AI Highlights — the whole picture — 2026-10-25 (Sun)

Germany's Generali is reported to be cutting 800 jobs by 2030 and reorganizing its business around AI. U.S. Department of Agriculture IT executives say human oversight of AI will be difficult to maintain due to attrition. AI adoption is expanding into employment, finance, and mental health applications. The difference between a decision to rush its introduction and a decision to postpone it is not so much the magnitude of the effect, but rather whether it is possible to maintain people's supervision and support. This half-day article can be read to confirm whether these conditions are met in the areas of employment, financial markets, emotional support, safety monitoring, and diagnostic imaging.

Personnel reductions are progressing and supervisors cannot keep up

Germany's Generali is reported to be cutting 800 jobs by 2030 and reorganizing its business around AI (Menafn). In the United States, concerns about mass layoffs due to AI have led Congress to explore ways to support workers (The New York Times). The layoff movement has figured in both corporate planning and political agendas. On the other hand, the numbers are not aligned in one direction. Based on a survey by the U.S. Chamber of Commerce, it is reported that many small and medium-sized businesses in Oklahoma have reported an increase in employment due to the introduction of AI (KSWO 7News). Even if the same AI is employed, the way it is employed differs between the restructuring plans of large companies and the workplace of small and medium-sized enterprises. It is dangerous to apply one or the other example to your company, ignoring your company's size and industry.

When reductions proceed, the issue will be less about the number of people and more about what the people who remain will be able to do. An IT executive at the U.S. Department of Agriculture said "human oversight" of AI will be difficult to maintain due to staffing cuts (FedScoop). It has been pointed out that even if AI is introduced with the assumption that humans will be responsible for checking the output of AI and preventing errors, that assumption will collapse if the number of people responsible for that role is reduced. From this point of view, the decision to estimate the cost-effectiveness of introduction based solely on the number of people that can be reduced must be reconsidered.

The International Labor Organization (ILO) is also focusing on worker rights, asking who gets a say when AI runs workplaces (Devdiscourse). The question here is whether there is a system in place for workers to express their opinions when AI makes decisions about placement and evaluation. Securing supervisory staff and giving workers a voice may seem to be separate issues, but both relate to the same issue: whether humans can continue to be involved in AI decisions.

One example of this is in Ireland, where FIT and Intel partnered to improve the AI skills of workers (Silicon Republic). Upskilling will be necessary for both those changing jobs and those moving into supervisory roles. However, the existence of alliances alone does not guarantee that training will be able to keep up with the speed of reductions. Which will go first: the decision to reduce the number of people or preparations to support those who will remain or move on? The order in which they are introduced will be the deciding factor whether to hurry up with the introduction or to postpone it. This question appears in the same form in the financial and psychological realms outside of employment.

Overreaction and pressure reflected by the money market and banks

Recent economic indicators indicate that there is a cautious view on AI-related investment and trade. According to a summary compiled by www.christophe-barraud.com, global AI-related trade is under pressure. Japan's inflation rate exceeded the Bank of Japan's target, while the Eurozone's Purchasing Managers' Index (PMI) fell short of expectations. On the financial side, there are concerns that AI agents will shake up bank deposits, and CNBC reports that this may be an overreaction on Wall Street. Since prices, business confidence, and deposits are all discussed at the same time, companies implementing AI are forced to consider not only the expected effects, but also the external environment, such as financing and customer movements. One of the deciding factors in whether or not to rush ahead with the introduction or not is whether or not the concerns are inflated beyond the reality.

On the other hand, transactions in the pharmaceutical and medical data field have not ceased. According to tokenpost.com, Tempus AI and Pathos AI have signed a neural data deal worth up to $55 million. marketscreener.com reports that Tempus AI has signed a multi-year agreement with Bristol-Myers Squibb to license multimodal oncology data. According to Indian Pharma Post, Tricog Health will expand its global footprint with Sanofi and OMRON Healthcare in an effort to move from electrocardiogram (ECG) to AI-powered care. There are contracts with clear amounts and terms, and the parties include major pharmaceutical companies and medical device companies.

These two movements show that even the same AI has different bases for making decisions depending on the area. In financial markets, concerns are first reflected in prices and views, and the question is later whether Wall Street's reaction was excessive. In data transactions, conditions are fixed in advance in the form of contracts in specific areas such as neurology, oncology, and electrocardiography. However, what we can learn from the current materials is the conclusion of the contract and the expansion of deployment, and it is not clear from these reports alone how the verification and support of people will work after the introduction.

If there is an area where transactions can proceed even when market conditions fluctuate, then the decision to introduce it will be less about the strength of the economy and more about whether people can continue to check the results in the field. In the next section, we will look at the realm of the mind to see if this confirmation and support is actually maintained.

More people are looking to AI for emotional support

Young people are starting to use AI to support their mental health, Fast Company reports. Fox News discussed why more people are drawn to Meta's AI Muse. Meanwhile, The Independent reports that a psychiatrist has pointed out the dangers of asking AI for mental health advice. If you put these three cases side by side, the spread of use and experts pointing out the dangers are progressing at the same time. When readers decide on a company's products and how to use them at home, it is not enough to rely on growth in the number of users as a basis for introduction. It is necessary to first confirm that the use of the term to ask for advice is viewed as problematic by psychiatrists.

The Social Market Foundation says there are both opportunities and costs to AI that mimics human empathy and intimacy. Mechanisms that mimic empathy provide easy options for those looking for someone to talk to. On the other hand, the view is that there are costs associated with imitation, and those who introduce them will also have to bear them. The decision here is whether to expand based on the opportunity or to narrow the scope based on the cost.

This question is not limited to medical or counseling settings. CBS News reports what a Minnesota religious leader has to say about the impact AI will have on religion. Boston University is announcing a new exhibition at 808 Gallery that explores the imperfect beauty of prints in the age of AI. Even in two separate fields, religion and printmaking, the emergence of AI is forcing us to reconsider what it means to be human.

As the movement to rely on AI for emotional support becomes more widespread, the extent to which humans can supervise that support will make the difference between a decision to rush its introduction and a decision to forego it.

Drawing the line regarding safety holes and awareness/cooperation

External audit finds major holes in OpenAI's safety oversight (the-decoder.com). This means that an external audit discovered a flaw that was not visible during internal inspections. For those implementing AI, it is one thing for a business operator to explain that it is conducting safety monitoring, and another is to have a third party confirm the results. If you are deciding whether to rush ahead with the introduction or postpone it, you now have a basis to check not only the provider's instructions, but also the presence or absence of an external audit and the content of the findings.

At the same time as this hole was reported, there were three discussions about where to draw the line. The first is the consciousness of AI. National Review, which featured an essay published in Nature about AI consciousness, argued against giving machines rights. The discussion of whether something is conscious or not and the judgment of whether or not it should be granted rights can be treated as separate questions. From the perspective of operators, no matter how the debate over consciousness progresses, whether or not to continue to place humans in charge of AI output remains a matter that must be decided by the company itself.

The second is cooperation between countries. The Bulletin of the Atomic Scientists argued that nuclear security could be an opening for U.S.-China AI cooperation. The idea is that even though two countries are at loggerheads, it would be easier to create a forum for discussion if there is a common interest in nuclear control. The third is medical regulation. According to Radiology Business, the Society of Radiology has issued an opinion as the FDA moves to regulate generative AI. A group of experts involved in diagnostic imaging in clinical settings has a direct voice in shaping regulatory proposals. For companies, a prerequisite for planning implementation is to follow the opinions expressed by academic societies and industry groups in their field even before the content of the regulations is finalized.

If we look at OpenAI's audit results, discussions on consciousness and rights, U.S.-China cooperation, and academic opinions on the FDA, we can see that the parties that draw the line are divided into businesses, commentators, nations, and professional organizations. The question of who will inspect the drawn lines remains unanswered. In the next section, we will look at the extent to which there is a system in place to carry out such inspections.

Expanding the verification of image diagnostic models

In breast cancer, a model was reported that combines radiomics of multi-phase tumor and peritumoral dynamic contrast-enhanced MRI with conventional features to predict the amount of tumor-infiltrating lymphocytes preoperatively (BMC Medical Imaging). For endometrial cancer, preoperative risk stratification using a combination of multiparametric MRI radiomics and deep learning was investigated (Frontiers in Oncology). Regarding the liver, a study has been published to noninvasively differentiate hepatocellular carcinoma from mass-forming intrahepatic cholangiocarcinoma by combining a deep learning model of dynamic contrast-enhanced MRI with a clinical model (Frontiers in Oncology). All three aim to increase the amount of information available to make a decision before surgery or biopsy. Those considering the introduction of a model need to first decide not only the accuracy figures, but also the stage of medical treatment in which the model will be used.

Reports using CT and pathological images also continue. A Transformer-based model was used to predict the prognosis of lung neuroendocrine cancer. The study is a two-center retrospective cohort (Frontiers in Oncology). For lymph node metastasis of early gastric cancer, preoperative risk was assessed using multimodal Transformer using whole slide images of endoscopic biopsy (Annals of Biomedical Engineering). In muscle-invasive bladder cancer, an interpretable contrast-enhanced CT-based cross-attention transformer model was used to predict progression-free survival. This is presented as a multicenter study (Abdominal Radiology). This includes retrospective designs and reports that are limited to two facilities. Whether it can be used as is in daily medical treatment is something that needs to be confirmed through separate verification. On the other hand, the emergence of designs that promote interpretability is directly linked to the oversight issue of whether doctors can inspect model output.

The applications of image AI are expanding beyond medicine. A study using a CNN model to classify abnormal brain MRI images of dogs was published in Scientific Reports. There is a study comparing supervised and unsupervised methods for extracting brain tumors from MRI images (Romanian Reports in Physics). In a field closer to agriculture, a lightweight AI prototype that uses a single camera to detect and remove bad apples was tested in an online sorting process (Bioengineer.org). The same image analysis technology is used for human diagnosis, animal diagnosis, and fruit sorting. Therefore, it is necessary to decide separately for each application how much of the process should be left to machines and what should be checked by humans.

As more verifications are conducted, the success or failure of implementation will be determined by who can continue to check the output and how, rather than the accuracy itself.

What each section has in common is that compared to the speed at which AI is being introduced and used, it is easy to confirm whether human supervision and support can be maintained later. With employment cuts moving forward, the question became what responsibilities those who remained would be able to take on. In the capital market, it has become necessary to determine whether concerns exceed reality. Regarding emotional support, the use of it expanded and experts pointed out that it progressed at the same time. Safety monitoring required third-party confirmation, and image diagnosis required a decision on what stage of treatment to place. Whether to rush ahead with the introduction or postpone it can be determined not only by the magnitude of the expected effects, but also by whether a system for supervision and support has been established before the introduction.

Q1: With the current number of people, do you have enough people and time to continue checking the results of work entrusted to AI even after implementation? Q2: Do you include external confirmation results and experts' comments as conditions for approving the use of AI, rather than relying solely on the provider's explanation? Q3: Are the expected effects of introducing AI compared with the costs and personnel required to maintain human supervision and support in the same judgment forum?

Finally, three questions

- Does the current number of people have enough people and time to continue checking the results of work entrusted to AI even after implementation? Q2: Do you include external confirmation results and experts' comments as conditions for approving the use of AI, rather than relying solely on the provider's explanation? Q3: Are the expected effects of introducing AI compared with the costs and personnel required to maintain human supervision and support in the same judgment forum? - Rather than relying solely on the provider's explanation, are external confirmation results and experts' comments included as conditions for approving the use of AI? Q3: Are the expected effects of introducing AI compared with the costs and personnel required to maintain human supervision and support in the same judgment forum? - Are the expected effects of introducing AI compared with the costs and personnel required to maintain human supervision and support in the same judgment forum?