Artificial intelligence (AI) is being deployed across domains and creating a wealth of opportunity.But who will benefit from the AI economy’s new sources of wealth and employment? As with past innovations, those with the capital to invest in new technology and the high-skilled workers developing new AI are primed to benefit. However, low-skilled workers within the AI workforce experience a duality of short-term opportunity and long-term challenges.

Growing access to computers is creating a global market for AI services. This is because AI applications are digital and thus easy to copy and transmit (e.g., emailing or downloading software). Nearly anyone can access the AI economy. In fact, the 2019 AI Index Report from Stanford University demonstrates—perhaps surprisingly—that the most hiring for AI-related jobs is not occurring in China or the US, but instead in Singapore, Brazil, Australia, Canada, India, Turkey, and South Africa.

Although AI products are shareable, AI development is costly. AI algorithms are built on cutting-edge mathematical theory, access to massive often-proporietary datasets, and require warehouses of computer servers on which to train the algorithms. Those who can invest in data and computing infrastructure or have the skills required for theoretical AI development clearly have a role in the AI economy, but what opportunities exist for low-skilled workers—especially in developing economies?

Today’s AI applications are typically supervised which means they require data labelled with “ground truth.” This means that the AI economy requires workers that can label data. For example, social media platforms use supervised computer vision algorithms to identify and remove inappropriate images. However, some images make it past the AI and are subsequently flagged by users. These flagged images are then reviewed by a global workforce of moderators, labelled as inappropriate or not, and then fed back into the AI system as labelled data that can improve automatic detection moving forward. As these social media platforms grow, users create and upload more content which increases demand for workers to perform content review.

Labelled data is an opportunity for the global pool of low-skill workers. Just as software is globally transmittable, so too can tech companies transmit data and receive labels from workers anywhere in the world. This work typically requires only routine human abilities (e.g., identify a person’s gender from an image) which, combined with the digital nature of the work, makes it an ideal opportunity for employment in developing economies.

However, the current paradigm of supervised AI may be shifting to a new generation of reinforcement learning algorithms. These AI may be trained on labelled data initially, but eventually learn only based on real-time feedback. One of the most notable examples is DeepMind’s Alpha Go which beat world champions in the game of Go. This game has many possible board positions (e.g., a typical game with 150 moves 10 360 possible moves) which makes it infeasible to label all possible moves given the current board and makes it impractical for an AI to search all possible outcomes systematically. Alpha Go succeeded by repeatedly playing against itself and learning through experience after initially learning from historical games. In fact, the current iteration of this AI, called Alpha Go Zero, is able to outperform the original Alpha Go system without any labelled training data at all.

The rise of reinforcement learning AI may diminish demand for labelled data and, thus, diminish demand for the routine workers who produce labelled data. However, those with investment capital and the high-skilled workers who develop AI still stand to benefit. If low-skilled work is no longer required in the AI economy, then how can global leaders—particularly in developing economies—create decent work for all? AI will always need computing infrastructure on whichto train and knowledge-workers who design new algorithms. Continued relevance in the AI economy requires investments that bring both aspects of AI into the local context. This might seem costly at first but several aspects of AI offer hope.

First, almost all AI-related learning materials are digital and available online for free. Even if local universities are not among the global elites, the academic AI community is strongly committed to open-source knowledge. Github is a website that hosts coding tutorials and machine learning code that anyone can download and implement. The Arxiv hosts AI research reports and is open-source. Online communities, including Stack Overflow and Math Overflow, are open platforms where AI researchers and engineers engage with one another. These free online platforms enable even developing countries that may not have world-class educational institutions to access nearly the entirety of technical AI knowledge. Governments can take advantage of this by bolstering internet infrastructure and encouraging local universities to use these free resources.

Creating the computing infrastructure required to train AI is the greater challenge. World-class computing systems are prohibitively expensive. For example, Deepmind’s Alpha Go Zero ran on 4 Tensor Processing Units and cost around $25 million. However, international partnerships can spread out this upfront investment. The challenge is then for partners to share limited computing resources, but this may be reasonable in the case of AI. Training an AI algorithm typically requires significant upfront computing resources. But the trained algorithm is easy to distribute and subsequent retraining is typically less computationally intense. For example, the auto-correct typing service on most smartphones is an example of pre-trained AI running on the phone even though the phone itself does not contain the computing resources required to train such an algorithm.

Although AI is most strongly associated with Silicon Valley and world-class universities, it also represents global opportunities for decent work and economic growth. However, AI is nascent and evolving. Short-term opportunities may succumb to long-term challenges for those without advanced education in countries without significant computing infrastructure. But the global nature of AI combined with the openness of the academic AI community means that targeted education and partnered investment in infrastructure can stabilize decent work in the AI economy of the future.