When we talk about artificial intelligence and its potential, especially after the advent of models such as GPT-4, Stable Diffusion or Midjourney V5, the debate about the impact of these technologies on employment always takes place. While it is true that AI is still far from completely replacing us, the truth is that it can potentially impact some jobs. And GPT-4 was not going to be the exception, as OpenAI states in a paper published a few hours ago.
Exhibition. The first thing to explain is what OpenAI understands by exposure for the purposes of this study. According to the organization, “exposure” is “a measure of whether access to GPT or a GPT-enabled system would reduce the time required for a human to perform or complete a task by at least 50%.” In this sense, there are three possibilities:
Eo no exposición: There is no or minimal reduction in the time required to complete the activity or task while maintaining equivalent quality, or the use of any combination of the capabilities described under the criteria below would decrease the quality of the activity or task output.
E1 or direct exposure: Using only the theoretical LLM (Large Language Models) or the GPT-4 described through ChatGPT can reduce the time needed to complete the DWA (Detailed Work Activity) or the task by at least half (50%).
E2 or LLM + exposed: access to the LLM alone would not reduce the time required to complete the activity/task by at least half, but additional software could be developed into the LLM that could reduce the time required to complete the specific activity/task with quality at least half. Among these systems we have access to imaging systems.
The database. To conduct the research, OpenAI researchers analyzed 19,262 tasks and 2,087 workflows described in the O*NET 27.2 database. This database includes 1,016 job positions with detailed descriptions of their tasks and processes (the aforementioned DWA).
The results. Given all this, the results of the OpenAI study are, to say the least, striking. One of the organization’s findings is that the highest paying and most skilled jobs are most affected, as are jobs related to programming and writing. The least are those related to science and critical thinking. The complete list can be seen in the table below.
Table with the most exposed jobs.
What does the table say? Basically, the jobs and their exposure level divided by different evaluation methods according to human evaluations and by GPT-4. Mathematicians, writers and authors, web and digital interface designers, reporters and journalists, and accountants and auditors, among others, are the most threatened, with 100% exposure. According to OpenAI:
“The occupations listed in this table are those in which we estimate that GPTs and GPT-powered software can save workers a significant amount of time performing a large portion of their tasks, but this does not necessarily suggest that their tasks can be fully automated using these technologies.
education is important. Contrary to the trend of thinking that AI will kill off low-skill jobs, the OpenAI study suggests pretty much the opposite.
“Our analysis suggests that people with bachelor’s, master’s, and professional degrees are more exposed to GPTs and GPT-based software than those without formal education. Interestingly, we also found that people with some college education but no degree show higher level of exposure to GPT and GPT-based software Looking at the table showing barriers to entry, we see that jobs with less exposure require longer training, which may offer lower pay (in terms of median earnings). ) once proficiency is acquired.In contrast, jobs that do not require on-the-job training or only require internships/residencies seem to offer higher earnings, but are more exposed to GPT “.
There are jobs that are still not affected. Although the top table is somewhat devastating, there is a table at the end of the paper where we can see the jobs that are not exposed in any way. These are, for example, operators of agricultural machinery, athletes, cooks, bricklayers, mechanics or installers of electrical lines, among many others. A striking contrast but one that makes a lot of sense, since GPT and other language models are based on text responses and cannot replace manual work.
Table with jobs without exposure.
A challenge for adults. While GPT-4 has the potential to significantly reduce the time it takes to perform a task, the truth is that OpenAI doesn’t see it as a complete replacement for workers. According to the organization, “although the technical capacity of GPTs to make human work more efficient seems evident, it is important to recognize that social, economic, regulatory and other factors can influence the actual results of labor productivity. As capabilities continue to evolve, the impact of GPTs on the economy is likely to persist and increase, posing challenges for policymakers in predicting and regulating their trajectory.”
A help, not a substitute (for now). OpenAI’s point of view is clear: GPTs have more value in optimizing than replacing, at least for now. The organization gives an interesting example: the legal profession. In this sector, “adoption may be driven by advances in some of the ethical and security risks associated with LLMs: bias, fabrication of facts, and misalignment, to name a few.” A lawyer cannot choose to rely entirely on the GPT result, but will have to review the original documents and do an independent investigation. AI has its biases and its limitations, and that means that it is still far from being a substitute in all jobs, which is why, for now, its value may lie in optimizing time.
We have seen it with the exams. GPT-4 achieves excellent results in the UBE (Uniform Bar Exam, the most popular test in the United States to be a lawyer), the LSAT (Columbia Law School entrance test), the USABO Olympics (biology) or GRE Quantitative (reasoning). and understand mathematical concepts), but fails at advanced programming, lacks creativity, and continues to make reasoning errors and inventing answers. The results seen to date are very good, but not perfect, and therein lies the crux of the matter.
Caution. While it is true that technological advances have affected certain jobs (an example is Amazon and recruiters), this coin has another side and that is the generation of jobs related to said technological advances. Without going any further, the arrival of GPTs and image generation models has given rise to the prompt engineer. Technology does not create or destroy jobs, it transforms them. GPT (or upcoming technologies) may impact certain jobs, but we don’t know what jobs will be created by that impact.
Cover image: Unsplash
In Xataka: How we will ensure that artificial intelligence does not get out of hand
