AI Highlights — the whole picture — 2026-10-08 (Thu)

It has been reported that HSBC will significantly reduce headcount in its wealth division for high-net-worth individuals in the UK and promote the use of AI. Around the same time, it was reported that Anthropic's claims and third-party test results from ClawSecure were at odds regarding the safety of AI. While replacement has entered the workforce plan, it has not yet been determined who will verify its performance and safety. What you will read in this half-day is the decision that goes beyond whether or not to include AI. Who will use AI, who will be responsible if an error occurs, and what confirmation procedures will be used? Below, we will follow the four ingredients in order: workforce reduction, image diagnosis, safety assessment, human resources and education, and see how decisions shift from the pros and cons of introduction to the design of roles and verification.

Reduction in personnel and the idea of treating AI as employees

It was reported that HSBC will make large-scale job cuts in its wealth division for high-net-worth individuals in the UK and move forward with the use of AI (Financial Times, Reuters). What we can confirm from the materials is that reduction and AI promotion are being talked about in the same context. The scale of the reductions and the breakdown of job types are not covered here. This incident shows that AI has entered the stage of replacing human work in finance. It can be said that discussions about its introduction have now moved to the stage where it will have a direct impact on personnel planning, after the demonstration tests have been completed.

However, the fact that replacement has begun does not alone determine what readers should decide next. FinAi News argues that financial institutions should treat AI agents as employees, not tools. If it's a tool, you can ask who uses it, but not who is responsible for it. If you treat them as employees, you will need to decide the scope of work they will be responsible for, how they will report to their superiors, how to confirm results, and who will be held responsible in the event of a mistake. This proposal raises a different question than the pros and cons of reductions. Rather than thinking about how many people to reduce, we are calling for a design that first considers what role AI will have and who will supervise it.

The same idea can be seen outside of finance. South Korea's Myongji Hospital is partnering with Lavoro AI to test robots for non-clinical hospital tasks (Korea Biomedical Review). It is important that the scope is limited to non-clinical work. Instead of delving into the heavy responsibility of medical treatment, he limits the scope of his work to peripheral tasks. This is not a story about reducing the number of people, but a story about testing how to divide work between people and robots. The HSBC case shows the aspect of ``replacement,'' and this test shows the aspect of ``division of roles.''

IMF Managing Director says AI is both a hope for growth and a danger (CNBC). This duality directly overlaps with the two cases above. Reducing personnel is close to the hope of growth through efficiency, and leaving things to responsibility without clarity is close to danger. The question of whether or not to adopt AI has passed; the question that remains is who will use AI and who will be responsible for the results. In the next section, we will look specifically at how to allocate these roles and how to verify them.

The accuracy of diagnosis and prediction is measured by interpretability and verification

Research to classify tumors and predict response to treatment using images such as ultrasound, CT, pathology, MRI, and endoscopy is now organized by organ. In ultrasound, it was reported that AI classified parotid gland tumors with high accuracy (Bioengineer.org). In CT, HSV-Net, which combines a state-space model and a self-supervised visual model with stepwise contour learning, is used for segmentation of pulmonary nodules (Frontiers in Oncology). In endoscopy, the performance of an artificial intelligence program to determine the depth of invasion of early gastric cancer is being verified (Endoscopy International Open). Although the types of images and organs are different, the common issue is that ``high-precision accuracy'' is not enough.

What I would like to note is that many of the materials place interpretability at the center of their themes and designs. Research predicting response to preoperative breast cancer treatment calls for explainable and interpretable artificial intelligence (Open Science Framework). For prostate cancer, an interpretable machine learning model is used to predict ISUP classification from MRI habitat images (BMC Medical Imaging). In advanced hepatocellular carcinoma, interpretable CT radiomics predicts the immunogenicity status associated with high-affinity neoantigens in patients treated with combinations of TACE, molecularly targeted drugs, and immunotherapy (Cancers). For pancreatic cancer, molecular subtypes are derived and predicted using tissue image models that incorporate genetic structure (Signal Transduction and Targeted Therapy). Designs that link image characteristics to the nature of the tumor region and genetic/molecular information are notable.

What these studies show is a shift from showing accuracy as a single number to showing it in a form that allows confirmation of why a decision is made. Once the evidence is clear, doctors can compare the output with their own findings and decide whether to accept or reject it. If the evidence is not clear, no one will be able to verify the AI's judgment and the responsibility will remain with humans. What readers should look at here is not so much the accuracy of the model, but whether the product or research includes who will verify the output and what steps they will take.

The perspective of questioning the verification system, rather than the pros and cons of introducing it, is not limited to medical care. Whether or not those using AI have the discipline to check the evidence will lead to the design of roles and responsibilities that we will see next.

Mechanism to externally verify safety claims

The National Law Review reports that Anthropic's safety claims are at odds with the results of a trial conducted by ClawSecure. What we need to look at here is not which is correct. Saying that a product is ``safe'' by the developer and coming to the same conclusion in a third-party test are two different things, and the two can differ. For companies implementing the system, the vendor's manual is only one of the deciding factors. Only when you compare the results of your own testing with the results of an external evaluation organization can you establish a basis for assuming responsibility.

This problem is not limited to individual companies. An editorial in the South China Morning Post argues that the US could cede control of AI governance to China and the EU. Who decides the evaluation standards and procedures is directly connected to which country's regulations the products are manufactured and used. It is not possible to say for certain whether the editorial's assumptions are accurate or not. However, if safety confirmation methods become established as a system for each country or region, there is a risk that adopters will not be able to choose for themselves the standards that their company will follow.

The movement to open up evaluation can also be seen in medical research and education. The study published in BioMedInformatics combines machine learning and single-cell analysis to identify ferroptosis-related grading biomarkers and prognostic signatures in gliomas. A literature review in the Eurasian Journal of Oncology and Radiology organizes the use of AI for endometrial cancer diagnosis, staging, and risk stratification. Both are fields in which the results are subject to clinical and other researcher verification, and discussions are based on the premise that the output of AI can be verified against external standards. On the education side, the Daily Bruin reports on UCLA's Digital Humanities major calling for debiasing AI and shaping the technology responsibly. The role of finding and correcting bias cannot be carried out solely within the developer.

Rather than relying on self-reporting, we develop people who can try things out from the outside, align standards, and point out biases. Only when these three things are in place can we actually answer the question of who will use AI and who will be responsible for it.

The foundation is human resources and literacy rather than equipment

TechTarget points out that the key to AI infrastructure is human resources rather than data centers. While investment in equipment tends to be talked about first, this article positions people as the ones who drive the infrastructure and turn it into results. This view overlaps with the movements of American universities. Baruch College wins federal grant to advance AI literacy and hands-on learning (Baruch News Center). It can be seen that investment targets are expanding from equipment to education. Meanwhile, expectations from the industrial side continue. It was reported that major overseas institutions have raised their evaluations of Chinese stocks and are paying attention to AI and emerging industrial fields (Global Times). The more the market has expectations, the more people who can use the technology will determine whether those expectations will come true.

However, skill is not the only thing required of the user. WKYT is covering how to deal with the anxiety that AI brings. Anxiety has a direct impact on the speed of introduction and how it is received on-site, so simply explaining the technology is not enough; how to deal with anxiety becomes part of the operation. At Countercurrents.org, the author argues that AI threatens the meaning of humanity. This is the author's assertion and is not confirmed as fact. Still, it can be seen that the discussion is shifting from talk of job replacement to the question of why people work. In the field of creative writing, LBB - Little Black Book goes beyond general-purpose AI and discusses the trend from the techniques of forgers to the brushstrokes of artists. The question is how to maintain the unique individuality of the writer, rather than the average output that does not belong to anyone.

The three questions may seem separate, but they are connected in that success or failure depends on the ability and values of the user. Literacy gives you the ability to not accept the output of AI at face value, and to separate the scope of what you should entrust to and the scope of which you should make your own decisions. How to deal with anxiety is a prerequisite for maintaining a state in which the field can report errors. Human interest in meaning and brushstrokes becomes the criterion for what should be entrusted to machines and what should continue to be handled by humans. When readers shift their judgment from the pros and cons of introduction to the design of roles and verification, the objects of design include not only equipment and procedures, but also how to train the people who will use it and the values that those people will rely on.

If we base our efforts on human resources and values, the next question is what specific tasks and disciplines these people will be responsible for.

The four materials so far have all pointed to the same point. We have already passed the stage of introducing AI, and the difference lies in deciding who will use it, who will be responsible for the results, and what disciplines will be used to operate it. Reports of layoffs indicate that we have entered a phase where AI will replace human jobs. Diagnostic imaging research has shown that not only accuracy is required, but also interpretability and validation. Discrepancies over safety made it necessary to compare the developer's explanation with external test results. Investing in human resources and education has shown that expectations turn into results only when there are people who can use them effectively. The time to decide whether to implement it or not is over, and now the focus is on design that allocates roles, determines verification procedures, and clarifies where responsibility lies.

Q1: Who checks the results produced by AI in my work and at what point, and who should I report to if I find an error? Q2: When there is a discrepancy between the vendor's safety explanation and the results of your own use or external evaluations, do you decide on which basis to allow use? Q3: Have you clearly determined the scope of work to be entrusted to AI and the person ultimately responsible for the results before implementation?

Finally, three questions

- Who checks the results produced by AI in your work and at what point, and who should you report to if you find an error? Q2: When there is a discrepancy between the vendor's safety explanation and the results of your own use or external evaluations, do you decide on which basis to allow use? Q3: Have you clearly determined the scope of work to be entrusted to AI and the person ultimately responsible for the results before implementation? - When there is a discrepancy between the vendor's safety explanation and the results of your own use or external evaluations, do you decide on which basis to permit use? Q3: Have you clearly determined the scope of work to be entrusted to AI and the person ultimately responsible for the results before implementation? - Have you clearly determined the scope of work to be entrusted to AI and the person ultimately responsible for its results before introduction?