AI Highlights — the whole picture — 2026-10-23 (Fri)

Barclays plans to cut around 5,000 jobs, mainly in support departments, citing AI as a reason. Around the same time, Axios reported that an uninvestigated AI agent was accessing U.S. bank accounts, and Google's education AI came under scrutiny from the Senate over concerns about children's mental health. AI is infiltrating people, accounts, and children's lives before there is even debate about whether or not to introduce it. Readers cannot verify for themselves the claims made by those implementing the product that it is safe and that it is effective. State laws and competition between countries are not set up to keep up with this pace. What we can see from this half-day movement is that we need to change our judgment not so much about the pros and cons of introducing the system, but about who should complete the preliminary inspections and who should check the results after the introduction.

Reducing personnel due to AI and relearning people

Barclays plans to cut approximately 5,000 jobs, mainly in support departments (business + IT). The New York Times reported this as layoffs due to AI-driven cost reductions (The New York Times). AI is cited as the reason for the reduction, and the number of reductions is shown in concrete numbers. What readers see here is that the results of the introduction have begun to be talked about in terms of how many people have been reduced. In that case, the number of reductions alone is not enough for those evaluating the introduction. It is also necessary to ask where the people in charge of the reduced jobs will go.

On the relearning side, the people responsible for the efforts are divided. India's IIT Madras has released a white paper covering AI jobs, global standards and reskilling (The Economic Times). In the US, Sundance and the University of Colorado Boulder will establish an AI education center for filmmakers (Longmont Times-Call). In Japan, Houou High School had students learn about AI through a special class in which they created videos to say "thank you" to their families (Minami Nippon Shimbun). The ages and occupations of the people studying at universities, film industry organizations, and high schools are different. On the other hand, it is unclear from this article whether people targeted for cuts like those at Barclays will be able to receive such education. Having retraining available is one thing, but having it available to those who need it is another.

The burden goes beyond just employment. Former New York Federal Reserve President Dudley believes that AI investment will push the Federal Reserve into a situation where interest rates can go up and down significantly (BigGo Finance). From this perspective, huge investments in AI will extend not only within companies that are reducing staff, but also to those who borrow and save money through interest rates. Costs, both in terms of employment and funding, flow out of the implementing company.

Companies will show the benefits of introduction in numbers. However, it is not possible to determine who will bear the costs and how much through companies' self-reporting. The next question is who will be responsible for the measurements and inspections.

AI agent that enters areas beyond the reach of inspection and A for children

An uninvestigated AI agent was accessing U.S. bank accounts, Axios reported in a scoop. What the headline conveys is that an agent who has not gone through any prior investigation has arrived at your account, the gateway to your money. Around the same time, the New York Post reported that Google's AI for education was being pursued by the Senate, raising concerns about children's mental health. Accounts and children are both areas where it is difficult for users to identify risks on their own. Even if the person implementing AI says that it is "safe," readers cannot verify the evidence for themselves. These two cases show that before discussing whether or not to introduce the system, we should ask who completed the pre-introduction research.

The same question is being asked on the state and national side. According to StateScoop, OpenAI has asked the governor to strengthen California's AI safety laws. This means that companies that are subject to regulation have requested stronger rules. However, it is not possible to determine what the request means just by looking at the headline. Until the details are confirmed, it can be read as an expression of the company's stance or its involvement in rule-making. The New York Times argued that A.I. also poses a risk to midterm elections. Geopolitical Monitor notes that China is setting America's tolerance for AI risks. The argument is that whether in elections or in competition between the United States and China, the extent of risk to be accepted can be decided without the user's knowledge.

The judgments that readers should make from this point are as follows. The cases are diverse: Axios deals with banks, the New York Post deals with schools, StateScoop deals with state laws, The New York Times deals with elections, and Geopolitical Monitor deals with competition between countries. Yet, in common, the people who decide on risk tolerance are not the same as the people who take the risks. Corporate self-reporting can favor decision-making, state laws are limited to one state, and competition between states can lead to higher tolerances. Therefore, when evaluating an AI service, it is necessary to check what and when a third party inspected it, rather than looking at functionality or price. If the subject of inspection is not visible, its introduction means that users will silently accept the level of tolerance.

If no one is determined to carry out inspections, responsibility will only be asked after damage has occurred. The next thing to look at is who will bear the cost and responsibility for the inspection.

AI in the field spreading to diagnosis and monitoring

AI is beginning to take on specific tasks in the field of diagnosis and monitoring. Google Research introduced an initiative to boost front-line maternal care with AI-assisted ultrasound examinations (Google Research). In telemedicine, PomDoctor leverages AI and wearables (FinancialContent). Outside of healthcare, Citri&Co and ORTH have partnered on AI agronomy services for fruit businesses (GlobeNewswire). What all three have in common is that they are used to deliver materials for decision making close to the site. What readers should be looking at is not so much what the AI shows, but who can verify the results in the field.

On the research side, results are coming in the same direction. CircumSpectrum, a computational framework that converts gene expression profiles into flux-based metabolic fingerprints, was published in Communications Medicine. PeerJ has a paper addressing a machine learning prognostic signature of lung adenocarcinoma based on senescence-associated secretory phenotype (SASP)-related genes and its association with the tumor immune microenvironment (PeerJ). Radiology Artificial Intelligence published a study on automatic classification of BT-RADS using a multi-agent system using a large-scale language model and convolutional neural network. Nature Communications' menu includes HRDPath, a patient-based, explainable deep learning architecture that predicts homologous recombination defects from whole-slide images. AI can make a wide range of decisions, from estimating prognosis to classifying images and estimating defects related to treatment strategies.

However, even within the scope of the paper title, there are studies that have different weights. Advanced Science's bladder cancer research conducted a multi-center study and cytology verification in multiple cancer types for a multi-task deep learning model that predicts lymph node metastasis and disease-free survival from cytology images (Advanced Science). The degree to which the field can rely on results differs between results obtained from one hospital and one cancer type and results confirmed across facilities and diseases. HRDPath's "explainable" design can also be read as an attempt to help doctors using the results trace back the evidence.

What determines the pros and cons of introducing AI is not whether or not AI is being used, but rather the extent to which the results are verified by external eyes. There is a difference in the level of inspection between the introduction of relying on self-reports, which we saw in the previous section, and medical research, which requires repeated verification. In the next section, we will consider who will be responsible for bridging this gap.

Drawing the line around emotions, creativity, and consciousness

Research into AI that reads emotions from speech is currently at the stage where it is being evaluated by academia. South Korea's Pyler was reported to have two papers on voice emotion recognition AI accepted by NeurIPS (Seoul Economic Daily). Machine estimation of emotions is already a major research topic. Meanwhile, on the creative side, the question is whether thoughtful AI art can compete with AI slop. The contradictions faced by creators are discussed (The Conversation). Additionally, a page entitled ``AI Girlfriend Recommendations 2026 for Beginners'' is being circulated under the name of the Hokuriku Regional Development Bureau. Although I haven't been able to confirm the text, just looking at the title alone shows that products that sell human-like responses are listed in general search results. What the reader should see is that the technology for reading emotions and the products that appear to respond to emotions have different mechanisms and different responsibilities.

Statements about consciousness are also divided. At the NDTV World Summit, Swami Sarvapuryananda is said to have said that AI can be intelligent without being conscious. The idea is that being highly intelligent and being conscious are two different things. In response, IEEE Spectrum reported that researchers are divided on whether or not to care about the welfare of AI models (IEEE Spectrum). There are no established standards for how much consideration should be given to AI, whether from the standpoint of separating intelligence from consciousness or from the standpoint of concern for welfare. Readers do not have to decide at the same stage the fact that AI behaves like a human and whether or not AI should be treated like a human.

There are also studies that measure bias in treatment. A study by Spain's UC3M (Universitat Carlos III of Madrid) showed that there is a gender disparity in media coverage of AI in Spain (Newswise). If there is bias in who talks about AI and who is talked about, there is a risk that the debate over whether or not AI should be treated as something close to humans will be shaped by biased voices. Papers on emotion estimation, AI art, debates on consciousness, debates on welfare, and biased reporting all question where to draw the line between humans and AI. The question of who should confirm this boundary and on what basis will be expanded upon in the next section.

Products that respond to reductions in numbers, account access, educational AI for children, diagnosis and monitoring sites, and research to read emotions. Although the areas covered are different, what each section has in common is that the subject who talks about the workings of AI and the subject who confirms it are the same. Companies spoke out about their results and safety, and in some cases OpenAI, which is subject to regulation, called for stronger state laws. Relearning efforts are proceeding separately in each region, and no one has been determined to verify the results on-site. Before deciding to introduce a system, the important question to ask is not so much performance or cost, but rather who will conduct the independent inspection and at what stage. Only introduction with a clear identity of the person in charge of inspection will provide a basis for users to make decisions.

Q1: Have you decided on the person and procedure to check the output of the introduced AI on site? Q2: Are pre-implementation surveys and post-implementation inspections conducted independently from the AI provider? Q3: Can you show the results and reduction figures due to AI in a way that can be confirmed by a third party?

Finally, three questions

- Has the person and procedure been determined to check the output of the introduced AI on site? Q2: Are pre-implementation surveys and post-implementation inspections conducted independently from the AI provider? Q3: Can you show the results and reduction figures due to AI in a way that can be confirmed by a third party? - Are pre-implementation surveys and post-implementation inspections conducted independently from the AI provider? Q3: Can you show the results and reduction figures due to AI in a way that can be confirmed by a third party? - Are the results and reduction figures achieved by AI shown in a form that can be confirmed by a third party?