Preply: BUILDING IMPACT THROUGH RESPONSIBLE AI

Borys Pratsiuk is an Engineering Manager at Preply. His background includes 10+ years of experience in AI and management. Starting as an Android developer, he later held key leadership roles, including Head of R&D at Cliklum and CTO at Scalarr. Now, Borys focuses on leading engineers, driving engineering success and bridging technical gaps to ensure seamless execution across teams.

Sergiy Barbashyn, as the AIEI president, had a great conversation with Borys recently. This discussion is especially timely as nowadays AI implementation is rapid and becoming more and more omnipresent in each industry. 

At the AIEI, we advocate for AI Ethics and Integrity. For us, Ethics is about human-centric values, and Integrity is about the uncompromising quality of technical implementation. We were curious to find out what values are core for Preply and whether they match with the AI Ethics. Borys Pratsiuk gladly shared some relevant information with us, so in this interview, we will dive into his insights and the experience of Preply.

Q – Borys, let’s start with some highlights about Preply. How can we describe the Platform in a few sentences?  

A – Preply is an international platform for private online lessons, founded in 2012, now having offices in Ukraine, the USA, Spain and the UK. The vision is to improve language as a learning experience by connecting students with independent tutors. Preply operates across 180 countries. Also, Preply provides 90+ languages and 163 subjects. There are over 150 000 active learners, 100 000 tutors and 3.5 million lessons per month. And all numbers are growing continuously. 

Q – Thank you. We have checked that there are other players in your market as well, such as Duolingo, Italki and Babbel. May you clarify the key distinctive components that distinguish Preply from these market players?

A – Not all of these companies we consider as direct competitors, as their services significantly differ from what we do. Preply actually helps people speak and unlock language capabilities. Today, Preply is roughly equal in size to or up to 5 times larger than its direct competitors. This growth was driven by an ambitious and bold strategy that integrated AI tools with a human-first approach in early 2023. 

Key innovations include a lesson recap, a self-practice app for vocabulary and a new AI avatar, which will help students to practice daily conversations. Unlike competitors, Preply provides students with a broad functionality even with the free version. And lastly, we prioritize the continuous development of our tutors. Our dedicated department helps tutors to level up their skills through formal certifications and quality approvals. By investing in our teachers’ expertise, we aim to upgrade the quality of lessons and results of every student.

Q – Moving closer to the AI topic, what are Preply’s key AI strategic directions? 

A – One of our main priorities is analyzing data from lessons to accelerate student learning. We treat this with the utmost responsibility, ensuring mandatory user approval. The data belongs to our users and we just save it in a safe way, analyze and use it to make students` learning with Preply faster. 

The second direction focuses on internal ecosystem automation. By integrating AI into coding, communication and other core business processes, we make operations at Preply faster, smarter and more efficient.

Finally, our applied research group works on its own models and leverages data to implement cutting-edge solutions like unique algorithms for analyzing pronunciation mistakes.

Q – Thank you, that is really impressive. Speaking of unique algorithms, our readers would like to know more about “matching algorithms” at Preply. How does the system connect students with the right tutors? 

A – Preply’s powerful matching algorithm analyzes a bunch of parameters to connect students with the ideal tutor. We evaluate how lessons are conducted, focusing on student outcomes and tracking regional patterns and learning habits. Also, we have noticed that if the tutor starts a lesson with a “good morning” or “how are you”, that improves student retention. And if a person is looking for a gentle tutor with good manners, we can easily connect this intelligence to starting a lesson with “hello”, “good morning”, “how are you”. Our system is evolving constantly, ensuring that we respect and leverage each tutor’s unique background.

Also, each tutor at Preply has some skills in specific areas, like finance, law, IT, medicine and some people specifically search for such a tutor for different reasons. It could be for enrolling in university, improving or changing a career, etc. Preply allows you to book just a couple of lessons with a tutor from a certain industry and go deep in the topics of your interest. 

We have a real case at Preply with a Japanese student who worked in a globalization company in Japan. Despite having a limited budget, she was interested in finding out how her colleagues from the US work in the globalization area. So, she found a tutor from Preply who works in a globalization company in the US and uses the sessions to dive into American modern practices. Eventually, she created a manual and following it was really useful in her work. That is not just about language transfer, it’s actually more about knowledge transfer that addresses specific professional or educational gaps. Just for reference, 90% of lessons now are still language learning, but we are expanding into new, different topics, such as public speaking, psychology and many others. We try to maintain diversity and encourage its development by welcoming professionals from many fields and from all over the world to join Preply as tutors.

Q – That’s a wonderful philosophy! You earlier mentioned the AI avatar and other AI tools at Preply. How do you divide tasks between those where human interaction is prioritized and those best optimized by AI? I would like to know what that synergy looks like at Preply. 

A – We were surprised, but when we started doing different pilots, we found out that people choose Preply because it’s a non-AI avatar space. Students like to talk human to human and we make it possible by connecting people. 

For lessons, we prioritize live human interaction, while AI is used for highlighting mistakes that might otherwise go unnoticed.  At Preply, we ensure that AI just copilots with humans, not competes or replaces humans. So we are exploring AI innovations for supplemental practice to help students improve some skills in between the lessons. For instance, with our AI avatar, you can easily practice ordering coffee. With the script, this tool guides the conversation and points you to continue dialogue in the same direction, making sure the student is not distracted by other topics. At Preply, we note that practising small talk with an AI Avatar in between the lessons shows really good results in language learning and in improving pronunciation for our students.

Another way we implement AI is transcription of lessons for the sake of history and as a data source. At Preply, AI can also be used to generate homework based on specific student mistakes. By reducing the time spent on manual exercise creation, we allow tutors to focus on reviewing and approving the most effective content. We believe that while AI provides a variety of tools, only the tutor truly understands what is best for the specific student’s progress.

Q – As we understand, AI can act as a copilot to help tutors and students with tasks like fixing grammar mistakes or summarizing. Nevertheless, how do you maintain a balanced dataset to ensure your AI tools are trained properly? Also, how do you avoid algorithmic bias?

A – We built our own grammar correction model that we use. While developing it, we have followed a truly scientific distribution approach. First of all, this dataset includes all groups of language levels. Then we balance this dataset by gender. So there should be an equal split between men and women. We also take a look at the global distribution. That should not be only people from the United States or from Europe, for example, but they should be randomly picked from all over the world. The same principle we follow about other parameters. 

By maintaining a balanced dataset that represents equality and diversity, we prevent algorithmic bias and ensure high performance in production. If we build an unbalanced dataset, it will not work in production because of wrong results. 

Regularly, we collaborate with OpenAI, which has been our trusted partner for a long time. In addition to our proprietary efforts, we leverage their API too. So we as a company are covering language with a higher coverage to bring experience faster to more and more people.

Q – Though AI can be a nice assistant, it may also hallucinate. Is it applicable to your work in AI implementation? 

A – So, in rare cases, hallucination might happen. However, we aim to avoid it by strictly controlling the main processes. For these needs, we have an internal evaluation framework that is based on manual review data. And the number of data points for evaluation is constantly growing. The special framework that we use is adopted by all teams at Preply, and that’s the key protection from possible hallucinations. However, we are still improving deeper ethical moments and our culture there to ensure the result is sufficient and no hallucinations are observed.

Q – That’s truly reassuring to hear. Moving from technical safety to community safety, Preply unites people for language education purposes. With that in mind, how do you maintain a safe environment to protect both students and tutors from inappropriate behaviour? 

A – Preply maintains a zero-tolerance policy toward harassment in any form. We have special ethical AI detectors that flag improper behaviour, allowing us to ban students’ accounts if their behaviour is improper for certain reasons. With the help of our special pre-selection process, we ensure harassment doesn`t come from the tutor side. At Preply, we feel the importance of guarding the morale of the people that we connect. And we involve AI for this purpose as well.

Upon receiving user consent for audio analysis, our systems can detect signs of harassment directly from voice patterns. Also, if there is, for instance, a 5-minute delay, we can at Preply raise the question to the tutor and ask if they have a problem. This preventive framework operates post-lesson to identify and flag issues automatically. Once a report is generated, it has to be reviewed by our specialized support team and they always have the final decision on what to do. 

We do not have mimics, behaviour, or other visual aspects detection. It would require real-time detection, which is possible only by asking for video consent, which many users are afraid to provide in terms of privacy.

Q – From a privacy standpoint, that makes total sense. Looking at the learning process itself from another angle, do you track student progress at Preply and do you use AI for this purpose? We are especially curious about how you measure the quality of a student’s progress, not just the quantity of lessons taken.

A – We have a dedicated team that is working on progress using some metrics over time. Progress measurement is a really scientific task to do. We are still in progress and are constantly trying to implement new metrics. 

By analyzing how students’ vocabulary changed from simple words to more complex ones,  our AI provides a clear picture of their growth. Preply visualizes this journey through personalized charts, mapping the student’s unique learning path and the results they have achieved during their learning with us. But measuring the progress of other parts of language, like fluency, grammar, listening and reading abilities, is a more complex task, which we are actively working on. 

Q – Well and lastly, what is your point of view on AI regulation and governance in the modern world? What do you think should be on the table: regulation, soft framework or maybe guidance?

A – Preply strictly adheres to legal regulations, ensuring full GDPR compliance and maintaining transparent consent protocols for data recording. Beyond legal requirements, we follow a strong ethical framework. Nowadays, we come to truly philosophical questions: should AI make a decision instead of humans? While, for instance, AI is crucial in identifying and blocking harassment, we believe that critical decisions affecting individuals should remain under human responsibility. For example, we do not support using AI to evaluate language levels for employee dismissal. On the contrary, we believe the purpose of AI is to empower and justify professional growth, not to automate decisions that negatively impact careers. And if we’re doing learning software, it should be used specifically for learning, not for firing people. 

Q – Thanks a lot for this insightful discussion. Let`s summarise with three notes from you on Preply’s approach with AI. 

A – Firstly, at Preply, we are using data for good reasons, specifically to unlock the student`s learning velocity. 

Secondly, we’re using data to protect users, so if something goes wrong, they can rely on Preply.  Preply will always be on your side to protect you from improper behaviour that compromises our terms of use or privacy policy.  For us, it’s crucial to follow all security and regulatory aspects for user protection. 

And the third one, we continuously develop new AI-powered products, aiming to improve our product, learning processes and to mature the company`s technological capabilities. So AI is our big friend and we are using it for good reasons.