A small exercise assessing the potential impact of AI and automation on dietetics
TL;DR: I did a small little exercise detailing how AI-based automation may affect clinical dietetics. This first involved pulling a set of 15 job duties from the O*NET system provided by the US Department of Labor. In table 1, I show line-by-line the tasks evaluated along with an assessment of the probability for its automation and displacement, based on the task’s complexity. This portion of the exercise made the assumption that the simpler and more routine the task, the higher the probability of displacement.
However, current AI systems can arguably already perform more complex tasks. Therefore, two additional economic scenarios were considered if AI progress were to advance further: displacing complex tasks would lower required expertise and likely depress wages; alternatively, displacing mostly simpler tasks would raise the skill floor, increase wages, but limit entry for junior workers. If these are the only 2 scenarios, the latter is the preferable path, but would be even better if the profession would increase investment in transitional training pipelines and expanded high-complexity roles to both maintain a wage premium as well as increased employment opportunities. The central takeaway is that leadership should probably more proactively evaluate potential positive and negative impacts of impending technological change on this profession beyond the initial considered outlined here.
-----
This year’s Super Bowl was a showcase of the three industries that seem to be doing well overall these past few years: gambling, glucagon-like peptide-1 (GLP-1) medications, and AI. Since the public release of ChatGPT in late 2022, the acronym ‘AI’ has now been used pervasively across everyday life.
Despite a nauseating push to adopt AI, anxiety regarding mass job displacement via automation by AI is developing amongst the broader workforce. The current dialogue on job displacement has arguably been the most intense within the software development space. While healthcare is generally technologically conservative, discussions on potential AI disruptions in healthcare led by business leaders (they are likely AI-optimists) have also picked up steam over the past two years.
While increased AI usage within healthcare have become increasingly scrutinized, most pieces on this topic have largely been within context of medicine and nursing only. The discourse regarding AI’s impact on jobs within the dietetics space is much less developed.
Therefore, this exploratory exercise will attempt to, at minimum, detail a construct of scenarios for which we can evaluate the expected impact of AI on the dietetics profession. The exercise will focus on assessing the potential of specific job duties to be displaced, discussing the dynamics of automation and job complexity, and providing a suggest path going forward. Additionally, for definition purposes, the term AI will be used quite broadly to encompass technologies ranging from LLMs to more specialized ML models.
Existing AI development within dietetics
It can be generally assumed that over the next decades, the underlying technology driving these AI agents will improve (maybe conservatively but still meaningfully). However, accurately forecasting its potential effects is also difficult. While reputable think-tanks have already tried to tackle this question, some pieces I feel have provided answers that are simply too general to be informative.
There are already some AI applications that have been developed within nutrition and dietetics, and implemented to varying degrees. These areas include: dietary assessment and food recognition, personalized planning and metabolic analysis, clinical decision making within chronic disease context, conversational agents for nutritional education, and mobile coaching. However, as of this moment, none appear to be capable enough for a full-on end-to-end replacement of what a competent dietitian can consistently accomplish. But if we assume that the drive to automate the jobs within dietetics is approaching regardless, developing a mental framework of what areas we can expect major change should be an urgent undertaking.
If AI comes for the dietitian, what part of your job will be affected?
In 2003, an influential MIT paper (The Skill Content of Recent Technological Change: An Empirical Exploration) measured the scale of impact of computerization on the landscape of labor. Given that jobs comprise a blend of routine and complex non-routine tasks, the authors argued that technological change induced through automation will displace more routine tasks while simultaneously augmenting the worker’s performance of the more complex non-routine tasks. Of course, existing AI capabilities are much more advanced, but even so, there’s some evidence that the jobs most affected thus far are those be those centered around lower skilled administrative work. Therefore, the first exercise here will be conducted under the assumption that, as AI technology advances, job tasks that are primarily simple, repetitive, and routine will likely be the first to go.
In order to concretely frame this exercise, it is necessary for us to define the specific tasks that comprise of a job. After all, the act of automation directly displaces the individual tasks within a job rather than the occupation itself. To do this within dietetics, we will first identify a set of job tasks of the dietitian and then systematically evaluate each task for its potential susceptibility to displacement relative to its complexity level. As ~42% of all dietitians work in settings that can be considered clinical, we will focus on this area in particular to limit the scope of this exercise.
The Occupational Information Network (O*NET) system was developed by the United States (US) Department of Labor, which provides a very nice breakdown of the various tasks and duties for workers of any particular job in the US. For defining the list of job tasks of the clinical dietitian, we will heavily borrow the list of duties itemized for the occupation code 29-1031 Dietitians and Nutritionists.
In the table 1 below, we list out each of the clinically relevant job tasks from the O*NET page for ‘Dietitians and Nutritionists’ and provide:
- Description of the Task
- Type of Task (Routine vs non-routine vs interstitial; complexity Level)
- Probability for displacement by AI technology (from very low to very high)
- Based on our earlier assumption, the harder the job task the more difficult it’ll then be for AI to displace (and therefore lower the probability).
- Notes (Rationale)
Please note that for task type, an additional category called ‘Interstitial Tasks’ was added largely inspired by this substack post. While this is written by a current economics PhD student, it brings up a good point about ‘off-the-grid tasks’ of jobs, which are tasks generally not captured in job descriptions (and thus difficult to capture in data for training an AI model). However, these tasks are fundamentally crucial to completing the job well. These are tasks, for example, which may require personal hands-on management of the patient, and can reveal patterns and contextual clues of this patient’s overall situation that are not fully captured in the electronic medical record. These are tasks in which the dietitian, through the development of long-standing relationships with other healthcare co-workers, is able to gather important patient care related clues during informal hallway conversations that can help tailor the nutrition plan of care more effectively.
Table 1 link.
For each of these 15 tasks, I essentially placed a probability of displacement in consideration of the complexity of the task as well as contextual factors that may prevent/augment the displacement from occurring. Feel free to review and point out concerns/thoughts.
As we continue to get newer iterations of these AI models, its ability to take on more complex tasks will likely increase. In an example that is particularly relevant to this discussion, Google’s DeepMind lab just released their AI Co-Clinician Initiative designed to be a novel patient facing apparatus that supposedly is aiming to function as a ‘collaborative’ member of the healthcare team.
What about a scenario if AI can also complete complex tasks?
In the last part, we made the major assumption that AI will mostly displace simpler and routine tasks, while the tasks that are more cognitively (or physically) complex would be harder to displace. However, this is unlikely to remain constant given the current capacity of some AI systems having already achieved the ability to help solve some very complex problems. Therefore, it would also be valuable to forecast the downstream effects on employment and wages under a scenario in which AI becomes capable of displacing both simple and complex tasks.
In another recent working manuscript (yet to be peer reviewed), David Autor (lead author of the earlier MIT study) and his co-author tried to provide an updated assessment of how advanced automation may impact a job, where expert-level tasks can also be impacted.
They posit two theories:
- If the tasks displaced within a job are the relatively complex tasks, the required expertise to perform the job will decrease and subsequently the relative wage of the job will also decrease as the supply of qualified workers will increase (given the lowering of the expertise level).
- If tasks displaced within a job are relatively simple tasks, the required expertise to perform the job will increase and subsequently the relative wage of the job will rise as the supply of qualified workers will decrease (as the job is tougher since there only remains more complex tasks left to do).
- This is essentially the scenario for the assumption we made in the last part for table 1
While even someone with basic knowledge of supply and demand could come up with these 2 frameworks, they can still provide a valuable lens to guide our evaluation of a more dynamic model of AI-based automation.
Under theory 1, if higher complexity tasks such as providing medical nutrition therapy to a highly complex patient can be displaced, then the remaining jobs within dietetics would become simpler, allowing more junior skilled employees to enter the competition pool, thus driving wages down. This of course, implicitly assumes that the jobs are not fully automated away and human labor is still vital to the institution providing the employment.
Alternatively, under theory 2, if lower complexity tasks such as routine charting were to be replaced, then the remaining jobs within dietetics would become more complex, limiting the pool of qualified labor (temporarily), and thus driving wages up.
Under these contrasting scenarios, the preferred path would arguably be under theory 2, where simpler routine tasks are displaced away while higher complexity cognitive tasks remain preserved. This path would not only induce a wage premium, but could also fundamentally build a rigor-intensive reputation that creates a higher baseline wage floor.
However, this theoretical wage premium isn’t without tradeoffs, as it is induced by increased entry barriers and therefore would limit overall access for junior workers.
To maintain long-term viability, the field should develop a more comprehensive transitional training pathway for early-career dietitians (with pay), enabling them to develop advanced skillsets that will be necessary for a more rigor-intensive job. On a more macro level, the profession should also expand both the number of available positions and increase the scope of complexity (research, informatics, analytics etc) to ensure a sufficiently diverse concentration of high-complexity tasks capable of both sustaining the wage-premium as well as overall future professional relevance.
These are neither simple endeavors nor particularly innovative ideas; however, some type of workforce transformation strategy should be considered given the broader trajectory of recent hi-tech developments.