Emerging Technologies as a Cause of Shadow Work

In “Shadow Work: The Unpaid, Unseen Jobs That Fill Your Day,” Craig Lambert argues that the upsurge in unpaid tasks (shadow work) that we do on a daily basis increasingly erodes our free time while eliminating jobs and reducing the cost of products & services.  Today, it is we who fill the tanks of our cars, bag our groceries, make travel reservations, drive our kids to school, and manage most personal banking operations and a myriad of other activities. Thirty years ago, paid professionals did it.

In many instances, we prefer and choose to do shadow work: by making our own vacation arrangements, we feel that we arrive at the best equilibrium between the value we get and the budgets we can afford; by using software to do our own taxes, we increase our chance to file them correctly at a lower cost than that afforded by a tax professional.  We pay a price in terms of time, but we feel that that price is justified by the value we get in return.  However, in other instances, doing shadow work is neither a preference nor a choice; taking, for example, additional responsibilities for the same compensation after a company goes through a downsizing amounts, in most instances, to doing unpaid work.

 In addition to the democratization of expertise, the information dragnet, and constantly evolving social norms, Lambert identifies technology and robotics as one of the four major forces that explains the upsurge in shadow work. But in my reading, Lambert seems to primarily emphasize the role that mature technologies have in enhancing the amount of shadow work we do. What I would like to argue here is that emerging technologies are a more perfidious cause of shadow work.

Let’s discuss two emerging technologies in more detail, in the context of two services: Speech Understanding in customer call support centers and Machine Translation in language translation services. During the last decade, both technologies have improved in leaps and bounds. As a consequence, they are incorporated more and more aggressively into larger solutions.

Nowadays, when we call a vendor to ask for support, we are most likely greeted by a machine that is not smart enough to replace a customer support agent and understand what we are saying, but smart enough to take us through a pre-programmed sequence of choices and sub-choices in an effort to fit our question into a category for which the answer can be provided automatically. The machine cannot understand reliably a question such as “what is the balance on my credit card?” so it takes us through a long list of questions: “if you are calling about your bank account, press 1; if you are calling about your grandmother, press 2; …; if you are calling about your credit card, press 9;” Once we press 9, we get other questions “if your credit card is blue, press 1”, and so on. After we pressed keys or spoke digits for 5 minutes, we may get the answer we are after; be sent back to the original menu to start from scratch because we might have missed saying the right number at the right time; or offered the chance to speak to an operator. The five or ten minutes that we spend on the phone interacting with a call center software system is emerging technology caused shadow work. We do that work because automatic speech understanding technology is not good enough to figure out as quickly and reliably as a human what we actually need.

Doing 5-10 minutes of emerging technology-enabled shadow work every time we make a call is annoying. But doing shadow work every day, as part of our regular job, is more than annoying; it is unfair. Let’s examine the shadow work professional translators do when working on post-editing jobs.

Three decades ago, translation work was easy to compensate fairly. When an organization wanted to translate a document, it counted the number of words in the document, negotiated a translation price with a professional translator and paid for the translation service when the translated document was delivered. Two decades ago, things got a bit more complex. As enterprises started storing their translations into large databases, called translation memories, they were able to identify whether sentences in a document to be translated were translated before. That technology is still used today. 

When, let’s say, Fitbit Charge HR was released in January 2015, the product documentation that came with it had many sentences that were part of the Fitbit Charge product documentation, which was released in November 2014. In preparing for the Fitbit Charge HR release, source sentences that matched 100% sentences that were translated before could be simply retrieved from the translation memory and the translator could be paid only to review the extent to which a translation produced several months back would still fit into the context of the new documentation. And source sentences that had a very high, but not perfect, match with sentences that were already stored in the translation memory could be also translated at a discounted price since the amount of work was lower. If Fitbit had, for example, a Spanish translation for the sentence “Radio transceiver: Bluetooth low energy”, which came with the Fitbit Charge documentation, it would have been reasonable to pay less for translating “Radio transceiver: Bluetooth 4.0”, which came with the Fitbit Charge HR documentation. A professional translator would have needed only to replace the translation of “low energy” with “4.0” in order to obtain a perfect translation.

 As Machine Translation has improved, localization departments and language service providers have started to use it more and more as part of their internal processes. Today, they typically pre-translate a document with an automatic translation engine and ask their translators to post-edit the outputs at a discount. Instead of being paid a full price, professional translators are paid the discounted price for each word in each sentence they translate (see the left side of the leading Figure). The problem is that this type of pricing is not in sync with the reality because Machine Translation systems cannot provide, by design, consistent quality. For some sentences, they produce very good translations, for which applying a discount is justified. But for many sentences, they produce low quality translations that cannot support increases in translation productivity and for which applying a discount is not justified. Today, all deficiencies associated with the adoption of an emerging technology – machine translation – are absorbed by professional translators: they end up doing unpaid, emerging technology caused shadow work (see the right side of the leading Figure for a graphical illustration).

Doing 5-10 minutes of shadow work whenever we call a customer support center is not the end of the world. Some of the customer support cost reductions that are enabled by we cooperating with an emerging technology could, most likely, be passed on to consumers via reduced prices. However, doing daily unpaid shadow work to accommodate an emerging technology is a completely different matter. Nobody could feel right when they spend 15-20% of their workday doing un-compensated work.

I am in the process of compiling a list of cases in which the adoption of an emerging technology creates shadow work. If you want to share such cases from your own industry, please comment on this post or email me directly; your insights are appreciated.