why it will probably remain just a tool that does inefficient work more efficiently

ChatGPT is a remarkable technological development^[1], capable of writing compelling prose that comes across as natural, coherent and knowledgeable.

But it has its limits, and can be made to say silly things. I managed to get it to say that 450 was larger than 500^[2], and others have made it claim that 1lb of feathers weighs the same as 2lbs of bricks^[3].

ChatGPT also cheats. While sometimes sounding impressive, it will make up citations to give the illusion of academic rigour. And it plagiarises. Ask ChatGPT to suggest some town names for a fantasy story, and suddenly you’ll be in the familiar territory of Tolkien’s Middle Earth.

Yet despite these “flaws”, there is great excitement about what ChatGPT will be able to achieve and what it can produce. In the media sector for example, Buzzfeed is planning^[4] to use ChatGPT to create online quizzes, and the newspaper owner Reach has already published articles^[5] written using the technology.

But as well as excitement, there are also fears – as is often the case with AI developments – that ChatGPT will bring mass redundancy to certain sectors of the economy.

It’s a common divide when scientific advancement is rapid. Whenever a new technology comes along, there is talk of productivity gains and automation^[6] and debate over whether people will be better or worse off.

Some economists^[7] argue that technology increases productivity without threatening mass redundancy because it creates new jobs. But there is never any guarantee^[8] that the new jobs are will be as well paid, secure or fulfilling as the ones which have been lost to automation. Workers have every reason to suffer from “automation anxiety^[9]”.

This perspective also assumes that the jobs being automated were actually necessary jobs. Otherwise, automation does not necessarily mean greater productivity.

The late anthropologist David Graeber’s compelling and controversial^[10] theory of “bullshit jobs”^[11] began to highlight this. His idea was that a large number of (mostly) office jobs are essentially pointless; that even the people doing them feel they contribute very little to society.

So let’s say ChatGPT starts to take on more roles with an organisation – writing invoices, formatting data, organising spreadsheets or compiling those quizzes. If those jobs exist because of bureaucratic inefficiencies, automating such work will not raise productivity – because the work was unproductive to begin with.

Nor does it mean that office work for humans will disappear. Managers will surely have limited interest in replacing the people who work for them with artificial intelligence. Some argue^[12] that top managers revel in managing large teams because it gives them prestige and authority. Many employees can also make firms look more legitimate^[13], which might have strategic benefits.

So white collar work will continue. If anything, it will become more nebulous; there will be more requests for a “quick Zoom calls” or a meeting over coffee. This is because tools like ChatGPT will be able to do the administrative work of these workers (like drafting an invoice), but making these workers redundant will not necessarily benefit their bosses.

But the biggest obstacle for ChatGPT, in terms of its impact on our places of work, can be gleaned from the thoughts of the management systems expert Stafford Beer^[14]. He argued that it is better to “dissolve problems than to solve them”. Put simply, he saw that well designed computer systems anticipate problems and dissolve them at the outset. Poorly designed systems simply firefight as problems emerge.

In Beer’s treatise Designing Freedom^[16], he expounds on what we might call the “efficient inefficiency” problem. This states that there is no productivity gain when technology is used to do more inefficient things (being more efficiently inefficient). Something known as the “Solow paradox”^[17] – that computers have a smaller effect on productivity than we might expect – also makes sense from this perspective.

ChatGPT makes Beer’s argument – dissolve rather than solve – more important. Certainly, a use of ChatGPT is to fix formulas in spreadsheets. But if the spreadsheets are unnecessary, this won’t benefit anyone.

The risk of over thinking the significance of ChatGPT is that it could easily end up as a prime example of using technology inefficiently to do an inefficient thing more efficiently. The task itself may still not be worth doing. And doing inefficient things more efficiently just means you can do more inefficient stuff – which compounds the problem.

New technology should aim to reevaluate entire systems, rather than the automation of individual tasks. If problems are only solved, rather than dissolved, future problems become baked in, and benefits decline.

ChatGPT is a powerful tool, and the potential benefits of AI are as yet not fully understood. But there is a substantial risk that the legacy of such technology is not more unemployment – but a proliferation of bullshit.

References

1. ^{^} remarkable technological development (theconversation.com)
2. ^{^} 450 was larger than 500 (twitter.com)
3. ^{^} same as 2lbs of bricks (twitter.com)
4. ^{^} Buzzfeed is planning (www.theguardian.com)
5. ^{^} has already published articles (www.standard.co.uk)
6. ^{^} productivity gains and automation (www.aeaweb.org)
7. ^{^} Some economists (www.aeaweb.org)
8. ^{^} never any guarantee (hbr.org)
9. ^{^} automation anxiety (www.tandfonline.com)
10. ^{^} controversial (journals.sagepub.com)
11. ^{^} “bullshit jobs” (libcom.org)
12. ^{^} Some argue (www.tandfonline.com)
13. ^{^} look more legitimate (gwern.net)
14. ^{^} Stafford Beer (www.theguardian.com)
15. ^{^} Andrey_Popov/Shutterstock (www.shutterstock.com)
16. ^{^} Designing Freedom (monoskop.org)
17. ^{^} “Solow paradox” (cs.stanford.edu)