For the last decade, the prevailing wisdom has been that advances in Artificial Intelligence will lead to large-scale automation and as a result many jobs will be lost and entire professions are at risk of becoming irrelevant.

I think there is quite a bit more to consider than the above suggests.

Different Types of Automation using AI

I can think of at least three different types of automation made possible by AI:

1. Augmenting the abilities of specialists

AI-driven tools can be very valuable to experts. These tools can: automate repetitive tasks; act virtual assistants or achieve outcomes that were previously prohibitively expensive or complicated.

Github Copilot is a great example of both a virtual assistant and the automation of repetitive tasks: it can help software engineers produce code faster but the responsibility of ensuring that code works as intended is ultimately up to the human using the tool.

AlphaFold, RoseTTAFold and similar models can be used to predict protein structure. Other methods for determining protein structure, such as X-ray Crystallography, are expensive and time-consuming.

This type of automation doesn’t make experts redundant it makes them more effective.

2. Empowering the layperson

Here AI is used to provide tools to allow the layperson to do things that were previously beyond their ability: translate text (Google Translate), write code (ChatGPT), generate images (Stable Diffusion, Midjourney, DALL-E 2 etc).

These tools are transformative in many ways. It can feel magical to behold the output of an algorithm that looks like something an expert made, especially in the early stages of experimenting with a new tool.

Are these outputs fit-for-purpose in such a way that will render the experts who previously produced them redundant? This depends on the intended purpose. When trying out a new tool there often isn’t an explicit application so it is easy to forget that producing something is not quite the same as producing something for a specific purpose.

Translation

Consider translation: translating from one language to another is something that previously required a deep understanding of both languages and of the nuances of the text being translated. It was (relatively) expensive. As a result, only certain types of texts were translated. These would be texts where translation was either necessary or economically viable. In both cases a high standard of translation is required because the costs of a poor translation may are undesirable.

While I don’t have any data to back this up what I expect is that if you could plot the total number of documents (or number of words) translated each year you’d see an increase of many orders of magnitude in the years since the introduction of Google Translate when compared to the years prior. However, if you could see two different groups for this plot: documents (or words) translated by humans and by algorithms that the previously noted increase would almost entirely be made up by algorithmic translation and that there would little, if any, decrease in human translation.

This should be expected. For the types of texts where it was important (specifically where the cost of human translation is low relative to the risks or profits at stake) to have high quality translations that humans are still the preferred translators.

There are so many types of documents where it doesn’t matter that much if the translation is poor or the nuances are lost. Web pages or user generated text are great examples. It would never be economically viable to have humans translate comments on Instagram but it makes good enough sense to do it on demand using an algorithm. If the output is confusing and mediocre hopefully the user will not be too surprised.

The net effect of machine translation is to make it viable to translate text in new contexts rather than rendering experts redundant.

Photography

Stock photography shares some similarities with generated images. Vast libraries allow you to search for professionally executed photographs, ready to use. The specific mode of interaction with DALL-E 2 et al is slightly different but the general dynamic is similar.

So why haven’t stock libraries rendered professional photographers redundant? And why would we expect similar dynamics for generated images?

Both stock photography and generated images can deliver images that fall within certain criteria. You can get some images out of these tools but in many cases what people need are images of specific things, people or places. If you are a business owner and you want to show your specific goods and services. You wouldn’t use stock imagery of a coffee shop to advertise your coffee shop.

Again we see a similar dynamic to translation: in instances where you would previously have paid for the services of a professional you will likely continue to do so because the outcome is more important. Stock libraries can be fantastic for instances where it wouldn’t necessarily make financial sense to use a professional, such as sprucing up a blog post with some eye-catching imagery.

In many cases I see AI tools growing the pie by making it possible to use previously expensive outputs in new contexts.

3. Full automation that doesn’t require a human at all

AI is used to automate tasks at a level of sophistication that eliminates the need for human participation entirely. Spam filtering and labeling objects in images come to mind as examples.

In many instances this type of automation may be enabling something that was previously unfeasible to do at scale using human labour.

Experts remain valuable

The second category and third category are easily confused and often taken to be the same thing. Just because new tools exist that allow laypeople to do some things previously only accessible to experts doesn’t mean that the experts are no longer required. Far from it: many experts will valuable when the task being done is important in some way.

We should be wary of visions of the future that look merely like anautomated version of today.