AI Highlights — the whole picture — 2026-11-08 (Sun) Daily News

In the stock market decline, banking stocks and tech stocks both sold off on concerns about AI. In the workplace, seven people in Blockworks' research department have reportedly left the company, and white-collar employees are reportedly leaving the company voluntarily due to concerns about their employment. Expectations for AI are not only a reason to buy, but also a reason to sell and quit. None of this is the result of one day or one company's actions, and it has not been determined that AI is the sole cause. Still, both the market and the workplace are now at a stage where they are confirming what they will actually be entrusted with before they even set expectations. The question readers should keep in mind as they read through this half-day article is not how large the capacity is, but where human involvement and explainability remain.

Expectations for AI fluctuating in the market and workplace

Concerns about AI first appeared in stock prices. CNBC reported that concerns about AI were a factor in the decline in bank stocks following Thursday's plunge in stock prices. NBC News also reports that US stocks have fallen sharply, with tech stocks selling off on the back of election results and AI concerns. This means that stocks that should have been bought based on expectations for AI are being sold for the same reason in different industries, banking and high technology, at the same time. Both articles cover intra-day price movements and do not confirm that AI is the sole cause. Still, we can see that investors are starting to make buying and selling decisions based on AI.

The same movement is occurring in the workplace. According to KuCoin, Blockworks has pivoted to AI products and seven people in the research department have left the company. The Guardian reports that white-collar workers are leaving their jobs voluntarily, fearing AI will destroy jobs. The working side is taking the lead in the intellectual work department of research, and in a wide range of clerical and professional positions. This means that people are deciding to leave even before the company assigns the job to AI. However, a survey of over 1,500 people published by DIAMOND Harvard Business Review found that AI cannot cure workplace loneliness, and that human connections are valued. This will give both those who leave and those who stay a chance to reconsider the elements of their work styles that are difficult to replace with AI.

Macroeconomic discussions are also taking place in parallel. On WIBC 93.1 FM, a university professor discussed the future of the economy by comparing AI, inflation, and the national debt balance. The outlook for the economy as a whole is being questioned, with stock prices leading to capital movements and job separations leading to the movement of human resources. With expectations and concerns moving simultaneously in terms of both funding and employment, it has become difficult for readers to make decisions based solely on the assessment that ``AI will have great power.'' We need to look at how the scope left to AI and the scope left to humans are reflected in concrete results such as stock prices, personnel, and workplace connections. In the next section, we will explore how this distinction is actually tested in the field.

The human hand behind autonomy and the limits of connection

AI agents are often talked about as beings that perform tasks automatically, but the reality is a little different. CNBC reported that humans are working behind the scenes at agents for startups such as Dazzle. Even if it appears to be autonomous on the screen, if some of the results are supported by humans, the implementer should check not only ``what the AI did.'' It is also necessary to check who was in charge of which process. Similar questions arise in the field of dialogue. The AI buddy app "Tolan" introduced by Quasa.io has voice conversation and memory functions. Psychology Today highlighted new research showing that LLMs may be more empathetic than humans. However, what can be gleaned from the materials is the possibility that the responses may appear empathetic. It is one thing to have an empathetic response and another thing to be able to connect with people, and verification of the latter has not yet been shown.

At home and at school, questions about the division of roles become more concrete. According to Quasa.io, the founder of Fitbit has launched a family health management platform called Luffu. Health is directly connected to the lives of each family member. Even when AI provides support, the criteria for users to choose is which decisions they can leave in their hands. In education, The Shorthorn reported on a monthly roundtable discussion on AI education. The theme of the second session was "Building AI Literacy." Continuing to provide opportunities to learn how to use AI itself assumes that the person receiving the output of AI requires judgment.

Given these trends, it is difficult to decide on introduction based solely on the outlook for capacity. Unisba Media reports that Anthropic's CEO warned that "we don't know what's going to happen." Given that the person in charge of development says that the future cannot be foreseen, it is safe to evaluate systems that incorporate human work, such as the example of Dazzle, conversation apps such as Tolan, and home-use platforms such as Luffu, based on where people are currently involved, rather than expectations for future performance. The position of human involvement lies at the same time as the question of who is responsible for the results and how to explain them. In the next section, we will consider where to place this explanation.

Verification of explainable AI advancing in medical care and cancer research

Health authorities in New Zealand are considering using AI to prepare patients for surgery and refer patients to radiology, reports Healthcare IT News. Around the same time, a systematic review and meta-analysis evaluating AI assistance in finding tumors with upper gastrointestinal endoscopy in randomized controlled trials was published in the World Journal of Gastroenterology. Health authorities consider AI at the beginning of their work, and meta-analyses place AI in the most rigorous form of verification in the form of randomized trials. In both cases, the question is not just ``how much responsibility should I give?'' but also ``Which tasks should I be entrusted with and under what conditions?''

In pathology and imaging research, the premise is to provide evidence in addition to predictions. The Science Progress study uses data from multiple institutions to predict the receptor status of breast cancer from pathological images of H&E staining. The model is a hybrid of a vision transformer and a convolutional neural network, with cross-attention fusion and explicit explainability. Sensors' brain tumor MRI research combines 3D deep learning with layer-by-layer explainability and is titled "From classification to clinical confidence." A prostate cancer study in the Journal of imaging informatics in medicine addresses how the value of bpMRI-derived image representations for risk stratification varies depending on the clinical context. This means that the evaluation of the same image expression will change depending on the situation in which it is used.

The target of analysis is not limited to images. Current gene therapy bladder cancer research prioritized candidate biomarkers using integrated multi-omics analysis and interpretable machine learning. LungCURE, announced at the ACM International Conference on Multimedia (ACM MM), is a benchmark for measuring real-world multimodal clinical reasoning for precise diagnosis and treatment of lung cancer. In both cases, the subject of verification is not only the validity of the conclusion, but also the process of reasoning that leads to it.

When readers look at medical AI results, they should look for explanations and validation designs next to accuracy numbers. This is because the scope of what can be explained determines the scope of what people can confirm and accept. If the evaluation of abilities moves to the presentation of evidence and the design of verification, the next question to ask is who will be responsible for this and at what stage.

Design and industrial restructuring to select implementation

Drex, a decision model that returns only a probability distribution without generating sentences when given a choice, was proposed as a public model of four different scales in an arXiv preprint (own site). The method of having students write long chains of thoughts is highly accurate. On the other hand, during the actual test, there are problems such as the answer being unstable, the process of extracting the answer from the output being required, and the cost and waiting time increasing. According to the authors, the maximum Max shows an accuracy close to that of a generative large-scale model. Another preprint, Route Anything, uses a single framework to sort candidates for language, images, videos, sounds, and robot operations. We measured the current state of growth with 374,401 scored runs across 25 benchmarks (own site). Both are reports that have not yet been peer-reviewed and are not final evaluations.

The movement to selectively introduce products is also being talked about outside of research. There is an article (Analytics India Magazine) that discusses the potential of small language models to change the way companies adopt AI. The National Museum of Korea's AI strategy was also featured in CIO magazine (CIO). Rather than using large models uniformly, the idea of deciding the scope of use according to purpose and cost is emerging in both research and companies/facilities. For the reader, the question shifts from which model is the most capable to what size and method is appropriate for the job.

The choice of introduction as a strategy can also be seen in the reorganization of industries. Personalis' Board of Directors (PSNL) unanimously supported the proposed merger between Tempus AI and Personalis. It is reported that up to 50% of the consideration could be paid in cash (Stock Titan). In gene therapy manufacturing, Forge's Piteri-Molina talked about AI, Ohio State, and the future of manufacturing (BioSpace). Should AI be developed in-house, should it be incorporated through a merger, and how should it be incorporated into the manufacturing site? These are management decisions that are separate from the size of ability.

The choice of what to entrust and how much to keep on one's own is inseparable from the involvement of people and explainability as discussed in the previous section. The design of selection is also the design of deciding where to place responsibility.

Conditions for minimizing harm and cross-country commitments

First, let's look at who will decide who is eligible to advance and under what conditions from the scientific side. A one-page essay published in the journal Science argued from the standpoint of the Lakota tradition of science, which places the minimization of harm as a condition of science, that AI in its current configuration does not qualify to move forward in terms of both ethics and science (Science, own website). Behind this are four concerns that have arisen as the use of generative AI and agent-based AI expands. Displacing human involvement, exceeding human supervision, ignoring restraints, and causing irreparable harm to the environment. The author does not deny the possibility that future AI will be consistent with indigenous science. The question is not whether AI itself is possible or not, but whether the current configuration satisfies the conditions. In terms of bias, research has also found that AI's reasoning process reproduces human stereotypes, according to Bioengineer.org. If bias remains not only in the output but also in the process of inference, the explainability discussed in the previous section is more useful for judgment than as a means to boast of one's ability.

Next, there is a movement to draw lines that should be protected in terms of rights. According to freebeacon.com, a senior United Nations official called for AI to be banned from "digital rape" and given "reproductive rights" (freebeacon.com). This requirement raises questions in a different direction than the condition of preventing harm. While the Science magazine article was based on harm to people and the environment, the UN executive's request is based on the idea of creating rights and prohibitions for AI. Readers need to distinguish between the words ``safety'' and ``ethics,'' but the content changes depending on whether the object to be protected is humans or AI. Please note that this request is based on the reporting on freebeacon.com, and it is best to check the details of the request in the original statement.

Questions have also been raised about whether lines can be shared across borders. Asia Society's argument is that India should build an interoperable AI security mechanism with BRICS (Asia Society). War on the Rocks asks whether the US and China can build trust in AI and stability. The way agreements are concluded differs between the multi-nation framework of BRICS and the bilateral relationship between the United States and China. However, the two essays overlap in that they call for a form in which safety standards can be verified across countries, not just within each country. Since the parties who draw the line are divided among researchers, United Nations executives, the framework of emerging countries, and the two major powers, readers' judgment will depend on who will confirm the promise and what procedures they will use, rather than the size of the promise. The design of a verification system is also required of each organization that chooses to introduce the system.

The situations involved are very different: stock prices, job turnover, behind the scenes of agents, medical tests, model design, and conditioning from the scientific side. However, what each chapter has in common is a shift from the stage of placing expectations on AI to the stage of verifying what it can be entrusted with. The fact that humans support some of the agents' achievements, that responses that appear empathetic and connections with people are different, that the scope of entrustment is being ascertained through randomized trials and evidence, and that the question of whether the conditions for minimizing harm are met are all pointing in the same direction. Judgments cannot be determined solely by the size of one's ability. The criteria for choosing to introduce a system are whether it is possible to leave out which processes people will be involved in, how far they can be explained, and under what conditions they will be entrusted with the task.

Q1: Is your workplace in a position to explain which processes of the results produced by AI were carried out by humans and which processes were carried out by AI? Q2: As a condition for allowing the scope to be entrusted to AI, what level of verification and evidence is required, and who verifies it? Q3: As expectations for AI change, what criteria will you use to maintain or reassess your investment in human involvement and explainability?

Finally, three questions

- Is your workplace in a position to explain which processes were performed by humans and which processes were performed by AI among the results produced by AI? Q2: As a condition for allowing the scope to be entrusted to AI, what level of verification and evidence is required, and who verifies it? Q3: As expectations for AI change, what criteria will you use to maintain or reassess your investment in human involvement and explainability? - What level of verification and evidence is required as a condition for allowing the scope to be entrusted to AI, and who is verifying it? Q3: As expectations for AI change, what criteria will you use to maintain or reassess your investment in human involvement and explainability? - As expectations for AI shift, how do we maintain or reassess our investments in human engagement and explainability?