AI Highlights — the whole picture — 2026-10-09 (Fri)

Investments in AI are now producing results in three places simultaneously. Israel's HoneyBook cut 14% of its workforce in an AI-led restructuring, and the US FOMC minutes reported a shift in price drivers from tariffs to AI investment. On the other hand, research on diagnostic models and reports that AI explains emotions still have limitations in terms of what is verified and what is not. Talk of competition over the speed of introduction stands out, but that alone is not a basis for judgment. Throughout this half-day article, the question that comes to mind is not so much what AI can do, but rather the question of how much of the work, judgment, and relationships that AI should entrust to humans, and where to draw the line for humans to continue to take over. Each section below lays out facts to help you think about where to draw that line.

Reduction in personnel and disappearance of middle managers

Israel's HoneyBook cuts 14% of its workforce in an AI-driven restructuring (Calcalist). In the United States, Goldman Sachs warned about the future of employment, saying that ``AI will eliminate middle management positions'' (NDTV). Furthermore, the British journal Technology Record discusses whether agent-based AI will provide an opportunity for banks to rethink their operations. What all three cases have in common is that the targets of reductions and warnings are not limited to on-site work, but are beginning to reach the middle levels of the organization, those responsible for receiving and communicating instructions and coordinating progress. However, the figures cited here represent only one company's reduction rate, and Goldman's warning is merely a view of the future. It is too early to accept this as a complete picture.

What is relevant to readers' decisions is that the effects of introducing AI cannot be measured solely by one indicator: reduction in personnel costs. When positioning AI as part of a company-wide reorganization, as HoneyBook did, it is necessary to decide which judgments and adjustments to pass to the AI and which to leave to humans, before deciding on the number of employees to be eliminated. Regarding the review of banking operations, Technology Record treats it as an opportunity to restructure existing procedures, rather than adding AI to them. Future results will tell whether the warning to middle managers was correct. Still, you can start taking stock of the tasks your company's managers are responsible for, dividing them into coordination, judgment, and development, without waiting for results.

On the other hand, around the same time, there was also a movement to find new jobs and re-study. IT services giant Kyndryl plans to open an AI innovation lab in Frisco, Texas, hiring more than 300 people (Community Impact Newspaper). Brown University has opened an AI literacy course for teachers who create classes based on AI (Brown Daily Herald). The former indicates that jobs are being created elsewhere at the same time as the cuts, while the latter indicates that teachers are the first to start relearning. However, it is unclear from the materials whether Kyndryl's more than 300 positions will be filled. The jobs that will decrease and the jobs that will increase will not necessarily be for the same people or in the same region.

Therefore, simply looking at the increase and decrease in personnel and looking at the deductions is insufficient as a basis for making a decision. What should be looked at is what to entrust to the remaining and new employees in an organization with a thinner structure, and what they should learn in order to do so. This line leads to the next question of how far we can leave it to AI in areas where results are difficult to verify, such as diagnosis and interpretation of emotions.

Impact of investment on prices and competition between countries

The FOMC minutes reported by Kitco indicate that interest rates are expected to be raised again within the year, and that the factor driving prices has shifted from tariffs to AI investment (Kitco). This means that the main variable used to explain prices has changed from trade policy to AI investment. Up until now, a company's spending on equipment and computing resources has been interpreted as a business decision for each company. It is now included in the records of meetings that decide on U.S. monetary policy and is also involved in the outlook for interest rates. Those implementing AI also need to factor in the possibility that financing conditions and price assumptions may change when formulating their own investment plans.

When investment becomes a policy variable, its speed can be seen as competition between countries. Foreign Policy discusses the possibility of China overtaking the United States in the spread of AI (Foreign Policy). What is being compared is not cutting-edge performance, but how widely used it is. Around the same time, the Oklahoma Attorney General asked Congress to regulate the AI industry (oklahoma.gov). While federal monetary policy measures the impact of investments on prices, state law enforcement officials are calling for regulation. Additionally, Homeland Security Today has an article titled ``AI and the Security Economy: A Structure We Won't Accept'' that discusses AI from both security and economic perspectives (Homeland Security Today).

By arranging the four materials, we can see that the forces that urge investment and the forces that restrain or direct it are operating simultaneously in different places. Central banks are looking at the impact on prices, state attorneys general are asking Congress to regulate industry, and national security debates are questioning the boundaries between economics and national defense. No single entity has yet decided who determines the speed. For readers, the deciding factor is less about the speed of adoption and more about how sensitive their investments are to interest rates, regulation, and international competition.

If this tug-of-war over control continues at the national and industrial level, the same kind of line will be drawn individually within organizations as to which tasks should be entrusted to AI and where should be checked by humans.

Accuracy and practicality of cancer diagnostic models

The European Medical Journal examines both the promise and limitations of using AI electrocardiograms to classify acute myocardial infarction. As the title suggests, models for cancer and heart disease have produced results in various areas, but the results vary depending on the application. Below, we will list the studies cited as material in order and discern what has been shown and what has not been verified.

The first thing that catches the eye is a model that reads the properties of molecules from images. A study in Communications Medicine used deep learning to infer clinically important molecular subtypes of pancreatic cancer from routine histopathological images. The Frontiers in Oncology study uses a preoperative MRI-based deep transfer learning radiomics model to non-invasively predict meningioma brain invasion. A study in the International Journal of Computer Assisted Radiology and Surgery compared three modeling techniques for predicting the extent of resection for planning parenchymal-sparing liver resections. All have in common that they extract information from existing images without requiring additional tests or invasive procedures.

Next, there are models that are responsible for stratifying prognosis and deterioration. Frontiers in Medicine has developed and internally validated an interpretable machine learning model to predict clinical deterioration in newly diagnosed lung cancer patients. The Journal of gastric cancer presented a prognostic model to predict survival after upfront R0 gastrectomy in gastric cancer with a small number of peritoneal metastases. Scientific Reports describes a resource-saving, multimodal AI method for risk stratification after diagnosis of prostate cancer for at-home monitoring. npj Breast Cancer's MIRROR study is an attempt to use artificial intelligence to extract clinical data from electronic medical records in breast cancer clinical trials, and is aimed at the effort involved in running the trial rather than prediction.

However, caution should be taken in the verification stages indicated by these issues. The lung cancer model has been internally validated, and performance at other institutions or populations is not known from this title. Prognostic models for gastric cancer are limited to patients with peritoneal oligometastasis who have undergone upfront R0 resection. The fact that the European Medical Journal separately examines the promise and limitations of AI electrocardiograms for acute myocardial infarction shows that it is difficult to directly apply the results to the on-site classification. What readers should see is what the model predicts, in which populations, and to what extent it has been tested.

This way of reading leads to the decision to stop deciding whether or not to introduce the technology based on binary choices, and to divide the scope of responsibility for each research. Estimation and data extraction from images are easy to use as aids, but the more thorough the verification is, the more decisions are made on diagnosis and treatment. In the next section, we turn to the more difficult-to-examine area of emotional interpretation.

The question of entrusting intimacy and emotions to machines

The movement to seek intimacy from AI is already being lined up as a commercial product. The China Daily Hong Kong edition (chinadailyhk) reports under the title ``AI Pet Market: Cat Meows for Lonely People'' that pets that respond with cat meows are being sold to lonely people. Bioengineer.org reported that the AI has learned to describe emotions during conversations with structured precision. AI is becoming equipped with the ability to read and explain emotions from people's words. Meanwhile, the Financial Times featured Sherry Turkle's stark warning against the artificial intimacy created by AI "peer relationships." The more widespread it becomes as a means of filling loneliness, the more important the question becomes as to whether it can be used as a substitute for relationships between people.

However, what readers are asked here is not what AI can do, but what it is entrusted with. Even if explanations of emotions become more precise, the means to verify whether the explanations are accurate are not necessarily as well-equipped as in the case of diagnostic models. A meowing pet or AI talking partner could alleviate the loneliness of those who use it. At the same time, as Turkle's warning points out, it is necessary to separately look at how a relationship based on the assumption that the other party is a machine changes expectations for humans. Whether or not to introduce it depends not on the superiority or inferiority of the functions, but on what the users want from the relationship.

Old ideas are trying to answer this question. Bioengineer.org introduced an argument for using the virtues of 13th century theologian Aquinas to guide the ethics of AI triage. Triage is the decision to decide who to direct limited medical resources to first. His position is to seek not only rules and calculations, but also the virtue of the person making the decision. First Things also discussed ``How does Scholastic philosophy think about AI consciousness?'' The question of whether humans are conscious or not is a prerequisite for considering whether we can trust AI to interpret intimacy and emotions. Neither argument can be read as suggesting that the responsibility for decision-making can be transferred to machines just because their performance has improved. However, what I have listed here is the title and introduction of the article, and I do not go into the details of each discussion.

Intimacy and emotion are areas where it is difficult to measure accurately with numbers. That's why, unless you first decide on the scope of responsibility, the speed of spread will become a substitute for judgment. The next question is whether this line should be left to individual choice or whether it should be left to organizations and institutions to draw the line.

Although the four sections deal with different areas, they reach the same point. Retrenchments and warnings to middle managers indicate that AI is beginning to extend beyond front-line operations to mid-level decisions and adjustments in organizations. Investment has entered into policy variables such as prices, interest rates, and competition between countries. The results of diagnostic models vary depending on their use, and even if they become more precise in explaining emotions, there are limited means to verify their output. In other words, the speed at which AI's capabilities are increasing is not progressing at the same rate as the extent to which its results can be verified. Therefore, rather than whether the implementation is early or slow, the difference between organizations is whether they decide on what to entrust to within the scope of what can be verified, and leave the rest clearly for human judgment.

Q1: In what areas of your work do you accept the output of AI without being able to verify it yourself? Q2: Who can later verify the results of decisions made by AI and by what criteria? Q3: Apart from the single indicator of reducing personnel and costs, on what basis does your company decide what to entrust to AI and what to leave to humans?

Finally, three questions

- In what parts of my job do I accept the output of AI without being able to verify it myself? Q2: Who can later verify the results of decisions made by AI and by what criteria? Q3: Apart from the single indicator of reducing personnel and costs, on what basis does your company decide what to entrust to AI and what to leave to humans? - Who can later verify the results of decisions made by AI and by what criteria? Q3: Apart from the single indicator of reducing personnel and costs, on what basis does your company decide what to entrust to AI and what to leave to humans? - Apart from the single indicator of reducing personnel and costs, on what basis does your company decide what to entrust to AI and what to leave to humans?