabulary you care about and then help develop comfort actually using it a natural way.

We think that using this vocabulary at the centre of all AI interactions helps to keep the core app experience emotionally resonant. You're driving, while Coral assists in convenient ways to give you more exposure.

Coral uses AI in the following specific ways:

• Generating "focus" conversations and vocab, to help build momentum and exposure
• Suggested elaboration and extension of your vocab
• Conversations
  • AI responses
  • Missions
  • Single-message feedback
  • Response suggestions
• Text-to-speech and speech-to-text throughout the app
• Translations

Nothing is added automatically and content merely acts as a focus for your own studies. By actually creating and practicing vocab, you decide what sticks.

The Future

Coral actually started as just a simple way for me to listen to TTMIK while commuting. I used Google Translate to practice speaking and Anki for flashcards. After combining these, it slowly expanded to what it is now. We want to continue in this spirit by keeping up with AI advancements, while focusing on learner feedback.

The basics, like AI responses and voices, will continue to improve, but we can also embed AI more gracefully into current workflows and make completely new ones.

We imagine keeping the core app as a self-driven toolkit, while adding community-based features, like some crowd-sourcing, human review and maybe even a teacher backend.

Our focus now is improving the tools we have to give the best learning experience possible.

Appendix

1. Some research , just to highlight that it is happening, but still catching up as the tools are so new (search & summaries via some Claud sessions, then I checked abstracts & intros for accuracy):
   • Lyu, Lai & Guo, 2025. International Journal of Applied Linguistics 35(2). Meta-analysis of 31 studies, 41 effect sizes. Chatbots had a medium effect on L2 learning, g = 0.608. Effects varied by whether the chatbot was mobile, its modality, whether generative AI was involved, and what the comparison group got. Abstract does not state which direction those moderators cut.
   • Hou & Min, 2026. ReCALL 38(1). Open access. Three-level meta-analysis of 16 studies, 89 effect sizes, on speaking specifically. Moderate effect, g = 0.61. System type, meaning constraint, and speech versus text output mattered. Intervention length, corrective feedback, proficiency level, and learning location did not. Authors flag generative AI as future work. The addendum only adds a corresponding author and changes no numbers.
   • Wang, Cheung, Neitzel & Chai, 2025. Review of Educational Research 95(4). Meta-analysis of 28 studies, 70 effect sizes. Positive effect on language learning performance versus non-chatbot conditions, g = 0.484. Results varied by educational level, language level, interface design, and interaction capability.
   • Kızıl, Klimova, Pikhart & Parmaxi, 2025. Journal of Computer Assisted Learning 41(2). Systematic review, not a meta-analysis, of 33 studies from 2020 to May 2024. No effect sizes. Speaking and vocabulary showed the clearest gains, plus individualized learning, immediate feedback, and more willingness to communicate. Explicitly calls the research base insufficient, with gaps on age, proficiency level, the teacher's role, and classroom integration.
   • Kittredge et al., 2025. Frontiers in Education. Two pre/post surveys, 385 learners total, one month of using Duolingo's generative AI features. Self-reported confidence rose on six of seven measures in the second study. No control group, no proficiency measured, English speakers learning Spanish or French only. All six authors work at Duolingo, which funded the study.