Marcia,
You bring up several excellent points. Let me attempt to address them while also remaining focused on your main thesis: there are some functions that humans perform that AI will never be able to accomplish (specific example: AI cannot correct human mislabeling). It is important to remember that we are searching for evidence of a fundamental difference between humans and machines that will hold for all times and not simply a difference in the current capabilities of humans and machines. If humans currently outperform machines in a specific task, that is no guarantee that it will necessarily stay that way. To prove your thesis, we need a method to establish that permanent differences between humans and machines exist such as: current status/trends in performance indicate subhuman performance for the foreseeable future or, better yet, fundamental barriers or limitations in performance exist that will never allow machines to outperform humans for a specific role.
While researching AI topics, one can’t help but notice the chronic assumption that machines may replace some jobs, but not the jobs in my field; my field is certainly too skilled/creative/complex/human centric for machines to accomplish or even understand. Truck driving is too skilled for AI to accomplish; autonomous truck routes now exist with many more on the way. Music requires too much creativity for AI; AI is currently generating a wide variety of music and even finishing symphonies started by the world’s greatest composers. Science and engineering are too complex for AI; AI/automation is already pervasive in science and engineering and its use is rapidly expanding. Rhetoric scholars must understand and connect with humans in ways that AI never could; AI behind social media and recommendation applications “understands” humans at an ever-improving level and AI companion apps like Replika (over 10 million users) are connecting with humans at an ever more fundamental, and perhaps scary, level. All of these “skills” have been developed in the last couple of decades and the performance and implementation of AI is accelerating. This level of performance has been achieved without artificial general intelligence which is expected to be available within the next couple of decades. With the general ability to assess humans and the ability to perform all of these specific functions, it seems likely that AI will achieve a superhuman ability to assess a human’s capability to perform these specific functions. That is unless there is some fundamental roadblock that will halt or limit the capability of machines. I haven’t found any that are relevant for the next couple of decades.
You are absolutely correct that combining the opinions from a large number of individuals often times, but not always, leads to more accurate results. This is absolutely an important factor to include in the discussion. To properly weigh its impact however, it is important to estimate the practical number of individuals involved for both the human caseand the machine case. How many humans are typically involved in the process of evaluating or labeling another human? In a few cases, voting in a US presidential election for example, a very large number of people are involved, 154.6 million in the 2020 election. In the vast majority of cases, however, the number is somewhere between 1 and 100. How many: teachers are involved when one is given a grade, coaches are involved when one is selected for a team, university employees are involved when one accepted into a college, interviewers for job opportunity, committee members for awards, bank officials for loans, potential partners for relationships, judges/jury members for legal issues..? Typically, between 1 and 100. Now, how many machines can be used when evaluating a human? Just one? There are multiple types of AI and multiple countries/companies working on individual AI solutions, but there can also be multiple instances of AIs. Imagination time… Company A creates an AI human evaluation tool and sells the same tool to 10 different hiring agencies. At the moment that the software is first installed on each computer, each installation is essentially the same. With time, however, the “individual” AI tools will grow apart as they are trained on different data sets (different “educations”) and applied for different roles (different “experiences”). After several years, each AI “individual” could be brought together and each used to evaluate the potential of a human for a specific role. Afterwards their votes could be tallied and summarized into a group consensus. Essentially the same as in the human evaluator case. Ultimately, what would limit either the total number of AI individuals or the number of AI individuals that can be dedicated to a specific task? At a first glance, it appears that there are more fundamental limits to the available number of humans than on machines.
In summary, machines are currently evaluating/labeling humans for a broad range of roles and their performance is rapidly improving. The current trends, further bolstered by the predicted availability of artificial general intelligence, point to a time in the next couple of decades when machines are superhuman at evaluating the potential of humans. There are no fundamental differences between combining votes from individual humans and votes from individual machines to modify the expected outcome. At the present time, there are also no identified roadblocks to halt the accelerating progress of AI. In the end, one of the potential impacts of AI may actually be to demonstrate that humans are indeed understandable and are essentially biological machines.
