AI Highlights — the whole picture — 2026-10-14 (Wed)

Kroger has announced plans to reduce the number of employees at its headquarters by approximately 3,000 people, while using AI to reduce redundant work. At the same time, research on predicting HER2 expression in breast cancer from mammography has progressed to an area close to diagnosis. There are already situations where the results of AI are used in practice, such as organizing tasks and supporting medical decisions. On the other hand, given the speed with which it was introduced, it has not been determined who will check whether the results are correct and who will take responsibility if an error occurs. The topics of this half-day have in common something that cannot be measured by efficiency figures. In each of the following sections, we will look at where the procedures and responsibilities for verifying results are determined and where vacancies are located, in the order of employment, diagnosis, practice, regulation, and creation.

Reduction in personnel and professionals losing opportunities to grow

US retail giant Kroger has announced plans to cut approximately 3,000 jobs at its headquarters in line with the expansion of its use of AI (Reuters). Yahoo News UK reports that this reduction will come in the form of AI that reduces duplication of work. The targets are back-office departments at the head office rather than on-site stores, and the cases in which AI is directly affecting employment appear not in terms of new products or sales floors, but in the reorganization of overlapping tasks between departments. For those considering introducing AI, the difference between its effectiveness and its impact on personnel is not so much what AI can do, but rather whether or not it can identify which of its own operations are redundant.

However, it is possible that the duties that were being reduced were also opportunities for those in charge to gain experience. GeneOnline reports that Sanzenbacher analyzed how AI automation could change the apprenticeship model for the trades. The question that arises is that if redundant and routine work is reduced, there will be fewer opportunities for younger workers to acquire skills while helping their seniors with their work. There is also movement on the training side. MSU Denver received recognition for preparing students for the AI era (MSU Denver - Early Bird). Universities are making progress in teaching students how to interact with AI, but whether they will be able to develop systems for training people in the field after they find employment at the same speed is another question.

Furthermore, the efficiency of a single company cannot be separated from its external economic environment. According to CNBC, the IMF's chief economist talks about inflation and AI risks and rising borrowing costs. In a situation where borrowing costs rise, companies will more seriously reconsider their personnel expenses and investment decisions. How cuts like Kroger's will interact with interest rates and inflation cannot be gleaned from individual company announcements alone.

Decisions to reduce the number of employees or to review the training system are based on the premise that the AI output is correct. The question of who verifies that premise leads to the next issue of verification and responsibility.

Introduction progresses in diagnosis and drug discovery, remaining verification homework

The AI market for breast imaging diagnosis is expected to exceed $1.8 billion in 2033 (Grand View Research). On the research side, the scope of prediction and classification is expanding. BMC Medical Imaging focused on low HER2 expression with a two-center study using mammography-based deep learning radiomics to predict HER2 expression in breast cancer. The same journal also includes a study that uses CT habitat radiomics and deep learning to predict the intra-airspace spread of early-stage lung adenocarcinoma before surgery. The International Journal of Drug Delivery Technology reports on a self-evolving digital twin ecosystem aimed at early detection and precision management of breast, lung, colon, and ovarian cancers. In both cases, the targets are not only finding lesions, but also determining receptor expression and preoperative progress, which are close to treatment strategies.

There is movement in the same direction in the field of the heart and dementia. Medical Economics' Morning Medical Update reported that an AI model has improved the accuracy of cardiac ultrasound diagnosis. IGC Pharma announced PRISM, an AI platform that advances Alzheimer's disease risk prediction into action and commercial development for brain health (BioSpace). This suggests that prediction research is moving toward the implementation stage, such as behavior and sales.

On the other hand, there are also evaluations that question the selection of the method itself. A paper in PeerJ Computer Science comprehensively compared conventional machine learning and deep learning for brain tumor classification using MRI, and questioned the versatility of deep learning. This means that the premise that using deep learning is not always a good thing has been demonstrated from the evaluation design side. What the implementer should check is which method will produce what degree of accuracy based on the target disease, image type, and facility conditions. The fact that the prediction of HER2 expression was shown in a two-center study also indicates the need to separately confirm whether the same results can be obtained with other facilities and devices.

Reports of improved accuracy and prospects for market expansion provide support for adoption. However, whether this can be achieved with the patients and equipment of the own facility is another question. The problem of the verification system remains: who picks up on errors and who takes responsibility when they occur.

Penetration into business practices and AI infrastructure within companies

In the insurance field, AI is beginning to enter into individual business decisions. Enlyte introduces a mechanism to predict risk scores from unstructured notes in physical therapy records (Risk & Insurance). On the medical side, Medicare home health care costs $0 out of pocket, but it has been pointed out that AI-generated discharge summaries cannot replace face-to-face consultations (24/7 Wall St.). While AI that deciphers records can speed up insurance assessments, practitioners are already questioning where to draw the line: it cannot replace on-site examinations and judgments.

Even in areas such as research and development and manufacturing, implementation progresses in specific work units. FRONTEO's medical paper search AI was introduced at the Suntory Global Innovation Center (Iyaku Tsushinsha). Research has also been reported on detecting crop pests in real time using explainable deep learning (Bioengineer.org). Mazda explained the technological progress of its "empathetic AI" at the Japan Mobility Show (Yahoo! News). The situations in which it is used differ depending on the industry, such as searching for literature, detecting pests, and having conversations in cars. However, in all situations, the common process is that a person receives the output of AI and takes the next action.

Funding and foundation building to support such on-site deployment is also progressing in parallel. Vertex Ventures led a $3 million seed round raised by AI science startup Drylab (Tech in Asia). As the number of individual uses increases, the internal infrastructure and funding flow to accommodate them will also be established. Once the infrastructure is in place, the speed of implementation will increase, but at the same time, the decision-making body for deciding which tasks and to what extent AI output can be adopted will become more likely to be dispersed from site to site.

What the reader should look at, rather than the number of installation sites, is whether the document clearly defines the procedure for checking the output at each site and where responsibility should be placed in the event of an error. As the number of locations used increases, the next decision that must be made is whether to leave the procedures to individual sites or to centralize them as an organization.

Regulation/supervision and geopolitical pressure

Banking regulators are increasing their scrutiny of AI, and Reuters reports that this is due to growing concerns about financial stability. There are also movements in the same direction in Europe. According to PPC Land, the EU draft law makes the statute of limitations for sanctions against Google and Meta's AI restartable every five years. This can be seen as a design that leaves room for sanctions to be imposed over time and maintains the effectiveness of regulations. For companies implementing AI, the scope of oversight will not only include systems currently in operation, but also records of past operations. The basis for decisions and how long and at what level of granularity the logs are kept are additional conditions that determine whether or not to introduce them.

Regulation is aimed at more than just financial stability. 조선일보 reports that AI-based price discrimination is a topic of controversy in the retail industry. The question of whether it is fair for consumers to be shown different prices for the same product has come up in front of consumers. In the election field, ABC10 reports that AI-generated campaign ads are attracting attention in California's gubernatorial race. The authenticity of the information that reaches voters is beginning to be treated as a question of how technology is used. Furthermore, on the security front, Reuters reported that Nvidia's share of AI semiconductors in China is declining and Huawei is closing the gap. The semiconductor competition between the United States and China has turned the AI supply chain itself into a political issue.

These three developments require new confirmation from those making decisions about the introduction of AI. In addition to checking whether efficiency has increased, it is necessary to check in advance whether the method for determining prices can be explained, whether the source of the generated content can be shown, and whether there are any restrictions on the sources of the semiconductors and boards used. Regulatory pressures take different forms in different countries and sectors, but they all come back to the question of how and by whom the results will be verified. In the next section, we will look at the current state of implementation from the perspective of the mechanisms that support verification.

Fluctuations surrounding humanity

Regarding the debate over whether AI is conscious, Modern Ghana quotes a cognitive scientist as saying, ``No one can say for sure.'' Since experts avoid making definitive statements, there is little basis for readers to answer the question ``Does AI feel?'' right away. It is not the presence or absence of an answer that influences the decision. Without answers, products and services incorporating AI will be used first.

In the field of creation, the question of how to measure value is being asked in a more concrete manner. According to The Cool Down, a Tasmanian knitter's beanie hat went viral, prompting a copycat called the "AI Slop" to appear on Etsy. One-of-a-kind items created by the creators themselves and items made to resemble them are lined up in the same sales area. It will be difficult for buyers to differentiate items based on appearance or price alone, and information about who made them and how they were made will become an independent part of the value of the item. There is a similar movement on the relationship side. Hong Kong International Airport features an "AI companion", where the conversation partner itself is provided as a product. When a machine is responsible for interacting with people, what constitutes value is not uniformly determined and differs depending on the user's lifestyle.

On the field side, Chunichi BIZ Navi reported on ``Family-owned business in the age of AI: Ohno Knife Seisakusho's challenges.''

Protect safety

(above)" is posted. Knives are items that are directly linked to the safety of the person using them. What the family-run manufacturing company is trying to protect in the age of AI is reflected in the title, "Protecting Safety." As indicated by (above), the article is a continuation, and the conclusion has not yet been presented. For manual labor sites, the issue is not whether or not to use AI, but rather who is ultimately responsible for safety.

What all four cases have in common is that there is little that can be said with certainty. Even experts cannot say for sure whether they are aware of it or not, and it is up to sellers and buyers to draw the line between counterfeits and genuine products, the value of companions is determined by individual users, and manufacturers treat safety as their own responsibility. When decisions are left up to the individual's own lifestyle and field, the next question is who will provide the means of verification and responsibility to support those decisions.

What is common in each of the sections so far is that, rather than the speed with which AI produces results, it is not decided in advance how the results will be verified and who will undertake the work. Due to the personnel reduction, it became difficult to see the training function that had been carried out by the duties that had been cut. In diagnosis and insurance assessment, the accuracy of predictions and the process by which humans confirm them have been treated separately. Regulators are moving toward designs that can even cover past operational records. When it comes to price and creation, information about who made it and how it was made has become part of its value. The success or failure of an introduction cannot be judged solely by the number of man-hours saved or accuracy. We are now at a stage where the criteria for judgment will be whether or not the verification mechanism and the place of responsibility can be demonstrated at the same time as the introduction.

Q1: Among the tasks you are responsible for, can you specifically name the processes in which the output of AI is checked by humans and the processes in which it is passed without checking? Q2: Regarding AI decisions, for what period, at what granularity, and by whom should past operation records be kept in a state where they can be reproduced? Q3: Can you compare and explain the man-hours reduced by introducing AI and the costs added to the verification and accountability system using the same metrics?

Finally, three questions

- Among your duties, can you specifically name the processes in which the output of AI is checked by humans and the processes in which it is passed without checking? Q2: Regarding AI decisions, for what period, at what granularity, and by whom should past operation records be kept in a state where they can be reproduced? Q3: Can you compare and explain the man-hours reduced by introducing AI and the costs added to the verification and accountability system using the same metrics? - Regarding AI decisions, for what period, at what granularity, and by whom should past operational records be kept in a state where they can be reproduced? Q3: Can you compare and explain the man-hours reduced by introducing AI and the costs added to the verification and accountability system using the same metrics? - Can you compare and explain the man-hours reduced by introducing AI and the costs added to the verification and accountability system using the same metrics?