～ Realities and Countermeasures in the AI Era ～ Can Disinformation Be Prevented?
Interview Article | 2026-06-18
7 minute read
What does it mean to build mechanisms that protect society from disinformation through collaboration between AI and humans?
As AI becomes increasingly embedded in society, disinformation—such as deepfakes and fake news that are difficult for humans to detect—is spreading widely and causing visible negative impacts.
How should we confront this new threat? In particular, what should companies, media organizations, and platforms do?
In this session, we welcomed designer Takayuki Fukatsu and Shota Tajima, CEO of StoryHub Inc., as guest speakers, and discussed countermeasures against disinformation and the challenges of social implementation.
How Far Has Disinformation Advanced?
Minamizawa: Today, we will be discussing under the theme of “Protecting society from AI‑generated disinformation.” To begin, I would like to share the current state of disinformation.
Changes in the “Depth = Quality” and “Breadth = Volume” of Disinformation
Nakayama: In the video shown here, real and fake images are mixed together. Can you tell which ones are fake? The correct answer is that four are fake and two are real. As this example shows, fake images and videos that are indistinguishable to the general public are spreading.
Nakayama: In addition, disinformation such as fake news is having a negative impact on society and the economy. The scale of economic damage is significant, and risks have been pointed out globally. In Japan as well, initiatives to counter disinformation have begun to move forward.
Fukatsu: Thanks to generative AI, it has become much easier to overcome language barriers, and the difficulty of the Japanese language—which used to function as a barrier—is no longer as effective.
Nakayama: Disinformation carries not only the risk of being deceived, but also the risk of unknowingly disseminating sophisticated fake content that has been created or modified using generative AI. There have already been cases where articles mistakenly included disinformation and were later retracted, which means that those who disseminate information also need to be aware of the risks they face.
Challenges Arising from “Asymmetry”
Minamizawa: Earlier, Mr. Fukatsu, you were able to identify fake videos with a very sharp eye. From the perspective of a platform operator, how do you view the current state of disinformation?
Fukatsu: I am involved in operating “note,” and if we let our guard down, disinformation can easily overflow. How to prevent it has become a major challenge. Disinformation is asymmetric, particularly in terms of cost. Stopping its spread requires careful, layered design, which in turn raises concerns about increasing costs.
Minamizawa: Mr. Tajima, you are in a position to support media publishing and are directly facing countermeasures against disinformation. What kinds of changes are you seeing in media operations on the ground?
Tajima: At StoryHub, we provide tools that support media article production and corporate storytelling. In practice, “fact-checking” and proofreading and copy-editing carried out in the field take far more time than people might imagine.
For example, to publish a 2,000–3,000-character web article, reporters, editors-in-chief, and proofreaders collectively spend around six hours on checking tasks, and this cost places pressure on business operations.
Moreover, as monetization of web media becomes increasingly difficult, the number of media organizations that cannot afford to allocate sufficient resources to proofreading and copy-editing is growing. As a result, how to balance profitability with trust and quality has become a key challenge for media organizations.
Disinformation has an “asymmetry” in which verification and countermeasures require far higher costs than generation. This imbalance has been exacerbated by AI in some respects. On the other hand, depending on how AI is used, there is also the potential to reduce the cost of countermeasures.
Fukatsu: From the consumer’s perspective, I have started to think that it may be acceptable to give up on checking every single fact and instead conduct fact-checking selectively, based on impact. Information that does not affect one’s life can be consumed regardless of whether it is true or false. On the other hand, the idea is to concentrate fact-checking resources on information that has a high impact on health or daily life.
Nakayama: What to target for fact-checking is crucial. As one example, the Japan Fact-Check Center has indicated a policy of focusing on domestic events in Japan, placing importance on issues that have a wide reach, a deep impact on people, and a sense of proximity. If fact-checking is to be focused, deciding where to place that focus may be the key.
Minamizawa: Determining how to assess impact is difficult.
Fukatsu: Personally, from the perspective of “YMYL (Your Money or Your Life *),” I believe that higher importance should be placed on information that affects health, daily life, or careers, as well as information that could damage communities over the medium to long term.
*YMYL: Categories that may directly affect people’s money, health, safety, or well-being.
Why Can’t Disinformation Be Stopped?
Minamizawa: Earlier, the keyword “asymmetry” also came up, and I would like to clarify where, specifically, the difficulty of dealing with disinformation lies.
What Kinds of Asymmetry Lie Beneath Disinformation?
Nakayama: We will look at the disadvantages that arise from the “asymmetry” in which the impact of disinformation is not proportionate to the effort and cost required for countermeasures, examining this issue from three perspectives.
【Three Types of Asymmetry That Create Difficulties】
① The Asymmetry of “Fabrication” — Lies are easy to create, while corrections are difficult.
② The Asymmetry of “Transmission” — There are reports that false and misleading information spreads six times faster than truth and facts.
③ The “Cognitive” Asymmetry — Disinformation can persist even after being corrected, and rebuttals can sometimes have the opposite effect.
Nakayama: The asymmetry of “fabrication” means that, for example, it is easy to say “Mount Fuji has erupted,” but to correct that claim as false requires verifying it with data, which entails a substantial cost.
Next is the asymmetry of “transmission.” In many cases, disinformation is sensational and appeals to people’s emotions, creating a feeling that it must be shared with others. As a result, it spreads overwhelmingly faster than truth.
Another aspect that makes countermeasures difficult is “cognitive” asymmetry. Once people come to believe disinformation, it is extremely difficult to shift them to a different understanding. Psychological research has pointed out that being confronted with rebuttals can actually make people more entrenched, producing the opposite effect.
Barriers Go Beyond Asymmetry — Challenges of Data Access, Transmission Layers, and Business Models
Nakayama: There is also the issue that data from social media, where disinformation is widely circulated, is not easily accessible because it is closely tied to platform operators’ businesses. While regulatory responses are progressing in the EU, many researchers still feel that accessing social media data remains difficult. The combination of barriers to data access and the asymmetry inherent in disinformation itself makes countermeasures more challenging.
Fukatsu: Another concern is the number of layers involved. Even if a newspaper conducts thorough fact-checking before publishing a news article, distortions can occur at the level of social media influencers. It is not possible to guarantee information clearance at each transmission layer, and ultimately, fact-checking ends up being done at the browser layer where users actually view the content.
Tajima: In the past, media operated in a “direct” manner, with newspapers delivered straight to households. Today, however, news influencers read articles from newspapers and repost them based on their own interpretations.
Fukatsu: They create summary blogs, upload them to YouTube, and the content gradually becomes distorted.
Tajima: In newspaper articles, every single character has meaning, and there is a journalistic convention in which credibility is expressed through nuances such as the question particle in phrases like “whether or not.” However, once even a single layer is inserted, those conventions are ignored, statements become definitive, and the content can effectively turn into disinformation.
Nakayama: Even when the original information is correct, differences in interpretation, the way it is conveyed, or subtle nuances can transform its meaning, inadvertently giving rise to disinformation.
Fukatsu: There is also the view that disinformation is “profitable,” and that economic incentive is another complicating factor. In web media, page views often translate directly into revenue, and because of structures linked to CGM (Consumer Generated Media), stirring controversy and attracting attention can be more financially incentivized than determining whether something is true or false.
This is what is known as the attention economy*.
Minamizawa: So the mechanisms designed to generate profit are encouraging the increase and spread of disinformation.
*Attention economy: An economic system in which people’s “attention” and “interest” themselves have monetary value.
How Far Can Technology Play a Role?
Minamizawa: Through the discussion so far, we have seen that disinformation involves a number of complex and difficult issues. With that in mind, I would like to introduce Fujitsu’s initiatives by using demonstrations of AI-based countermeasures against disinformation and deepfakes, to explore how far technology can go in addressing these challenges.
Countermeasure Example 1: Detecting Online Fraud Through “Impersonation”
Nakayama: We created a demonstration in which someone impersonates our company’s president, Tokita, and instructs a funds transfer. Not only the video, but also the attached invoices and receipts are fake.
When such sophisticated attacks that provide multiple types of information emerge, it becomes difficult to counter them with a single technology alone.
This makes it necessary to adopt multi-layered countermeasures, such as detecting deepfakes, identifying whether attached documents are forgeries generated by generative AI, and cross-checking text within documents against web information and external sources to flag inconsistencies.
Minamizawa: Could you share your honest views on how this technology should be utilized going forward, and whether there are still areas where it falls short?
Fukatsu: It really depends on the use case. Ultimately, it comes down to cost. For example, if it costs five dollars to take down a single fake video, that would be difficult for platform operators to sustain.
Tajima: From a UX perspective, it may be more user-friendly to present the risks in the form of labels rather than pointing out errors directly.
Indicating areas where something is likely to be incorrect, or where a mistake would have serious consequences, and then selecting those as items for human review could be a practical approach.
Fukatsu: Rather than guaranteeing safety, there is an option to design systems that inform users that something “may be risky.”
Countermeasure Example 2: Multi‑Perspective Fact‑Checking in Article Production
Nakayama: We created a demonstration that combines our technology with StoryHub’s feature for fact‑checking during the final review stage of article production. The material used was an article from our owned media outlet, “PR note,” and we intentionally included some disinformation for testing purposes.
When fact‑checking is applied to the text, it cross‑references surrounding information and source materials to check for inconsistencies. The red alerts indicate images generated by AI, and our AI‑generated image detection technology presents the likelihood of them being fake in the form of a score.
Nakayama: As a countermeasure against the risk of unintentionally publishing content that has been edited by AI, we believe that having a mechanism in which AI supports humans at the checking stage is effective. This approach is also useful in cases where, even though an image itself is genuine, a photo from a different past event or a placeholder image is published as is.
Fukatsu: That’s interesting. How do you detect it?
Nakayama: We detect inconsistencies between the context of images and text. By analyzing how an image has been used in the past—including through web searches—and comparing it with the surrounding textual information, we can identify inconsistencies and indicate the possibility that the image was used in a different context previously. This helps prevent mistakes caused by reuse, as well as cases where past photos are presented as if they depict something happening now.
Tajima: That sounds promising. Image reuse is quite common in media. It is fine when humans do this intentionally, but as AI increasingly proposes materials, I think it is interesting that this allows for double-checking discrepancies between human intent and AI suggestions.
Nakayama: It also seems useful from the perspective of recovering from AI misguidance.
Tajima: Having a mechanism in which AI double-checks mistakes that humans make without malicious intent will become increasingly important.
Fukatsu: I would like to treat this as a starting point and move toward detecting even more sophisticated manipulations. For example, cases where a photo is cropped to remove a person, or made to look as if it was taken in a different location.
Nakayama: With regard to cropping, I think it is relatively easy to detect when images have been used in the past. It is possible to identify elements within an image and run searches based on those elements. For newly created materials, in situations such as political events where multiple media outlets photograph the same event from different angles, approaches such as detecting differences across articles about the same event and checking for consistency can be considered.
Fukatsu: From a UX perspective, in addition to making the basis for detection visible, “time” is also important. If fact-checking finishes in 0.1 seconds, people will use it frequently, but if it takes five minutes every time a button is pressed, it will be skipped.
Nakayama: On that point, fake images can now be identified dramatically faster than with our initial algorithms. There is a trade-off between speed and detection accuracy, so we are advancing technology development to optimize that balance.
Fukatsu: Another idea I would like to experiment with is labeling not the results of fact-checking, but the fact that something has “not been fact-checked.” Taken to the extreme, if all news on social media were labeled as “unchecked,” then checked in order starting with those that receive the most traffic, and gradually displayed as “checked” after a few minutes, it would enable speculative execution and help reduce the stress of waiting.
Toward Social Implementation
Minamizawa: Up to this point, we have received valuable insights into the points to keep in mind when implementing technology in society. Fujitsu is also advancing initiatives toward social implementation, so I would like to have Nakayama introduce these efforts.
International Consortium “Frontria”
Nakayama: As discussed earlier, a single technology is not sufficient to address all challenges, and the perspectives and items that need to be checked vary depending on the use case. For this reason, it is difficult for our company alone to tackle the wide range of disinformation, and cross‑industry collaboration is essential.
With this in mind, we launched the international consortium “Frontria,” and have begun initiatives to work together with a diverse range of companies and research institutions to address negative impacts related to AI safety and trustworthiness, with a particular focus on countermeasures against disinformation.
Players who share the same vision bring together their IP—such as advanced technologies, content, and licenses—and we are creating a forum where participants from the user layer across various industries can also join, engage in discussion, and carry out activities collaboratively.
Nakayama: At present, more than 70 organizations are participating (as of the time of this discussion, 77 organizations as of the end of May 2026). Themes are set for each industry, and people from those industries work in groups together with engineers whose expertise aligns with those themes.
From the engineering side, we provide technologies to each community in forms that can actually be used, and we also hold hackathons where participants build applications together.
We have seen eye‑opening applications proposed by participants from a wide range of companies, and concrete initiatives are steadily moving forward.
Key Principles of Building Effective Mechanisms
Minamizawa: When moving toward the social implementation of countermeasures against disinformation, what points should be kept in mind, and where should efforts begin? I would appreciate your advice.
Fukatsu: At the service infrastructure level, the first principle is not to rely on people’s diligence or goodwill. We cannot assume that everyone will press a fact‑check button every time, or select suspicious content and right‑click on it. We need to design systems based on the assumption that people will not do these things, and think about how to make them function under that premise.
I call this the “law of laziness.” People are drawn to easy solutions, and diligent solutions—such as those that require frequent right‑clicking—do not take root. What is desirable is a simple and effortless mechanism in which constraints at the regulatory or contractual level are built into the platform itself, and APIs are automatically called when a timeline is displayed.
Nakayama: Not relying on people’s goodwill as a premise—that really resonated with me. As researchers, we tend to design systems with the expectation that “users will press the button,” but the reality is that they often do not.
Fukatsu: One thing I would like to implement on note going forward is a mechanism that puts the brakes on information that is spreading rapidly. Like a circuit breaker in the stock market, when a certain fact or rumor begins to spread explosively, it would temporarily slow down that spread and create a pause—a kind of dam.
However, to compete with existing page‑view and viral distribution models, UX design and service design will become increasingly important.
Tajima: Business models are gradually shifting away from an era dominated by page‑view–based advertising models, and media organizations are moving toward collecting trust instead. From here on, even if articles are mass‑produced at low cost using AI, the resulting returns will be low, making it difficult to generate profits.
From a slightly different perspective, I think it is important to build mechanisms for creating content that does not require fact‑checking in the first place. Until now, writing articles and reviewing them has involved significant costs. Going forward, if high‑quality primary information is collected, AI can take care of the rest. Rather than publishing content written casually, focusing on gathering solid primary information may make the return on investment clearer.
Minamizawa: That may represent the way media ought to be.
Tajima: In the future, media organizations may primarily focus on building databases of facts, while users themselves decide how to compile and consume that information.
Fukatsu: It is a healthy practice for media organizations to guarantee facts, and for influencers or YouTubers to provide explanations using those facts as sources. It would be good if facts could be connected in a way similar to a blockchain.
Nakayama: With regard to the “information chain,” there are also research approaches that observe how information spreads and transforms. As information becomes incomplete, is selectively excerpted, or has emotions exaggerated, it may still remain connected as a set of facts, but at some point the chain breaks.
I think it would be interesting to build mechanisms that detect where this happens, or that provide incentives to those who connect facts correctly.
Tajima: If missing information within a given fact, or the angles from which information has been omitted, could be made visible, it would also be possible to consider models that complement those gaps.
Making It Work as UX and Services
Minamizawa: Finally, could you share your views on what is most important in turning technologies that have been put into practical use as countermeasures against disinformation into excellent user experiences and trusted services?
Aiming to Increase Authentic Information
Fukatsu: Fundamentally, I think there are two key points: making things “effortless,” and changing the underlying economic principles. It tends to work well when media organizations guarantee facts and users simply read them, but it does not work as well when media publish large volumes of content and expect users to verify the facts themselves. One approach is to remove users’ burdens at the browser layer, the media layer, or both.
The other point is how to break existing economic incentives. While it is difficult to design systems in which sensational articles are unprofitable, I believe that mechanisms which increase reward scores for people who share fact‑checked articles, or who post reliable information, can work to some extent. Ideally, trustworthy influencers would be given badges and rewarded with higher compensation.
Nakayama: Even if we accept that disinformation circulating is unavoidable to some degree, there are certainly ways to create incentives for increasing truthful information.
Networking Trusted Services
Tajima: I have the sense that we are coming full circle back to the era of SEO. In the same way as “PageRank*,” where value is evaluated based on the number of citations, I think it is important to connect trusted services with one another and build a network.
In business‑to‑business contexts, there is a motivation rooted in brand governance to partner with companies that are considered trustworthy. As with the consortium discussed earlier, services operated jointly by trusted companies can potentially increase their social value.
*PageRank: An algorithm developed by Google to measure the importance of web pages, interpreting links as votes and ranking search results accordingly.
Minamizawa: We have now reached the end of our session. Thank you very much for joining us today.
Wrap up
This talk session highlighted that countermeasures against disinformation are not merely a technical challenge, but a systemic issue involving asymmetric cost structures, business models, and UX design.
As there are clear limits to what individual organizations can do on their own, cross‑industry approaches—such as the consortium “Frontria” introduced in this session, where knowledge and technologies are shared across sectors—may prove to be a key factor going forward.
Takayuki Fukatsu
Designer, CEO of THE GUILD
Shota Tajima
CEO, StoryHub Inc.
Takayoshi Nakayama
Fujitsu Research
Sara Minamizawa
(Facilitator)
Fujitsu Design Center
