Dr Ruth Shillair[1] and Bill Dutton
Discussion about the potential implications of AI has often focused on the loss of jobs. However, a key impact of AI and AI agents is more around the speed of completing tasks. For instance, before AI, a market analysis of a business idea might take days to a week to complete. Someone had to gather the data from reliable sources and compile findings into a comprehensive report, writing an executive summary and perhaps making tables and charts to illustrate key themes. Currently, with AI assistance, this can be done – literally – in a few minutes. Even if the data sources and conclusions are verified by an expert human, the output time is a fraction of what it was without AI.
As this speed of output is seen as the new “norm” then AI will become foundational to white-collar and academic work processes. It is wonderful that this work can be done more efficiently, allowing businesses, governments, academics, and individuals to move faster and make informed decisions. However, there are potential downsides to this efficient and potentially speedy process.
Unverified conclusions
Since AI can produce output that upon a cursory review looks ‘pretty good’, and sometimes even amazing, it is too easy for users to pass on the results without careful review of the output. Are data and sources logical and accurate. Since large language models (LLMs) are at their core, probability models, there is and always will be an error rate. The error rate might vary based on the reliability of the data and the sequences of runs might compound the error rate. Thus, the issue of AI “hallucinations” or making business decisions based on false data is a serious risk. If the humans in the chain feel under time pressure or feel that careful checking of the output is not necessary, it is easy to effortlessly “paste and post” the results. There is little personal investment, as there was comparatively no time expenditure. But if weak conclusions move up the food chain, you could see major failures for governments, businesses, and individuals.
Lack of training opportunities for humans
Another potential issue with acceleration of speed is a lack of opportunities to train new workers. New employees are often assigned the “gopher” work of finding data from various sources and bringing their initial analyses and findings to the more experienced workers for discussion and further analysis. This process gave entry level employees the chance to develop analytical skills, and they could learn much by perusing various sources and searching for the most accurate and up-to-date materials. There might well be far fewer economic incentives for businesses to train entry level workers if AI is seen to do it faster and cheaper.
However, if we rely on those with years of experience to the be sole “verifiers” for AI output, might we lose the pipeline of trained people to make that reality check. Given these conditions we should probably think about how to incentivize the training entry level workers. This would reinstate the pipeline for trained information workers who not only offer needed verification checks, but more importantly can offer creativity and see new methods to improve processes and also be more agile with using IT innovations.
Different outcomes
Any two people who write a report or summary of an event are likely to arrive at different results. They literally read, see, and hear different points based on their different backgrounds and education – in this sense-each person is analogous to an LLM.[2] Not surprisingly, if an AI agent and a human independently do the same information processing task, they are likely to arrive an different outcomes.
For example, after a recent forum, the participants were promised a summary. One of the authors of this post set out to write a personal summary and discussion of the event while colleagues set off to use AI to draft a summary of the same event. A challenge. AI produced the output before the human got started (speed), but the two reports were dramatically different with the AI version far weaker (in our opinion!) albeit useful in catching points missed by the human author. As IT reconfigures the ways we do things, it also changes the outcomes of those activities (Dutton 1991).
That observation might suggest that AI is best used to complement and support people rather than replace them. Working in collaboration with AI rather than using AI to complete a task is likely to take more time as well. In the early days of automation, people were worried or excited about computers replacing people. New forms of leisure were being considered to occupy workers with nothing to do (Linder 1970). The realities were far from the expectations then and probably now.
This means there are clear roles for entry level staff and researchers as well as more time from beginning to the end of many tasks that are not time constrained. With AI, as with earlier innovations in information and communication technologies, early expectations are often at odds with the realities of IT meeting people and organizations – but most often for the better.
References
Dutton, W. H. (1999), Society on the Line: Information Politics in the Digital Age. Oxford: Oxford University Press.
Linder, S. B. (1970), The Harried Leisure Class. New York: Columbia University Press.
[1] This post arose from a conversation between Bill and Ruth Shillair, Assistant Professor and Director of the Media + Information Master’s Program in the Department of Media and Information in MSU’s College of Communication Arts and Sciences.
[2] This analogy of humans with LLMs we owe to Patrick Sweet in a recent conversation.
