AI Highlights — the whole picture — 2026-10-26 (Mon)

While AI drives employment numbers, the same technology is also powering election information, romantic and child-rearing advice, cancer predictions, and military and pharmaceutical operations. IT giants are reported to have cut nearly 200,000 jobs worldwide, but the scale of the cuts is just part of the changes taking place. Only 8% of American adults get election information from AI chatbots, and if a major AI dating service goes bankrupt, users' partners will be deleted. Regardless of whether the numbers are large or small, humans are increasingly receiving answers from AI and using them as the basis for decisions and relationships. What I would like readers to take away from this half-day is the question of how far we can leave things to AI, and where we should draw the line. The sections below show how the line is positioned differently in different industries and situations.

Personnel reductions take the lead, and the system and economy follow

Indian economic newspaper Livemint reported that major IT companies have cut nearly 200,000 jobs worldwide due to AI transformation. The scale of reductions has already exceeded that of any one company or country. Al Jazeera highlighted that the UK is leading the way in layoffs related to AI and discussed the reasons for this. In the financial sector, Barclays has been reported to have both layoffs due to AI and excessive concerns about bank taxes (Seeking Alpha). Similar movements are occurring in different industries and regions, such as IT, the UK, and banking. Therefore, it is necessary to view AI and employment not as management decisions for one company, but as phenomena common to multiple industries.

This movement is also spreading outside of employment. The IMF cited both high energy prices and AI as burdens in its assessment of the challenges facing Asian economies (India Today). While AI transformation is a cost-saving measure for businesses, it also creates an adjustment burden for the economy as a whole. Furthermore, in the case of banks, the topic of layoffs is discussed in the same context as the institutional issue of bank taxes (Seeking Alpha). When reading a company's reduction plans, unless you look at them in conjunction with tax and regulatory developments, you will not be able to tell whether the decision was made independently or whether it was influenced by the system.

The same trend is appearing in medical organizations. According to Hospitalogy, the CEO of Navvis, head of healthcare, said that while value-based care (VBC) has moved towards outcomes, LEAD has been put on hold. He then talks about preparing for risk models that could become mandatory. What's going on here isn't the AI itself. Payment and evaluation mechanisms are changing, and organizations are being forced to restructure accordingly. Even if the mandate is postponed, the need to be prepared remains. This point overlaps with the example of the United Kingdom, where personnel reductions proceeded faster than institutional improvements.

When lined up in this way, the number of people and the ratio of reductions are not necessarily the result of each company's own decisions. It is influenced by external conditions such as economic evaluation, tax system, and medical payment methods. Those looking at the numbers need to check under which system the numbers were generated. The next question to be asked is not so much the size of the reduction, but rather how much of the on-site decision-making, creative work, and clinical work should be left to AI.

Information and educational settings where bias and trust are questioned

Only 8% of U.S. adults are using AI chatbots to gather information about the 2026 election (Pew Research Center). The number of users is still small. However, if the answers received by that small number are biased, then the small number is no reason for reassurance. In TechCrunch's review, when faced with a choice, Grok chose the Republican Party, while OpenAI chose the Democratic Party. Both models tend to be used assuming neutrality. Still, if forced to do so, they will reveal their position. What readers should check is not whether the chatbot will answer. In a highly partisan question like an election, the question is whether the answer can be used as a basis for making decisions.

In the field of education, the numbers show that there is a lack of preparation. Washington state school AI policies remain 2010-era and unprepared for the future (The Seattle Times). Although students already have access to generative AI, school regulations still rely on language written at a time when AI was not commonplace. If the policy was old, teachers would have no choice but to make individual decisions, and what they would allow and ask of students would change from classroom to classroom. Trust here is not trust in the AI output itself. This is trust in whether the organization that decides how to use it is keeping up with the actual situation.

Movements related to security are also beginning to take shape in response to attacks. After the attack on Hugging Face, there were a number of movements surrounding AI safety. Tech Xplore organizes its progress in chronological order. With the shared infrastructure of publicly available models and tools being targeted, security is no longer just an internal issue at development companies. The question is how to draw the line between what users should check on their own and what should be left to the infrastructure operator.

In creative work, there are examples where responsibility is clearly indicated. Creative Assembly denies AI art in full version of Total War: Shogun 2, admits responsibility (Simulation Daily) What is important here is that they acknowledged that they used something created by AI, and then showed a willingness to take on the task. The common thread between biased models, outdated school policies, security measures taken after the attack, and production companies that admitted responsibility is that trust is built not on performance, but on clarity about who is responsible for what. In order to decide on the scope of work to be entrusted to, the person who will undertake the work must be visible. In the next section, we will see how this line is drawn in more personal, emotional and clinical settings.

The price you pay when entrusting your emotions and relationships to AI

The question of whether machines have emotions is prompting a debate over AI, MarketWatch reports. Although this debate has not been resolved, relationships are being built on the users' side without waiting for a conclusion. According to the New York Post, a major AI lover service is going bankrupt and users' AI lovers are expected to be deleted. The person with whom the user has been exchanging words on a daily basis disappears due to the operating company's management status. Regardless of whether the feelings are genuine or not, the responsibility is on the company to maintain the partner, and if the company collapses, the relationship ends. The Malaysian Reserve reported that AI is changing motherhood for the worse. The more situations in which child-rearing consultations and decisions are entrusted to AI, the more interactions between parents and children that should normally be concentrated may become less effective. This article deals with such changes as an issue.

There are also advances in the ability of machines to read emotions. According to The Decoder, Block has announced a new AI called Dessa that learns solely from audio. It is said that the voice itself captures emotions that cannot be captured through transcription. There is scope for it to be used in situations such as call centers and interviews where judgment cannot be made based on the content of the words alone. However, it is unclear from this report who uses the emotions read from the voice and for what purpose. The more accurate readings become, the greater the risk that those receiving the results will use them as material for making judgments about people.

On the other hand, there are also ways in which AI can support relationships between people. Oncology Nursing News provides an example of how breast cancer risk AI models can be used for triage and nurse-led patient guidance. Here, the AI only serves to estimate the risk, and the nurse is responsible for explaining the situation to the patient and addressing their concerns. While AI Lover and Dessa depend on AI to read their relationships and emotions, in this case the AI prepares the materials for human-to-human dialogue.

Looking at them side by side, whether or not you can entrust your work to AI is not determined solely by the AI's performance. Before using the device, it is necessary to check whether the person entrusted with it will not disappear due to the bankruptcy of the operating company, who will receive the information obtained from voices and words, and whether there will be a human being left behind. This confirmation will serve as the basis for determining the scope of decision-making and creative work that will be entrusted to AI.

Prediction accuracy of cancer treatment advanced with images and data

For breast cancer, attempts have been made to estimate the presence or absence of metastasis before surgery. A study was reported on using AI to predict breast cancer metastasis before surgery using ultrasound image interpretation (Bioengineer.org). Another study presents a calibrated static stacking ensemble framework for predicting breast cancer mortality (Archives of Breast Cancer). Calibration is the process of checking whether the probabilities produced by the model match the actual frequency of occurrence. Since it is related to whether predicted values can be read as probabilities, it is as important as accuracy when explaining to patients or consulting on treatment plans. In the field of pathological images, SCLC-SBT, a spectral bilinear transformer for classifying tissue images of small cell lung cancer, was published in Scientific Reports. Even if the type of image changes, it is not just the accuracy numbers that are important, but also how to read the output.

There are also studies that call for multi-institutional verification. Prognostic risk stratification of glioma using a pathological-based model also identified key genes and biological pathways involved in prognosis, and was published as a multicenter study in Acta Neuropathologica Communications. For gastric cancer, CT two-field deep learning radiomics that includes visceral adipose tissue is said to have improved preoperative prediction of lymph node metastasis. This is also a multicenter study (European journal of radiology). In hepatocellular carcinoma, there are studies using MRI radiomics to predict FAP expression and the immunosuppressive tumor microenvironment. These designs seek to see whether data from other hospitals holds true rather than results from one hospital. What readers should pay attention to is not the size of the highest value, but how much the results change when the facility changes.

The unit of analysis has been reduced to the cell. In bladder cancer, we integrated single-cell/spatial transcriptomics and machine learning to demonstrate the molecular signatures and key risk genes of malignant epithelial cells with high stemness (Discover Oncology). scGlycoAI is a framework that uses machine learning to analyze glycan and RNA sequence data of single cells, determine cell type, and search for glycans associated with malignancy (Glycobiology). From images to cells, the targets of prediction are expanding for each type of data. However, what these papers indicate is the possibility that the predictions will hold true, and whether or not they are at a level that can be used in clinical practice requires separate confirmation. What is being measured now is not the performance itself, but the distance to which predictions can be relied upon. The question of who draws that line, and how, appears in the same way outside of medical treatment.

Introduction to the field will proceed by industry based on evidence

The US Air Force plans to use AI to support aerial refueling, the Military Times reported. Aerial refueling is a mission with detailed operational procedures and safety standards. Using it as support is different from entrusting the entire mission to AI. The other is drug discovery. According to an introductory article published by Unisba Media on April 14th at 16:00, Novo Nordisk and OpenAI have entered into a partnership for AI drug discovery. The same article also mentions the official departure of Apple's head of AI, Gianandrea. The military and pharmaceutical industries have different standards to follow and the costs of failure. Even so, the places where AI can be used are limited to processes that can be supported within the existing business flow.

In medical surgery, the rationale for its introduction is beginning to be shown in numbers. A systematic review published in Cureus summarizes the prospective clinical evidence for robot-assisted intraoperative digital guidance in prostate cancer surgery and kidney cancer surgery. A prospective design is one in which conditions are determined and patients are followed before the results are known. They have a different weight as a basis for judgment than advertising slogans or cases seen in hindsight. Work is currently underway to confirm the effectiveness of intraoperative guidance for each organ, the prostate and kidney.

When these three are placed side by side, there are commonalities in how the introduction progresses. The Air Force is starting with support for aerial refueling, Novo Nordisk is starting with a drug discovery partnership, and surgery is starting with intraoperative guidance for robot-assisted surgery. In both cases, the targets are divided into parts of the business and the results are verified within that range. When looking at the implementation of AI in the field, it is more useful for readers to judge by looking at which processes it is used in and what kind of verification has been carried out, rather than whether it has been declared that AI will be used. Processes that have been extensively verified provide grounds for expanding the scope of responsibility, while processes that have undergone little verification remain cautious. The next question is where to draw the line where people do not give up their final judgment even after verification.

What each section had in common was that the results were determined not by the introduction of AI itself, but by where to demarcate the scope of entrustment. The size of the layoffs drew attention, but the same moves were spreading across multiple industries and regions, placing an adjustment burden on the economy as a whole. In information and educational settings, questions were raised about bias in answers and delays in making arrangements within schools, even if the number of users was small. In the area of emotions and relationships, the management status of the operating company influenced the relationships between users. In cancer treatment, conditions for reading output, such as calibration and multicenter verification, were treated alongside accuracy. In the military, pharmaceuticals, and surgery, the introduction was focused on processes that could be supported within existing procedures, and evidence was accumulating. Rather than deciding whether to use AI or not, we now need to decide in each situation what to entrust to AI and what to keep in hand.

Q1: Which processes in your work do you use AI output as the basis for decisions, and who has a system in place to notice when the output is incorrect? Q2: What standards and procedures are used to check bias and errors in AI output before and after introduction, and are records kept? Q3: What was the basis for deciding the scope to be entrusted to AI? Have there been any conditions and who will be responsible for reviewing the scope when the business or client situation changes?

Finally, three questions

- Which processes in your work rely directly on the output of AI for decisions, and who has a system in place to notice when the output is incorrect? Q2: What standards and procedures are used to check bias and errors in AI output before and after introduction, and are records kept? Q3: What was the basis for deciding the scope to be entrusted to AI? Have there been any conditions and who will be responsible for reviewing the scope when the business or client situation changes? - What standards and procedures are used to check bias and errors in AI output before and after implementation, and are records kept? Q3: What was the basis for deciding the scope to be entrusted to AI? Have there been any conditions and who will be responsible for reviewing the scope when the business or client situation changes? - What was the basis for deciding the scope to be entrusted to AI, and have the conditions and responsibility for reviewing the scope been determined when the business or client situation changes?