AI Highlights — the whole picture — 2026-10-19 (Mon) Evening News

Recent news reports surrounding AI tend to focus on the scale of investment and the rise in stock prices. However, as you read through each section, you will find that the point to look at is other than the size of the amount. Amazon has reportedly installed more than 1 million robots, and while cybersecurity jobs in the U.S. are down 26%, jobs for AI security engineers are skyrocketing. As funds are being invested in equipment and stock prices, the nature of employment, hiring standards, and methods of medical evaluation are changing in the field. What is consistent with this highlight is the view that it is not the investment amount itself that determines whether AI will produce results, but whether people and systems are able to keep up with the changes. In each of the following sections, we will examine the basis for this based on individual cases.

Flow of funds biased toward equipment and stock prices

ABC News & Headlines (Australian Broadcasting Corporation) reports that the sharp rise in AI-related stocks is creating bias in parts of the U.S. economy. Around the same time, The Korea Daily reported how Wall Street banks were approaching AI infrastructure investments with both growth and risk considerations. If we put these two side by side, we can see that the movement of funds is concentrated in two places. One is the stock price of AI-related companies, and the other is capital investment such as data centers. Because banks are involved in both as lenders and underwriters, they are simultaneously assuming that growth will continue and preparing for losses if it does not. What readers should look at is not whether the investment amount is large or not. The question is whether the funds are concentrated in a few areas and arranged so that if one area collapses, it will easily spread to other areas. (ABC News & Headlines, The Korea Daily)

This concentration also casts a shadow on discussions outside economics. The LI Press article, "The AI Dilemma Perpetuates an Arms Race," deals with the logic that once the race for AI begins, it becomes difficult to get off. The structure is such that each entity believes that as long as the other party continues to invest, it has no choice but to continue investing. The same holds true from the perspective of capital investment. As long as other companies and countries continue to invest, it is difficult to make the decision that only one company will slow down. If investment builds up not because of individual profit prospects but because of an inability to step back from competition, then concentration is the result of everyone aligning rather than being the result of careful planning. This contribution discusses this point within the framework of the arms race. (LI Press)

Another factor is the debate over preparations for when concentration breaks down. According to EIN Presswire, economist Schenker has published a book on post-AI recession prevention that has reached number one on Amazon. It is not possible to determine the future of the economy from the contents of books. However, the fact that books about dealing with recessions topped the list of best-sellers suggests that readers are becoming more interested in recessions, along with growth stories. ABC News & Headlines' story of skewed stock prices, The Korea Daily's look at banks, LI Press's discussion of the perpetuation of competition, and Schenker's book sales reflect the same fragility from a different perspective. The more money is tilted in one direction, the wider the impact will be if it deviates from that direction. (EIN Presswire, LI Press, ABC News & Headlines, The Korea Daily)

However, what we have looked at so far is limited to the financial side. Whether investment turns into results depends on how the content of work and the way it is evaluated changes at the site where the equipment is used.

Introduction of robots and replacement of job offers

Amazon has reportedly installed over 1 million robots, raising questions about the future of its employees (36Kr). During the same period, U.S. cybersecurity job openings have fallen by 26%, while job openings for AI security engineers have skyrocketed (SecurityWeek). The former is a movement in which on-site workers are replaced by machines, and the latter is a movement in which the required expertise is replaced within the same job category of information security. Both indicate that the total amount of work does not increase or decrease uniformly, but that there are simultaneously fewer and more job openings. Therefore, judging the future based solely on the job title or industry name may lead to misunderstanding the actual situation.

The measuring stick at the entrance to recruitment has also changed. Korean game companies are moving away from recruiting new graduates based on coding exams to evaluating their ability to use AI (finance.biggo.com). In the past, people looked at whether they could write code on their own, but now they are looking at their ability to use AI as a tool to achieve results. Movements in the same direction can be seen in daily work after joining the company. Israeli media CTech introduces how to organize work procedures to be entrusted to AI as "skills" and time-saving techniques for daily tasks. Being able to use AI is becoming a prerequisite, and the difference lies in whether or not you can organize your work into a form that can be handed over to AI.

There are two conclusions readers can take away from this. One way to do this is to separate and examine tasks within your company and your job that can easily be replaced by robots or AI, and tasks that will create new demand. As in the example of cyber recruitment, even within the same field, one is shrinking and the other is growing. The other thing to do is to review the evaluation criteria for human resources. As South Korean game companies have shown, if you change the content of the exam, the image of who gets hired will also change. Organizations that use traditional evaluation criteria may overlook people with the necessary skills.

In other words, the content of work and standards for evaluation are changing at a rate different from the scale of capital investment. Even if the amount of investment is the same, if human resources development and systems cannot keep up with this turnover, there will be differences in results. In the next section, we will look at the institutional movements that will influence this trend.

Evaluation of medical AI is measured in the worst situation

There are two reports that change the way we look at the evaluation of medical AI. Bioengineer.org reported that the cause of bias in medical AI may lie with clinicians rather than algorithms. KevinMD.com argues that medical AI should be evaluated based on the worst answer rather than the best answer. If the former is correct, the bias will remain even if the model is rebuilt. This is because humans receive the output of AI, interpret it, and decide whether to accept or reject it. If you take the latter position, no matter how high the average accuracy is, its meaning is limited. It is not the average answer that harms the patient, but the worst answer.

When these two factors are combined, it can be seen that the average value of accuracy is insufficient as a yardstick for measuring trust in the field. Even if the average is high, unless you can see the quality of the worst answers and the judgment habits of those who use them, you cannot decide whether to introduce them or not. The subject of evaluation is not the model alone, but the operation of the model and user.

The same thing is happening in industries other than medicine. Bioengineer.org introduced an initiative to synthesize pomegranate disease images using generative AI and use it for on-site identification. Synthetic images can be used as training materials for discriminative models, but whether they are useful for identifying diseases in the field can only be verified by the user. According to Asahi Shimbun, AI drug discovery company Elix and Axcelead DDP have signed a joint research agreement on protein degradation inducers. This is a partnership between a company that has AI and a company that handles drug discovery. Both cases show that the success or failure of implementation depends not only on the performance of the tool, but also on how it is incorporated into the process at the site.

As long as we talk about the results of implementation in terms of average values and investment amounts, these differences are not visible. The next question is whether each site can prepare a system that supports evaluations that assume the worst-case scenario and operations that include the users.

Dialogue with young people: Protection gaps surrounding AI

Can we apply the lessons learned from not being able to protect children through SNS to AI? The Hill in the United States addresses this point both as a ``lesson'' and an ``opportunity'' (The Hill). This means that the experience of SNS is the starting point for discussion of conversational AI. In South Korea, Infoseek reports that the National Assembly has pointed out that there are deficiencies in the protection of young people using AI chat. In the United States, a review of past failures and in South Korea, the National Assembly is reviewing the system, both of which are raising concerns about the lack of protection in the same field. Even if the countries are different, the point of view is the same.

On the other hand, the user side has already moved beyond the discussion of protection. The Berliner asks readers, "Would you like to go on a date with an AI chatbot?" and deals with the topic of considering AI as a romantic partner (The Berliner). RBC-Ukraine posted an explanation by an astrologer, ``Which zodiac sign would you like to talk to ChatGPT about?'' (RBC-Ukraine) There is an article published in a different country's media that talks about conversations with AI as an extension of familiar entertainment like horoscope reading. From romantic relationships to everyday conversation partners, conversational AI is starting to be talked about more like a human relationship than a work tool.

This is where the reader's judgment comes into play. The Hill and Infoseek acknowledge that protection mechanisms are lacking. The Berliner and RBC-Ukraine show that the way humans and AI interact is widely viewed as a topic of entertainment and romance. Juxtaposing the two, there is a gap in protection between the stage where conversational AI is evaluated in terms of its performance and the stage where it is used as a familiar relationship with young people. Will the design and supervision be prepared with young people in mind, or will it be required in advance? The points made by the South Korean National Assembly indicate that this question has passed the stage of debate and has begun as an examination of the system.

The magnitude of this gap cannot be measured by the amount of investment or the number of cases introduced. How quickly can people and systems catch up with the speed with which AI enters the realm of human relationships? Once again, that difference will be the yardstick that determines the success or failure of AI.

Drawing the line between safety standards and awareness/rights

Anthropic has reportedly set "injuries to more than 50 people" as the criteria for triggering AI safety measures, and this has been questioned (Firstpost). Since the numbers are concrete, it seems clear as a standard. However, when you think about whether a response would not begin if fewer than 50 people were injured, the question becomes clearer. When the conditions for activation are determined by the number of people, there is a risk that damage that is smaller than that or damage that is not easily reflected in the number of people will be placed outside the standard. When reading the standards set by businesses, it is important to look at what the numbers count and what they do not count, rather than the size of the numbers.

The ambiguity in drawing the line is not just a matter of safety. Anthropic has touched on the possibility that its AI is conscious, and the reporter believes it will have a major impact on people's lives (Asaase Radio). The question of how we should treat AI changes depending on whether we accept the possibility of consciousness. On the other hand, the yardstick for determining consciousness is not necessarily shared. The difficulty of the judgment itself is also evident in other situations. AI has determined that ``Salvator Mundi'' is not by Leonardo, but peer review is not yet complete (Martin Cid Magazine). Whether the judgment made by AI can be used as a conclusion or remains an unverified hypothesis depends on whether it has gone through the peer review process. Those receiving the judgment need to distinguish between the stages.

This question is also addressed in fiction. It is a ``human rights ace Attorney'' that depicts a trial in which an AI stands in the defendant's seat (d Menu News). In order to make an AI a defendant, it must first be determined whether an AI can be held responsible or have rights. However, in reality, the conditions for recognizing responsibility and status have not yet been determined, including the triggering criteria of more than 50 people being injured, the declaration of possible consciousness, and the determination of authenticity prior to peer review. It can be read that the reason a story works is because the author temporarily sets the conditions for it.

Without deciding what constitutes responsibility or status, even if we evaluate the performance of individual AIs, we cannot decide where to stop the evaluation and to whom it should be attributed. In the next section, we will look at how people and institutions in the field handle this undecidable part.

Concentrating funds, replacing jobs, evaluating medical AI, protecting young people, and drawing the line between safety standards and awareness. Although the fields are different, what each section has in common is that the results are determined by whether or not the mechanism that receives the technology is able to keep up, rather than the performance of the technology itself. Investments are biased toward equipment and stock prices, job openings are divided into those that are decreasing and those that are increasing, and the criteria for evaluating hiring has changed. In medicine, it has been shown that average accuracy alone is not enough, and it is necessary to evaluate the judgment of the user. There are still parts of protection and safety standards that fall outside the framework of numbers and systems. The success or failure of AI cannot be measured by the amount of investment, but rather by whether the way people work, how they are evaluated, and the lines of the system are restructured to match the speed of change.

Q1: Are they able to distinguish between the work that will be reduced and the work that will be increased in their jobs, not by the job title but by the specific content? Q2: Are the criteria for evaluating the output of AI set based on the worst case and the user's judgment, rather than average performance? Q3: Apart from the size of the investment, how do you check whether the development of human resources and the development of systems are keeping up with the speed of introduction?

Finally, three questions

- Are they able to distinguish between the work that will decrease and the work that will increase in their jobs, not by the job title but by the specific content? Q2: Are the criteria for evaluating the output of AI set based on the worst case and the user's judgment, rather than average performance? Q3: Apart from the size of the investment, how do you check whether the development of human resources and the development of systems are keeping up with the speed of introduction? - Are the criteria for evaluating the output of AI set, not just based on average performance, but also on the worst case scenario and the user's judgment? Q3: Apart from the size of the investment, how do you check whether the development of human resources and the development of systems are keeping up with the speed of introduction? - Apart from the size of the investment, how do you check whether the development of human resources and the development of systems are keeping up with the speed of introduction?