I’ve lived in Germany for around a decade now, and an enduring source of shame for me is my rather underwhelming German language skills. Like many language learners, I’ve reached the point where I can understand what’s said to me, make casual conversation, and handle everyday situations, but I’ve never really achieved the level of comfort with the language and the breadth of vocabulary that’s required for true fluency.
I’d love to improve my language skills, but I’ve struggled to commit the time and resources for intensive classes. But perhaps there’s a solution hiding in plain sight on my phone.
Perhaps AI can help.
Enter Google
Google recently announced a collection of three “bite-sized learning experiments” built using its Gemini AI. The experiments are publicly available, free to use, work in a web browser, and currently cover a modest range of languages — including, fortunately for me, German.
I was curious how the experience of using these experiments would differ from interacting with Google Translate. Like practically everyone else I know, I make extensive use of Translate. I especially appreciate the camera feature, which lets you snap a photo of a document such as a letter and see an instant, good-enough translation. I wouldn’t rely on these translations for formal documents or for work, but they are extremely useful for quickly getting the gist of long text passages.
The first experiment I tried, called Tiny Lesson, lets you input an activity such as “visiting a museum” or “going swimming.” It then shows you a number of words and phrases related to that activity, with the option to hear them pronounced as well. There’s also a tips section which gives you some pointers on grammar and sentence construction.
The results I got were fairly basic, more suited to language beginners, but everything was accurate and relevant. It’s nothing you couldn’t find using Google Translate — or, for than matter, and old fashioned paper phrasebook — but it’s convenient, and I can see it being useful for tourists who want to look up a few key words and phrases before heading out to an activity like visiting a restaurant, museum, or beach.
Getting the hang of slang
The second of Google’s experiments, and the one I was most excited to try, is called Slang Hang. It’s supposed to teach you about the more casual way that people speak to each other in real-life conversations, rather than the formal kind of interactions typically taught in language textbooks. The idea is to “generate a realistic conversation between native speakers and let users learn from it,” Google says, and it sounded like just what I needed to bolster my conversational skills in a practical way.
The concept is good, but the reality is strange, to say the least. The lesson generated a few plausible scenarios for me, like two people meeting in a book shop or acquaintances bumping in to each other in a cafe. You can see simulated dialogue between two people, and click the sound icon to hear the text spoken out loud. Unusual slang words or phrases are underlined, and you can click on them for more information and explanation. In the first few scenarios I tried I saw a mix of phrases I didn’t know and ones that I’ve heard on the street, which seems useful.
Then the experiment decided to show me a bizarre and confusing exchange that was supposed to be between two strangers in a park who are confronted by a pink pigeon. Puzzled by this, I asked a German friend for his take on it. He acknowledged that the language is technically correct, but said that there was something about the exchange that didn’t feel right, and that it read as stiff and stilted.
He also mentioned that one of the slang words in the dialogue — der Knaller — sounds like something his dad would say, which would be out of place among the more casual and younger-skewing slang that is typical of our city. This points to one of the biggest stumbling blocks for this kind of AI language teaching: whether or not a slang phrase is appropriate for a given situation depends a lot on the social context. Using slang correctly requires fairly sophisticated understanding of social cues, which AI obviously doesn’t have.
If you’re an English native speaker, you can get an idea of how odd some of this dialogue is by looking at the AI generated Slang Hangs intended to teach people English. I tried this out of curiosity and was presented with this slightly off kilter dialogue:
Is this technically correct? Yup. Is it using recognizable English slang? Sure is. Is this a conversation that real two human beings might actually have at any point in time? Ehhhhh. Maybe not.
Point and shoot
I imagine the most popular of the three experiments will be the final one, called Word Cam. You use it to take a photo with your phone camera, and it shows you the words for the objects around you, like a combination of Google Lens and Google Translate.
I was impressed by how well this worked, even identifying somewhat esoteric objects like specific pieces of gym equipment. If you click on one of the words presented, the app takes you to a screen showing examples of the word in use in full sentences. Here, again, I was impressed by the relevance and usefulness of the sentences, which were tailored to the specific kinds of things you might want to say about a particular object.
I’m not sure that this experiment will be useful, exactly — I can’t imagine too many situations where I’d need to know a word and would rather take a photo than enter an equivalent English term into Google Translate — but it is neat, and it does work well. It’s certainly fun to play with.
Am I speaking German yet?
The three experiments were entertaining to try out, and I can see some scenarios where they might be helpful, particularly if you’re brand new to a language and are trying to dip your toes into the water. But I wish there was more depth.
If you’re going to make full use of AI, I can envision a similar app which combines hyperlocal information about your particular region to teach you more relevant words. In most of Germany, for example, the word for “I,” “ich,” is pronounced something like “ish.” But in Berlin, where I live, people say it more like “ick.” It would be cool for a language app to recognize and share that information. Google’s app offers some localization for different slang in Germany, Austria, and Switzerland, but nowhere near this kind of fine-grained nuance of pronunciation.
I’d like to see more social information integrated too. I was testing during Germany’s Labor Day, which is famous for two things: rambunctious demonstrations, and all the shops being shut. A truly intelligent language app could give you information about current events or social traditions as part of the process of learning — and remind you to pick up groceries a day early, which I forget about every year. Without this deeper context, I’m not sure that the AI is adding much to the learning experience.
And there’s the biggest drawback of all to these kinds of apps, which is that listening and reading in a foreign language is one useful way to learn, but to really grasp a language you need to get out and talk to people. You need to hear real conversations, to fumble your words, to make mistakes, and then find a way to make yourself understood anyway. All the AI in the world can’t substitute for some genuine human interaction.