Social and ethical impacts of any technology presuppose its diffusion. That is why we need to understand the mechanisms underlying the spread of AI, and it is why one of our current research projects concerns AI diffusion in society. This post summarizes a recent study of ours in this project.
The AI-breakthrough has been an effect of better access to data, greater computational power, and increased connectedness. The ratio of scientific papers to inventions based on AI decreased from 8:1 in 2010 to 3:1 six years later, which shows that AI-research is increasingly being translated into commercial products. Already 2016, it was reported in Harvard Business Review that AI interest came mainly from Fortune 500 executives rather than startup founders and academics. A recent blog post addressed AI interest in different regions of the world.
A large part of the economic value of AI is driven by deep learning on structured data, and this type of AI is generally not visual or intuitive, and it includes, for example, algorithms that predict which products to recommend, according to Andrew Ng, former head of Baidu AI Group and Google Brain. Thus, it seems AI-applications have often been merged with other technologies to enhance some of their functions. An example is the introduction of machine learning algorithms to rank searches in Google’s search engine. This may explain why awareness of AI usage is relatively low; for example, a study based on a survey among 1,426 consumers in the U.S., U.K., Ireland, Germany, Mexico and Columbia found that a majority of the respondents, 63%, used AI without knowing it. Apparently, many consumers did not know that search engines and voice assistants use AI. These finding motivated the current study.
Has infusion been an important mechanism underlying the diffusion of consumer AI in society up to now?
The meaning of infusion in the Cambridge Dictionary is: “the act of adding one thing to another to make it stronger or better”. With this as a starting point, we propose the following definition of AI-infusion: the act of adding an AI-application to another technology with a separate function to make it stronger or better.
Thus, an infused AI has been added to another application that provides value to its users without the AI. Consequently, the presence of an infused AI-application is not obvious to many of its users, and it cannot be accessed without the technology it has been infused in.
Hence, AI may be infused in software, websites, or smartphone applications (apps) in their first-time installations or in one of their updates. An instance of AI-infusion was when YouTube began using algorithms based on deep neural networks for personalized recommendations in 2016. Another example includes the introduction of AI-driven design recommendations for slides in MS PowerPoint.
In contrast, a standalone AI is a product or a service that is fully dependent on AI to function at all, and which would be without any value to its users without this AI. A smart-speaker device, such as Amazon Echo, is a standalone AI by this definition, as it is dependent on AI for its entire purpose; it uses automatic speech recognition to transcribe the user’s speech to text, as well as natural language processing to understand and interpret the meaning of that text.
We investigated the scale of AI-infusion in two ways: first by identifying and classifying the presence of AI in the most common smartphone apps, and then by categorizing common consumer AI. We gave special attention to mobile apps because of the massive global use of the Internet and smartphones: half the world uses the Internet, and a fourth connects to it with their mobile only.
The average user in mature markets, such as the US and Canada, spends nearly three hours per day in mobile apps. Investigating apps thus gave us an understanding of consumer exposure to infused AI, while studying common consumer products gave us an idea of the proportion of AI that has spread by infusion.
To examine smartphone apps, we considered the fifteen apps with the highest mobile reach in the U.S. in May 2019 with usage data from Statista. Further, to study infusion in common consumer AI, we evaluated a list of everyday examples of AI-technologies put together by Emerj Artificial Intelligence Research, in which each application was categorized as either Infused or Standalone.
AI was present and had been infused in all of the fifteen apps that we studied (Table 1). Each of these apps would provide a function to its users without the AI. The infused AIs include recommendation engines, spam filters, and chatbots, for example.
|Rank||App||US mobile reach, May 2019 [%]||World installs [bn]||Has AI?||Infused?||Example of infused AI|
|3||Google Search||65.6||5||Y||Y||Search rankings|
|4||Google Maps||64.5||5||Y||Y||Predict arrival times|
|8||Google Play||52.4||5||Y||Y||Detect harmful apps|
|9||Amazon Mobile||47.5||0.1||Y||Y||Fashion search|
|11||Google Drive||38.5||1||Y||Y||Find relevant files|
|12||Pandora Radio||36.6||0.1||Y||Y||Music track recommendations|
|13||Google Photos||36.2||1||Y||Y||Image classification|
|14||Spotify||30.1||0.5||Y||Y||Music track recommendations|
Table 1. AI presence and infusion in the fifteen mobile apps in the U.S. with highest mobile reach in May 2019, obtained from Statista. (Y=Yes)
These apps have massive diffusion: each of them reached at least 28% of mobile users in the U.S. in May 2019 (Table 1). With 257 million smartphone users, this indicates that each of these apps reaches at least 70 million people in the U.S. Further, the number of global installations for these apps ranges from 100 million for Amazon Mobile, to five billion for YouTube on Android phones. The sheer magnitude of these numbers indicates that infusion of AI-algorithms in prevalent websites and smartphone applications is a dominant mechanism for diffusion of consumer AI globally.
In the investigation of common AI technology, the results showed that AI had been infused in two thirds (10 of 15) of the studied cases (Table 2). (Note that four of the AIs listed by Emerj Artifical Intelligence Research were excluded because they were categorized as enterprise rather than consumer AI.) For example, infused AI included algorithms for ranking content, object identification in images, and autosuggestion of emojis. This result thus shows that infusion is a very important process for diffusion of consumer AI. In fact, so far it is the process by which AI has reached a large part of the world’s population.
|AI-technology||Infused or Standalone AI?||Example of Infused AI function|
|Navigation apps||I||Predict commutes|
|Ridesharing apps||I||Optimal pick-up locations|
|Spam filters||I||Spam detection|
|Smart email categorization||I||Email classification|
|Mobile check deposits||S||–|
|I||Object identification in images|
|I||Auto-suggestion of emojis|
|Smart Personal Assistants||S||–|
Table 2. Day-to-day AI technology as listed by Emerj Artifical Intelligence Research (I=Infused; S=Standalone)
The presence of AI in society is ubiquitous and ambiguous at the same time. It is ubiquitous because so many of us are using AI on a daily basis, and it is ambiguous because its diffusion and functioning are hard to grasp to us users. It is not obvious how or when we are using AI. This is a problem, because we need to understand what is happening around us and to us.
In this study, we have introduced the concept of AI diffusion by infusion, which occurs when an AI-application is added to an existing technology that can function without the AI. This concept may capture some of the essential differences between AI’s diffusion and the diffusion of previous technologies, such as the Internet.
In summary, the results indicate that AI-infusion has been a dominant mechanism for the spread of consumer AI. Infused AI includes, for example, online recommender systems, spam filters, and auto-generated thumbnails in video streaming services. We recommend that future studies evaluate the social and ethical implications of AI diffusion by infusion.