There’s much hype about the ability of artificial intelligence to raise an organisation’s efficiency. But that is not necessarily so. Technology adoption is not simply a technical process. Stuart Mills and David A Spencer write that if AI is used to carry out what they call “bullshit tasks”, it will lead at most to “efficient inefficiency”. Organisations must look beyond existing tasks and consider how they should change.
There is no certainty that artificial intelligence, by itself, will generate any breakthrough in economic efficiency, despite the significant interest demonstrated by business leaders and policymakers. They believe products like OpenAI’s ChatGPT have great potential to increase productivity and contribute to economic growth. Whether this potential is realised, however, will depend on how AI is used.
In our recent paper, we argue that the hype around the economic effects of AI may be distracting from a key question, namely what kind of tasks the technology will undertake. If it performs inefficient tasks, it will simply lead to more efficient inefficiency – that is, it will compound inefficiency rather than tackle it. This outcome reflects the fact that the adoption of AI is an organisational process rather than just a technical one and that its capacity to raise efficiency is limited by human and social factors within organisations.
Most organisations contain considerable slack and managers at different levels often can create and maintain tasks that, while in their own interests, undermine organisational efficiency. David Graeber wrote famously about the existence of “bullshit jobs” and we would argue that the presence of bullshit tasks within organisations, if carried out by AI, may simply perpetuate inefficiency. AI may just speed up the delivery of what are inefficient tasks to begin with.
Economists argue that inefficiency will be removed by competition. Our view, however, is that inefficiency can persist due to the interests that managers have in maintaining inefficiency. The capacity of managers to tackle efficient inefficiency may also be limited by their bounded rationality. No manager can spot every inefficient task. In practice, most managers may be focused on adopting AI to substitute for existing human tasks, regardless of their efficiency.
We point to several examples of technologies being used to improve the efficiency of inefficient activities. Herbert Simon reports the use of teleprinters in the US State Department. The Department would often be flooded with telegrams, which would be slowly printed out using teleprinters. This slowness meant important information reached key decision-makers hours after it would have been most useful.
The solution the Department adopted was to buy more teleprinters, reducing the lag between the sending of a message and its printing. However, most of the printed messages were never read. The problem was not a technical one (more printers), but an organisational one involving people. Improvement in the effectiveness of the Department depended on organisational changes such as better training, greater delegation of decision-making and the building of stronger trust relations.
As technology companies have sought to develop AI products, the problem of efficient inefficiency is emerging. One study explored how programmers used an AI co-pilot when writing code. The study found that AI increased the amount of code written. However, it also increased the amount of code churn—broken code that programmers had to edit and fix. In this instance, AI might appear to make programmers more productive, since they are writing more code. But factoring in churn, it becomes less clear whether AI is improving efficiency or just doing something inefficient more efficiently (writing more bad code).
The point we would stress is that the efficiency gains from AI are not predestined but instead depend on its uses within organisations. Because of the power that managers wield and the inherent bounds to management rationality, AI will often be used to more efficiently perform inefficient tasks.
In their recent book, Power and Progress, Daron Acemoglu and Simon Johnson cast doubt on the efficiency benefits of AI. They restate the famous ‘Solow’ paradox, whereby economist Robert Solow noted that computers had impacted on so many areas of life, except the productivity statistics. Part of the reason for the paradox is that it takes time—perhaps a generation—for people to acquire the skills to use new technologies effectively. Computing could not increase productivity while people did not know how to turn the computer on.
Efficient inefficiency adds a further component to this productivity puzzle. People also need to learn how technology can be used, beyond the tasks that exist now, and how existing tasks should change. If technology is primarily assessed technically, on the basis of what it can do, rather than organisationally, on the basis of what is currently done, organisations are likely to encounter the problem of efficient inefficiency, and the economic benefits of AI are likely to be much reduced.
Dealing with inefficient efficiency ultimately calls for organisational reform, not just to tackle task inefficiency, but also to challenge management power. Democratic processes and institutions, including within organisations, are vital if AI is to work in ways that enhance incomes and well-being for all in society.
Sign up for our weekly newsletter here.
- This blog post is based on Efficient Inefficiency: Organisational challenges of realising economic gains from AI, in Journal of Business Research.
- The post represents the views of its author(s), not the position of LSE Business Review or the London School of Economics and Political Science.
- Featured image provided by Shutterstock.
- When you leave a comment, you’re agreeing to our Comment Policy.
