When we're talking about AI and it's benefits it's worth thinking about the idea of *substitute goods* and *complimentary goods*. How good a substitute is AI? Might it even be a compliment to language learning? # Definitions ## Substitutes A *substitute* is a good that can be used in place of another. Margarine is a substitute for butter. Record players were a substitute for live music. Automated switchboards were substitutes for switchboard operators. Substitutes can be *perfect* or *imperfect*. Margarine isn't a great substitute for butter in my humble opinion. I keep using real Irish butter on my toast. There is no risk of margarine replacing butter. It's an imperfect substitute. Automated switchboards were a perfect substitute for switchboard operators. In old films we see people saying, "Operator, please connect me to ..." and the operator woman (they were nearly always women) would disconnect one cable and switch it with another. Now a computer does this instantly. The computer was in every way faster and more reliable. The operator women were soon out of a job. Records were a great substitute for live music. Lots of musicians lost work because people would much rather listen to would much rather listen to the Beatles on a record than an amature playing at the local bar. But they weren't a perfect substitute: the Beatles playing live sounds better than the record, and the atmosphere of a concert is better than listening to a record at home. Records actually make superstars like the Beatles possible. You can't reach the superstar level of the Beatles by travelling from pub to pub. A tiny fraction of people can listen to you. Even if songwriters and musicians better than the Beatles existed before the invention of the record player, it was impossible for them to be as wealthy as the Beatles because few people could listen to them. ## Complements A *complement* is something that affects the demand for another good. * If flights to Paris from America become cheaper then demand for French language courses go up: more people can go on holiday in France and see the value in learning French. * Hamburgers are often sold with soft drinks. So if soft drinks become cheaper then it will be easier for business to offer a meal deal and hamburger demand will go up. * Cheap video cameras and Youtube increased the demand for video essays and personal vlogs. * If we jump back to our Beatles example, record sales can be seen as a complement to concert sales. International sales of records create fans of the Beatles all over the world; these fans now want to buy tickets to Beatles concerts. # What does this mean for AI and languages? ## Systems involving technology and humans have hard limits of efficency In physics class in school we calculated the stopping distance of cars at different speeds. The stopping distance is the distance required from the vehicle in front in order to stop safely. If a car is travelling at 90 miles per hour it will need to leave a much larger gap from the car in front than if it is travelling at 30 miles per hour. If a lorry is travelling at 30 miles per hour it will need to leave more space than a car travelling at the same speed, becaue it take it longer to stop. The three components of this equation are the speed of the car, the time it takes the breaks to slow down the car, and the reaction speed of the driver. The interesting thing about this equation is that initially better breaks improve the stopping distace massively, as the car stops much closer to the point the driver notices there is a problem, but soon the reaction time of the driver becomes a bigger and bigger part of the stopping distance. Even if the breaks are able to stop the car immediatly some distance is required to account for the driver's reaction time. The only way that stopping distance can be improved at this point is to take over the decision to break from the person. If an AI could work out when to break and we had perfect breaks then breaking distance could be zero. With translation an idea has to be: 1. Explained by a person in language one, 2. Translated into language two (whether by a human or machine) and; 3. Understood by the other person who speaks language two. It's that last step that is the hard limit of translation. Translations are always trying to point at the same thoughts using different words. But, except for the most trivial cases, you can't just translate one word for another. A translation will always have to make a compromise between style and meaning; it will always have to make allowances for cultural differences (choosing to ignore them or include a lengthy gloss of their meaning); and some ideas are simply expressed in a shorter or longer way than other languages. So, with any translation is going to take time for the other person to read or listen to the translation. If the translator goes for a more literal translation then it will take longer to say; if they go for a shorter translation then they will lose the sense of what the person said. At all times they will be limited by the person they are translating for, what they know of language one and how good they are at understanding meaning. This is true of even a superhuman AI translator. AI can *never* be a perfect substitute for learning a language. ## But what about an imperfect subsitute? AI might make dabbling in the language may become less valuable as you might as well use AI translation. In many ways this has already happened. In 1995 if you needed to understand the instructions for your TV in your Paris Hotel room you could use basic French to work things out. In 2023 you would just use Google Translate. Beginners French has been almost perfectly *substituted* by Google Translate. Many people travel to Japan knowing no Japanease and they're able to get by. But true mastery may become much more valuable. Just as record sales complemented concert tickets, imperfect translation may drive demand for perfect translation. Cross-language cultural and economic exchange will become more common. People will see the value of goods in other languages. People will have more friends and work collegues with people who speak other langugage. Through this exchange people will use AI more and more and see the limits of machine translation. They'll then see the benefits of having someone with true fluency. Imagine you're a small Japanese company and AI tech makes phoning a client in Phoenix, Arizona to talk to them about your widgets as easy as talking to someone in Tokyo. You get a ton of sales. Then three years later you've got a large amount of American business. However, you see the limits of the machine translation. You lose clients because the translation isn't quite right. You don't understand the culture perfectly. So, you try to hire a Japanese Marketer with perfect English. In this case, AI translation becomes a *complement* to demand for language learning.