AI Highlights — the whole picture — 2026-11-02 (Mon) Daily News

The introduction of AI is often talked about as a story of efficiency, such as faster processing and less manpower. However, when we look at the scene, a different picture emerges. At EA, there are testimonies that some employees were pressured to use AI after being fired, and there are reports that human workers are filling in the gaps when AI agents fail. The number of processed items displayed on the screen does not include this additional work. The burden has shifted to employment, on-site personnel, securing human resources, and where responsibility lies, but it is difficult to see where the burden is placed in the results of implementation. In each of the following sections, we will examine the distribution in the following order: people and investment, field relationships, responsibility and regulation, demonstration and verification, and the yardstick for agent research.

Behind the scenes as organizations rush to implement the system, the burden shifts to people and investment

An EA employee told gaming industry media Vertical Slice that he was pressured to use AI after being fired (NeoTeo). In a workplace where the number of people has been reduced, the role of mastering AI has been added to the remaining people. There are similar reports regarding AI agents. When the agent fails, human workers fill in the gaps behind the scenes (The Tech Buzz). The results of automation are visible in the number of processes processed on the screen, but the work of the people who pick up mistakes is not. When measuring the effects of implementation, it is necessary to check not only the number of cases completed automatically, but also the number of cases completed manually and who is responsible for the completion.

The movement of personnel is not just a matter of those implementing AI. A major US IT company has announced AI-related personnel cuts, and training providers are being asked to make proposals based on this (AICERTs). Reductions create demand for training, the cost of which is borne by the individuals who have been retrenched or by the receiving companies. Meanwhile, the University of Phoenix released a white paper summarizing the local AI talent shortage and retention measures (PR Newswire). At the same time that some companies are reducing staff, there are regions that are unable to secure human resources. Even though those who let go of human resources and those who are looking for them are in the same market, they do not overlap. If you do not include the cost of recruiting and retaining human resources in your implementation plan, you are only looking at one side of the efficiency calculation.

The financial burden is also on the financial side. Yahoo Finance highlighted three bank stocks that are supporting the AI data center investment boom. Huge capital investments are supported not only by operating companies' cash reserves but also by loans from financial institutions. The Atlantic discusses "What the economy is sacrificing for AI," asking what the shift toward AI is costing the economy as a whole. The same implementation plan is supported by the labor of people in the field and by the funds of lenders in the financial sector.

None of the burdens that we have seen so far are likely to be included in the introduction announcement. If this is the case, the next point to be ascertained is who is responsible for bearing the consequences of that burden.

Quality of relationships between people in charge of the field and people

It has been pointed out that electronic medical records have taken nurses away from hospital beds. Nurse.org asks, "Electronic medical records have taken nurses away from the hospital bed. Will AI do it again?" There is precedent for a tool that streamlines record-keeping to reduce the amount of time spent with patients. Meanwhile, shattered.io reports that the AI-designed phage therapy Omnii-1 has been administered to the first patient. Situations in which AI produces results in medical care are becoming a reality. However, the magnitude of the results and the increase or decrease in labor on site are two different things. When evaluating the introduction, it is necessary to check how the time spent by nurses on hospital beds has changed before and after the introduction.

The same perspective is needed when it comes to relationships between people. Junta de Castilla y León features a Japanese AI girlfriend app. This means that there are already products on the market in which a machine plays the role of a conversation partner. The University of Wisconsin–Madison has released a new study examining the impact of chatbot conversations on loneliness. Research is currently being conducted to confirm whether it alleviates loneliness. If the design of placing a conversation partner in place of a person before the results are available is becoming widespread, then the cost of verification will be borne first by the user.

Even in the field of creative writing, questions are being asked about the role that people have played. An AI picture-story show of Panta and his friends titled ``The real magic of Halloween is compassion'' is available on YouTube. The theme is compassion, the very quality of human interaction. Can the receiver receive compassion even if the creator is not human? This question remains open. The Santa Clara advocated ``no to AI art'' and expressed its opposition to leaving creation to machines. Regardless of whether one stands for or against it, it is important to note that the work of the person drawing or speaking, and the involvement they put into the work, are not counted as the cost of introducing it.

The precedent of electronic medical records shows that the benefits of increased efficiency remain in the record, while the lost involvement is less likely to remain in the record. Whether it's nursing, conversation, or creative writing, we need to measure whether the relationships between people are becoming weaker by the amount of time and effort we save. The next question is who decides how to measure it and who is responsible for the results.

Location of responsibility and drawing the line between fairness and regulation

If damage occurs due to the judgment or use of AI, who will be held responsible and who will be compensated? Several movements with different personalities come to the same question. Pope Leo XIV reportedly asked who would be responsible if AI were given legal personality and warned that human responsibility and victims should be given more weight (Yahoo! News). In South Bend, a workshop was held on how to counter bias in AI used in housing appraisals (South Bend Tribune). In the area of benefits, the Center for Civic Futures will provide a grant to Nava Labs to support efforts to transform state benefits systems with AI (StateScoop). Examinations and benefits are areas where if the results are incorrect, people may lose their housing or livelihood support. Rather than measuring the effectiveness of the introduction solely by the speed of processing, does the design include who will correct errors and how to remedy them? That's what readers should look at.

The question of responsibility also turns to those who create AI. The UK's Competition and Markets Authority (CMA) has launched a formal investigation into Meta for using content from news sites to develop AI. What is being questioned here is the standpoint of the user, the writer and medium of the article used for learning. Regulators are looking at not only who is responsible for the AI's output, but also who gets the materials from.

The same discussion on drawing the line extends to the human side as well. The Spectator discusses why we shouldn't try to "kill" AI. The New York Post reports that the group has warned that the influx of foreign workers into AI companies is a "national security threat." The former is an argument about whether to stop the technology itself, and the latter is an argument about who is eligible to work, and the targets are different. However, in both cases, it is up to people and the system to decide where to draw the line. When we encounter strong arguments, we need to see who the line burdens and protects.

Responsibility and demarcation ultimately remain with people and institutions. If so, the next question to be asked is whether the human resources and systems that can actually take on that role are in place at the time of introduction.

Demonstration progressing in the field of diagnosis and breeding, and remaining verification

In the diagnosis of otitis media, a new study targeting children is testing the use of AI to extend the use of AI to home use (Scienceline). In order to use it outside of the doctor's office, it is necessary to confirm that the results hold even when taken and operated at home. This research is stepping into that step. Regarding image analysis, biparametric MRI-based habitat analysis of the entire prostate gland was reported as a method to identify pathologically high-risk prostate cancer before treatment. Performed at 2 facilities with external verification (Cancer Imaging). For adult-type diffuse glioma, multiparameter quantitative MRI and 2.5D deep learning are used to assess molecular classification and overall survival (Journal of Magnetic Resonance Imaging). When reading these, before looking at the accuracy numbers, it's a good idea to look at whether the data was verified using data from another facility or at the site of use.

For hepatocellular carcinoma, we are using CT lesion-based models to non-invasively evaluate the degree of malignancy and predict the risk of recurrence (Advanced science). For upper tract urothelial cancer, a study has been published that uses machine learning to predict postoperative lung metastasis (Scientific Reports). Both deal with predictions used to make treatment decisions. Who responds when predictions go wrong has a direct connection to the responsibility we saw in the previous section. Meanwhile, a new multi-stage medical image fusion method combines an ordered pixel-averaging filter with a parameter-adaptive dual-channel weighted pulse-coupled neural network (Iraqi Journal of Science). This is not a diagnosis itself, but a preliminary technique for preparing images. However, if the preprocessing changes, the input to the subsequent diagnostic model will also change. Improving image fusion can move the premise of diagnostic model validation.

Research is also progressing in agriculture. Kashmir Agricultural University has won a joint India-US research grant for AI-based soybean breeding (Kashmir Life). Obtaining funding indicates the start of research, but does not indicate that the results have been confirmed in the field of breeding. This point is in contrast to medical papers that include external verification and home verification.

Research is progressing in both medicine and agriculture. However, those deciding whether to introduce the system need to read up on the situation and to what extent it has been confirmed. Any parts that have not been verified will be taken care of by the people in charge and the person in charge at the site. The next section deals with how to accept it.

A yardstick for agent research that measures memory and learning

Are dedicated storage systems more useful than simple RAGs that search through the raw history? Memory Tax, a pre-peer-reviewed preprint published on arXiv, attempts to answer this question by separating the gains and losses (own site). It has been thought that long-running agents need a memory system, but it was unclear how much of a benefit it would be compared to RAG. Using RAG as a standard, the authors called the gain generated by memory the memory benefit, and the loss the memory tax. Furthermore, many dedicated storage systems report that they often perform worse than simple RAGs. Adding memories doesn't necessarily improve your abilities, and unless you use the same yardstick to subtract what you lose by adding memories, you won't see any gains.

The same perspective applies to the learning side. Training working agents through supervised learning requires a large number of professional tasks that produce reliable role models. However, if the information on the task is missing or inconsistent, the teacher's role as a role model is hindered, and the quality of learning decreases. The preprint that reported on WorkFactory proposed a framework for planning information dependencies at the stage of assembling a task, running it on a trial basis, and fixing defects (own site). The authors created 112,686 tasks across 612 occupations. Although the number of tasks is eye-catching, the argument of this study is that the quality of learning is determined by whether the tasks are put into action. When looking at numbers that indicate the amount of training data, you need to be sure that the task was doable and consistent.

The third situation is where the agent is entrusted with post-learning of the language model. This ability was often measured by a single benchmark score. However, in actual post-learning, it is necessary to judge the progress in one field by comparing it with the decline in another field. The preprint that proposed ParetoRSI-Gym fixed data other than those selected by the agent and created an environment that measured performance in five areas: mathematics, coding, finance, medicine, and science (own site). The authors compare the generalization performance of the best-performing systems to the starting model. Please check the original paper for the specific magnitude of the difference. What we can see here is that rather than the highest score in one field, the question is whether or not we are giving up something when we look at five fields at the same time.

Although the three studies differ in their subject matter, they have in common that they measure the gain of the added value side by side with the loss incurred. Numbers that indicate an agent's ability can only be used to make decisions about introducing the agent, along with what criteria it is based on and what it is compared with. In the next section, we will bring this perspective back to the implementing organization and look at who is taking on the burden that is often placed outside of performance numbers.

The cases treated in each section differ in area and character. However, what they all had in common was that even if the results of AI were expressed in a single number, it was impossible to see where the burden would end up. When personnel were reduced or their positions changed, supplementary roles were added to those who remained. Another quantity that changed when recording efficiency was the amount of time spent with patients. The question of who would be in charge of correcting and remediating errors when errors were made was a question in examinations and payments. When it comes to diagnosis and image analysis, verification at another facility or location of use should be checked before looking at accuracy numbers. The mechanism of memory can only be evaluated by measuring not only the gains but also the losses. The success or failure of introduction should be judged not by the level of efficiency, but by how the burden and responsibility was distributed and to whom.

Q1: After introducing AI, are you able to understand the amount of work being completed by humans and the people responsible for it, along with the number of tasks completed automatically? Q2: If an error or damage occurs in the AI's judgment, is it decided before the use of the system who will make the correction and how to provide relief to those affected? Q3: Do the indicators for measuring the effectiveness of AI implementation include not only the cost savings, but also the person to whom the burden is transferred and the location of responsibility?

Finally, three questions

- After introducing AI, are you able to understand the amount of work being completed by humans and the people responsible for it, along with the number of tasks completed automatically? Q2: If an error or damage occurs in the AI's judgment, is it decided before the use of the system who will make the correction and how to provide relief to those affected? Q3: Do the indicators for measuring the effectiveness of AI implementation include not only the cost savings, but also the person to whom the burden is transferred and the location of responsibility? - If an error or damage occurs in the AI's judgment, has it been determined before the start of use who will make the correction and how to provide relief to those affected? Q3: Do the indicators for measuring the effectiveness of AI implementation include not only the cost savings, but also the person to whom the burden is transferred and the location of responsibility? - Does the index for measuring the effectiveness of AI implementation include not only the cost savings, but also the person to whom the burden is transferred and the location of responsibility?