The conversation around AI has spread from labs and companies to mainstream media, coffee shops, and street corners. Everyone everywhere has been talking about artificial intelligence. Recently a friend told me how she overheard a group of men talking broadly about the threat of artificial intelligence when one of them declared in all seriousness, “I want to punch AI in the face!”
Needless to say this would not be possible, since AI has no face, and yet it is easy to understand this response to the rise of machine intelligence. The destructive potential of AI is alarming. It can be misused to perpetuate bias, destabilize political systems, and promote inequities. Some experts believe it may threaten our dominance as a species. As humans we are naturally afraid of entities that could be as smart as or perhaps even smarter than we are. And while there are many potential sources of anxiety regarding AI, I suspect that a significant source of the public fear is economic.
We don’t want AI to take our jobs.
The influential futurist Roy Amara noted that as a society we often overestimate how technology will change the world in the short run and undersell its effect in the long run. This idea has come to be recognized as Amara’s law, and it is especially relevant when we look at the impact of AI on the labor market. The long-term impact of automation on job loss is extremely difficult to predict, but we do know that AI does not automate jobs. AI and machine learning automate tasks—and not every task, either. Certain tasks that make up a job may be ideal for automation, while a whole set of additional tasks may not. So, in most cases, the impact of AI on jobs will not be a one-to-one relationship, wherein a single AI system or even a group of task-specific AI systems are enlisted to eliminate a job entirely. Yet AI will certainly change many jobs.
Consider how automated coding tools are impacting the software development industry. The startup founder and computer scientist Matt Welsh, in a presentation to the Association for Computing Machinery, compared the arrival of these technologies to an alien spaceship landing suddenly in our backyard. Still, he noted that he was not laying off programmers at his company and replacing them with AI. Instead, he was all but demanding his team use the technology because, in his estimate, the tool made them 30–40% more productive. My MIT colleague Armando Solar-Lezama has suggested that the wider embrace of AI coding tools will put more emphasis on high-level awareness and structure of the codebase. While AI tools plant certain trees, we will need smart, educated people to think about the forest. These experts might need to adjust those plantings, too, since the tools make mistakes.
One of the questions that naturally arises from these sort of productivity gains is who will enjoy the benefits. Will these more productive employees earn higher wages because they are turning out quality work at a higher clip? Or will companies reduce their workforce and turn those efficiency gains into corporate profits? On one hand, there’s the optimistic view: increased productivity translates into higher wages for workers who can produce higher-quality work in less time, enhancing their value in the marketplace. This scenario assumes a fair redistribution of the gains from increased efficiency, recognizing the enhanced skillset and output of the workers. On the other hand, efficiency gains could lead to workforce reductions and cost-cutting measures, with the primary benefits accruing to the corporation.
Typically, researchers explore industry trends at an economy-wide scale or look at the largest companies, but small- and medium-sized enterprises (SME) will be impacted as well. While large businesses will be able to explore and invest in AI, smaller companies may not, which could widen the competitive divide and concentrate power further with the largest organizations. Yet there are some helpful solutions being developed. A working group at Global Partnerships in AI has introduced a portal designed to educate business owners on how they might benefit from intelligent tools and what AI capabilities and services are available for their industry. Companies will need to adapt, but how exactly will they work AI into their operations?
Consider for a moment a small interior design firm, which offers a good example of the potential division of tasks between humans and AI. The essence of design, which is rooted in creativity, empathy, and subjective judgment, will remain a distinctly human endeavor. But interior designers can take advantage of advanced technologies like AI generation, virtual reality, and data-driven analyses, freeing up their time by automating routine tasks while retaining the higher-level creative, strategic, problem-solving tasks for themselves. Generative AI could be used to produce design renditions, but the results will be the average of the world’s knowledge, not something new and fresh. So, let’s look at which tasks in the design workflow could be automated and which ones will remain distinctly human.
Tasks that are technically possible to automate:
- Mood board creation: AI can scan through vast databases of images and styles to create mood boards based on specific keywords or themes.
- Space measurements: With the use of advanced sensors and AI-driven tools, precise measuring of a space can be automated, reducing human error.
- Material and furniture sourcing: AI can search through online catalogs, databases, and inventory lists to find materials or furniture pieces that match a specific design, price point, or theme.
- Layout optimization: Given the dimensions of a space and furniture, AI can suggest optimal layouts.
- 3D visualization: AI tools can quickly render 3D models of design concepts.
- Lighting analysis: AI can suggest optimal lighting setups based on the room’s purpose, size, and natural light availability.
- Color matching: AI can provide color palette suggestions based on a primary color or mood input.
- Trend analysis: By analyzing online data, AI can identify emerging design trends.
- Inventory and order management: For designers who handle purchasing, AI can track inventory, reorder materials, and even predict future inventory needs based on trends.
- Feedback collection: Post-design, AI can automate the process of collecting and analyzing feedback from clients.
Tasks that are technically difficult to automate:
- Client interactions: While AI can assist with the design process, building and maintaining a client-designer relationship is inherently human. Understanding a client’s nuanced preferences, emotions, and vision requires in-depth personal discussions and deep human intuition.
- Conceptual design: The initial conceptual phase, wherein designers ideate and brainstorm, is rooted in creativity and intuition. While AI can provide data-driven insights and assist in the process with mood boards and rapid ideation or discovery of samples, the spark of originality is human.
- Cultural and contextual sensitivity: Designs often need to resonate with a client’s culture, history, or personal experiences. AI might not fully grasp these nuances. You need broad human knowledge and intelligence here.
- Ethical and sustainable choices: Making ethical decisions, like choosing sustainable materials or considering the socioeconomic implications of design choices, also requires a human touch.
- Problem-solving: Unique challenges can arise in any project. An experienced designer’s ability to troubleshoot and find innovative solutions is not easily replicated by AI.
- Aesthetic judgment: While AI can recognize patterns, the subjective appreciation of beauty and style is inherently human.
This is merely one example, in one niche industry, but it demonstrates how jobs are the sum of diverse and often complex tasks. So, if we back up and out and consider the larger economy, what other tasks could be swept up by AI? And how do we go about identifying them and understanding our own risk? In a 2017 study of the potential economic impact of automation, Erik Brynjolfsson and Tom Mitchell suggest breaking the question down into tasks that are suitable for machine learning and AI and those that are not. The demand for jobs that meet the former qualifications might fall as AI solutions develop, but this may be balanced by an increase in demand for jobs that cannot be done by AI. Furthermore, jobs in which people can use AI as an assistant, such as writing and programming tasks, may become more valued due to the associated productivity gains.
Erik and Tom provide a useful set of criteria for identifying tasks that may be suitable for machine learning automation. A task which has a large associated dataset is a good candidate, because an abundance of quality data will allow an AI solution to learn effectively. If common sense or knowledge of the physical world is required, however, then machine learning will not be ideal. Given that large network models are black boxes, AI will probably not be a viable option if you need your automated assistant to explain why it made a specific choice. They note that any task under consideration for automation also needs to have clearly defined goals and metrics—a quality often missing in tasks performed by knowledge workers, especially in the hypothetical case of the interior design specialist discussed above. Across various design domains, clients rarely know exactly what they want.
Finally, Erik and Tom add that any tasks requiring human-level dexterity and specialized physical skills will remain safely in the domain of people for the foreseeable future. Skilled workers who use both their heads and their hands, such as plumbers, electricians, and carpenters, need not worry about automation in the slightest. Robots will not be rewiring your home anytime soon.
Yet there is significant change on the horizon. How soon will these changes start to take place? There is no clear timeline, and the change could be slower than expected. My MIT CSAIL colleague Neil Thompson has studied this question and found that widespread, rapid adoption is unlikely for a very simple reason: it is expensive. The fact that a task can be automated by an AI solution that perfectly matches it does not necessarily mean that the AI solution will be adopted—especially not if qualified people are cheaper.
As an example, Neil looks at the job of baking bread. The baker’s work can be divided into multiple tasks, including mixing ingredients, kneading dough, and inspecting the results. Currently, automating the entire job would not be reasonable. I know this firsthand, as my students and I built an intelligent machine called BakeBot that can make cookies. Our robot was able to complete the different tasks successfully and without human help, but it would cost more than a luxury automobile.
Neil wisely set aside the task of preparing and kneading the dough, determining that these tasks will remain the purview of the baker. Inspection is another matter. AI has proven its ability to uncover hidden patterns in a wide range of applications, including quality control for manufacturing, medical image analysis for diagnosis, and food inspection to determine compliance with safety and quality standards. Inspecting loaves is absolutely within its range of capabilities. So, Neil looked at what a small bakery would gain by installing an AI-enhanced computer vision system to take on that task. He estimated that an operation employing six bakers earning a mid-five-figure salary would end up saving $14,000 per year by deploying an AI inspection system. Yet implementing the system might cost more than $1.7 million, and maintaining it would run to nearly a quarter of a million dollars annually. That little bakery would have to sell quite a few baguettes.
The specifics will vary depending on the industry, so Neil designed a helpful rubric for estimating and comparing the cost of automating a task versus leaving it to humans. The cost of automating a task with AI involves:
- Fixed costs, including engineering-related work such as implementation and maintenance
- Performance-dependent costs, including training and retraining the model and any associated tools
- Scale-dependent costs, which center on the compute costs required to actually operate the new system
The total expenses involved in deploying AI add up, and Neil has found that the overlap between tasks that could theoretically be automated and the tasks that are economically suited to automation is smaller than one might think. He has projected that the number of jobs at risk is only a small fraction of the more common estimates, which neglect to account for the detailed economics of automation. His research also reveals that if the costs associated with AI deployment drop quickly, automation will accelerate, but if cost improvements happen slowly, which is more likely, then automation will be gradual and take place on the scale of decades, not months. The fact that it can be done does not mean it will be done.
New technologies undoubtedly disrupt existing jobs, but they also create entirely new industries and the new roles needed to support them. The MIT economist David Autor led a study showing that 85% of employment growth in the last eighty years has been driven by technology. The same study revealed that 60% of workers today have jobs that didn’t exist in 1940. I suggest focusing less on whether AI will steal our jobs and more on how it will change our jobs today, tomorrow, and in the years ahead, and how you can familiarize yourself with these technologies and educate yourself to put them to work for you. If you are a young person or in the relatively early stages of your career, it would be prudent to begin exploring the technologies that are relevant to your field, trade, or role, and finding opportunities for upskilling or reskilling so you can capitalize on these changes. And if you own or run a business, you might want to think differently about how your organization operates and how you are going to make use of these new technologies.
Excerpted with permission from The Mind’s Mirror: Risk and Reward in the Age of AI by Daniela Rus and Gregory Mone.