AI Highlights — the whole picture — 2026-10-17 (Sat) Morning News

As the bank cuts 3,000 jobs, it explains that the purpose of AI is to improve efficiency, not replace it. In enterprise software, it is said that even if AI agents approach autopilot, human involvement will remain. In drug discovery, partnerships worth up to $1.5 billion are being formed, and in diagnostics, predictive models are entering decisions about individual tumors. The power to generate speed reaches the workplace, the laboratory, and the clinic at the same time. What I would like to pass on to the reader here is the view that the success or failure of implementation is not determined by high performance, but by drawing the line between which processes should be left to AI and which decisions should be left to humans. The company's explanation and the size of its funds are not the basis for drawing that line. In the main text, we will take a look at where the lines are drawn in each area.

Words regarding personnel reductions and changes in roles occurring on-site

In articles reporting on layoffs at banks, words of explanation appear alongside the number of job cuts. Axis Bank will cut 3,000 jobs in FY26. According to a report (Goodreturns), the bank aims to use AI to improve operational efficiency and not to replace human resources. In the tech industry, local media is also talking about where AI and layoffs will take jobs (KTVU). However, in this type of announcement, the number ``3,000 people'' and the explanation ``not a replacement'' appear on the same page. Words of explanation are not a basis upon which readers can rest their judgment.

On the corporate software side, a design that leaves people in touch is talked about as a specification. An analysis of Workday Rising 2026 (iTWire) reported that Workday's AI agents will move closer to autopilot while maintaining human involvement. As automation progresses, product design decisions become more important in deciding which processes to employ people in. What those considering the introduction should look at is not how many people will be cut. The question is what part of approval and exception handling remains in the hands of people.

Changes in the nature of work are more visible in the fields of nursing care, education, and housing support. An article (McKnight's Senior Living) featuring a report on nursing homes reported that the use of AI is supporting nursing care workers. Shizuoka Prefecture has introduced AI chat to prefectural housing complexes to help prevent single-person households from becoming isolated (Mainichi Shimbun). In education, SBCCD and BBOP Center have launched an AI literacy partnership at College Corps (Precinct Reporter News). All three of these cases can be read as examples of AI changing the scope of work performed by people and the abilities required of them, rather than replacing people. Caregivers will be able to entrust some of the recording and confirmation to AI, and residents of the housing complex will be able to first consult via chat. Those who learn will acquire the ability to fully utilize AI.

Therefore, the impact on employment cannot be measured solely by the number of cuts. Even though the word ``improvement in efficiency'' may be used in the same term, what happens to workers is different when reducing the number of people than when reducing the number of people involved in dialogue and making decisions. What will determine the success or failure of introduction is the line drawn between what should be entrusted to AI and what should be left to humans. We will next look at how this line is drawn in areas such as drug discovery and diagnostics, where the cost of error is high.

Funding and clinical implementation of AI advancing through drug discovery and diagnosis

Funds are already moving into AI-driven businesses. AstraZeneca has signed a partnership worth up to $1.5 billion with AI drug discovery company Incitro (BigGo Finance). Google.org will support a University of Toronto team with $2.3 million to build OpenPep, an AI engine for peptide drug discovery (Bioengineer.org). In finance, Sweden's Matters raised $5 million to launch an AI-native bank (The Next Web). In Japan, AI, shipbuilding, and drug discovery are being discussed side by side among the 17 strategic fields with a budget of 370 trillion yen (Nikkei Business Online Edition). It is important to note that the phrase "up to" $1.5 billion for the partnership does not mean that the full amount has been committed. Rather than the size of the amount, what readers should look at is whether the funds are structured on the assumption that the process will be entrusted to AI.

On the clinical side, implementation discussions are catching up with funding movements. Diagnostic imaging leaders discussed the extent to which AI automation can be extended to clinical settings in a webinar (Radiology Business). Research is the basis here. The Journal of Imaging has a paper that examines the impact of AI assistance on radiologists' performance in lung nodule detection using CT using a systematic review and meta-analysis of detection accuracy studies. What is important is that it measures the performance of doctors using AI, not the accuracy of AI. A study in the Journal of Neuro-Oncology used multicenter data to develop and compare a deep learning, radiomics, clinical, and integrated model for preoperative WHO grade prediction of meningiomas by MRI. Since the design is to compare stand-alone models and integrated models side by side, it can be said that this research asks what combinations can be used.

この二つの流れを並べると、判断の置き場所が見えてくる。創薬や新しい銀行では、資金が先に出て、成果は後から検証される。診断では、検証の問いが「AIが正しいか」から「AIを使った人の成績がどう変わるか」へ移っている。 For those considering introduction, it would be better to check that there is evidence that measures the performance of the person, rather than the amount of partnership or procurement, in order to decide the scope of entrustment.

The speed at which funds are raised is not the same as the speed at which verifications accumulate. The question of who will fill the gap leads to the next section.

Prediction models enter into individual tumor judgment

In breast cancer research, a new digital marker for evaluating CXCL9 expression and prognosis was demonstrated using a "pathomics signature" created using machine learning (CytoJournal). In lung adenocarcinoma, ATP2A2 has been identified as a factor promoting microenvironmental remodeling and immune evasion from AI-derived inflammatory and metabolic ecosystem signatures (Frontiers in Cell and Developmental Biology). Both are studies that extract indicators related to prognosis from pathological images and gene expression data. What they have in common is that they attempt to show not only the numbers produced by the models, but also to which biological factors those numbers are linked.

Research using images also shows the same direction. Metastatic castration-sensitive prostate cancer uses explainable machine learning to examine factors that predict whether patients will receive guideline-aligned treatment intensification (Frontiers in Oncology). For liver cancer, we combined intratumoral habitat and peritumoral radiomics to improve the accuracy of predicting the efficacy of initial TACE. This was a multicenter retrospective study (Frontiers in Oncology). For renal cancer, we use contrast-enhanced CT-derived extracellular volume fractions to preoperatively predict the WHO/ISUP grade of clear cell renal cell carcinoma using interpretable machine learning (Frontiers in Oncology). The scope of prediction, including prognosis, treatment selection, treatment effect, and grade, is expanding to each stage of individual judgment for each tumor.

The outcome of these studies is whether they can show the reasons for the predictions rather than the predictions themselves. At the stage of applying the model to individual patients, a model that cannot demonstrate evidence will not serve as material for physicians to consider, whereas a model that can be demonstrated will serve as material for physicians to either be satisfied with it or to notice errors. For those considering implementation, it is not enough to compare accuracy numbers alone. It is necessary to ask whether it is possible to present the prediction based on what basis in a form that can be confirmed by clinical personnel. Furthermore, research on liver cancer is retrospective. Readers should also note that the results have not been verified through actual medical treatment.

This trend is not limited to tumors. A study using AI to analyze 101,000 people showed that brown fat is associated with heart and mental health (News-Medical). As the scale of analysis increases, the speed at which relationships are found increases, but it is up to people to decide how to interpret those relationships and to what extent they are used in decisions. The fact that explainable predictions are the axis of evaluation also means that the responsibility for accepting predictions remains with the person who uses them. In the next section, we will look at how to divide this responsibility within the organization.

AI that continues to be affirmed and pressure from regulation and competition

AI that constantly affirms people has the risk of damaging interpersonal relationships. Forbes JAPAN reports concerns about blame shifting and lack of empathy. The idea is that if the user only receives answers that are convenient for them, there will be fewer opportunities to reevaluate their own behavior. On the other hand, the dialogue service side continues to expand. Universidad de Córdoba (UCO)'s offerings include an unfiltered sexual conversation chatbot for people over 18, and an "AI girlfriend love simulator" service with privacy controls (AI Girl Generator Online With Customizable Characters And Privacy Controls). Privacy settings have become a selling point, and how users' data is handled is part of the product. Those considering introducing it need to first check whether the AI is designed to approve anything, and who will manage the conversation records, rather than the number of functions.

Discussions regarding AI's ``thinking ability'' continue. Deccan Herald drew on the ideas of Purusha and Prakriti from Indian philosophy to address the limits of AI consciousness and the question of whether machines can think. Psychology Today discusses why real paintings look AI-generated. Both of these points indicate that people's sense of how to distinguish ``AI-likeness'' is wavering. The premise of using humans to determine whether an output is genuine or a product needs to be reconsidered.

Regulatory and competitive pressures are beginning to impose external constraints on usage. According to Aju Press, the US Federal Trade Commission (FTC) has launched an investigation into OpenAI and Anthropic following a security incident. The Mighty 790 KFGO reported that China has proposed adding 40 technologies in the areas of AI, biotechnology, and solar to the export control list. StateTech Magazine reports that AI in American governments is moving from demonstration to full-scale operation. As public institutions enter the stage of incorporating this into their daily operations, standards for procurement and accountability will become stricter. As seen in China's precision agriculture, which was covered by CGTN, AI continues to transform industrial settings. Supplier accident investigations and technology export controls highlight the fragility of designs that depend on specific companies or models.

The macro environment is also not irrelevant. Investing.com discusses the paradox of the US dollar and inflation, reporting that AI-related credit risks and rising energy prices are reshaping the outlook for global growth. Investment in AI is supported by both funding and electricity, and if these conditions deteriorate, the profitability of the introduction plan will be affected.

Positive dialogue, investigations, export controls, and credit risk all externally set the conditions for how much AI can be trusted. The next question is who will decide the scope of responsibility under these conditions and how.

Explaining personnel cuts, funding for drug discovery, predictive models that support diagnosis, and dialogue services that continually return positive responses seem to be topics in different fields. However, they all come to the same point. What and to what extent AI will be responsible will not be determined solely by technological capabilities. Whether or not the design side determines who will check the output and who will be responsible for the results will determine the trust of users, patients, and employees. The words of explanation and the amount of money are no substitute for the design. When deciding whether to introduce a system, you should first check the processes that will be left to people and where responsibility should be placed, along with speed and cost effectiveness. The clearer the scope of responsibility is, the easier it is to review and explain later.

Q1: Of the work you are responsible for, how do you differentiate between parts of the work where you can use the AI output as is and parts of it that you should check with your own eyes before using it? Q2: If there is an error or bias in the results produced by AI, who will find it at what stage, and how will it be stopped? Q3: When talking about the results of efficiency improvements through AI, to what extent can you specifically explain to people inside and outside the company the scope of judgment and responsibility that is left to people?

Finally, three questions

- Of the work you are in charge of, where do you separate the parts where you can use the AI output as is and the parts where you should check it with your own eyes before using it? Q2: If there is an error or bias in the results produced by AI, who will find it at what stage, and how will it be stopped? Q3: When talking about the results of efficiency improvements through AI, to what extent can you specifically explain to people inside and outside the company the scope of judgment and responsibility that is left to people? - If there is an error or bias in the results produced by AI, who will find it at what stage, and how will it be stopped? Q3: When talking about the results of efficiency improvements through AI, to what extent can you specifically explain to people inside and outside the company the scope of judgment and responsibility that is left to people? - When talking about the results of efficiency improvements through AI, to what extent can you specifically explain to people inside and outside the company the scope of judgment and responsibility that is left to humans?