The speed at which software and artificial intelligence (AI) is maturing is awe-inducing. One of my favorite examples of real-world AI is Tesla Autopilot. It’s an advanced driver-assistance system that has lane centering, adaptive cruise control, self-parking, automatic lane change and even the ability to navigate autonomously on limited access freeways. But Tesla also recommends that drivers keep their hands on the wheel when using these powerful automation features.
As Tesla customers hit the pavement worldwide, the Autopilot technology is learning, and the team at Tesla is developing software and hardware to realize a future where humans do not actually have to drive because the computer will know more than we do. In fact, Tesla predicts that once Autopilot is proven to be 200% safer than a human, a human’s input would increase risk to drivers and pedestrians. So, from an ethical perspective, one could argue that it makes sense to use Autopilot today with a human-in-the-loop approach, and once it is proven to be safer than a human, we should be ready to let the computer do the driving. My prediction is that we will look back in wonder that we let humans drive cars, much as we look back in disbelief at the use of leeches in medicine.
When it comes to translation and localization, machine translation (MT) is one of the foremost innovations that leverages AI. I’m curious to explore the ethics of MT because we are seeing such a seismic shift in how we think about AI in translation today. This has been the primary focus of interest for many practitioners and conferences in the industry for the past few years, and many companies have already started using MT, or at least begun testing it. But this is the decade where MT will become commonplace across the enterprise.
There have been major advancements and investments in MT by massive companies like Google, and growing companies like DeepL. While MT has been used quite liberally by the general public to get the gist of what a bed and breakfast offers in a foreign country, it’s now becoming more commonplace by companies that might typically engage in human translation.
The most significant concern about MT is that it can be woefully inaccurate. MT is predicated on the data and content that’s available, and in some cases, the content and data is inaccurate, incomplete and even gender-biased.
So there are two questions that come to mind when considering the ethics of MT. First: is the MT model being designed empathically? And second: how do people and companies use MT?
Design thinking and MT
Design thinking is a process for creative problem solving with the human experience at its core. It has three essential components:
Empathy — understanding the needs of those you’re designing for.
Ideation — generating a lot of ideas. Brainstorming is one technique, but there are many others.
Experimentation — testing those ideas with prototyping.
The results of design thinking make answers seem obvious and natural, but the reality is that the process behind design thinking enables the user experience to feel simple. As an example, have you ever heard of PillPack? It’s a medical prescription home-delivery solution that does everything from managing refills to coordinating insurance payments for their patients. They worked with Ideo to develop a brand strategy and product that sought to fully understand and embrace the customer experience. The business sold to Amazon in 2018 for a cool $1 billion.
Another company, Airbnb, was famously struggling to grow. They came to realize listings without photos weren’t being booked. So, the Airbnb team started photographing listings in New York City with their customers. It wasn’t long before they realized that photos offered their consumers the requisite validation for Airbnb hosts to successfully list their homes.
Design thinking has its place in MT, too. And that’s where ethics comes into the picture.
When we think about MT, our perspective is that we have to input more data than one might imagine to build a baseline model. Then we use statistical data analysis to measure the accuracy of the target translations. It’s at this point where empathy, ideation and experimentation come into play. This is also the point at which companies that develop MT models can establish a differentiated product.
That’s because a human has to be looped in during this process to make decisions that impact the long-term development of the MT model. Statistical outliers have to be removed from the engine, reviewed by a human and edited by a human where necessary. The widely recognized issues of inaccurate, incomplete or biased data can negatively impact a person’s experience with content if outliers and at-risk translations aren’t addressed by a human.
Of course, new technology always introduces risk. Developers and product managers who leverage design thinking must actually put themselves in the shoes of the consumer to understand how it would feel to be wronged by the very product being developed. This is no trivial task! Leaning on the human-in-the-loop methodology lends itself to building a better MT model and a competitive advantage in a crowded field.
How companies and people use MT
For decades, consumers have relied on MT engines to achieve a basic understanding for a worldwide product, service or information. Sometimes it works great. Other times the translations are questionable, or even downright dangerous. I’m sure you wouldn’t bet your life, or someone else’s, on the accuracy of MT if you didn’t understand the language.
This analogy can be used as a guiding principle for people and companies considering MT for their content today. If the translation is mission-critical or can put a life at risk, it’s worth investing in human translation.
That’s why Translators without Borders enlists human translators and interpreters to realize humanitarian communications worldwide. They’ve translated over 83 million words. Some quick math here: 83 million words times six characters (average number of characters per word) is about 500 million characters. That’s about $10,000 in MT — a lot less than the actual value of the professional translations donated by Translators without Borders’ 30,000 human translators!
But there’s a good reason for this: people’s lives are on the line. Some are seeking citizenship, others are facing life-or-death health humanitarian crises. It would not be ethical to depend on MT when the probability of an error introduced is significantly higher than if a human translator is assigned to the case. It’s not worth the toll on human life.
Another example: when the deadly Covid-19 (Corona-virus) spread in Wuhan, China, foreign governments donated medical equipment and documented procedures to help contain the epidemic. As reported by Global Times, a Chinese publication that is translated into English, volunteer translators chipped in to translating a word count of 272,600 characters (this is about 409,000 words in the English language). It would have been faster and less expensive to machine translate the documents and product descriptions, but it would not have yielded the intended results. Lives were on the line, and using MT for this work would have been unethical.
MT is a perfectly suitable solution for a lot of different content types, and it will become better for all content types in time. Companies today frequently use MT for lower-value content: certain help content, internal communications, low-priority texts that are infrequently read by consumers. If someone’s life isn’t at risk, the decision to use MT is largely a business calculation. How much risk are you willing to take for the potential reward?
How we think about technology, artificial intelligence and machine translation in our industry is a shared responsibility, just as Tesla and its owners consider its Autopilot a substantial innovation and achievement, but that still requires a human-in-the-loop approach.
New technology always brings with it unknown ethical challenges. We can be mindful of these challenges in two ways. We can engineer these products with a design-thinking approach: consider the end-user, and work with a human translator to actively improve the language learning model. Additionally, we can consider how and when MT should not be used: if the end-user’s life depends on the content, do not depend on MT.