Title: HCI Pioneer de Souza: When AI Reads for Us, Who Is Responsible for Knowledge? | ACM MM 2026
Body:


Have you ever had this experience: faced with dozens of pages of densely packed PDFs and genuinely unable to find the time to read them closely, you hand the paper to ChatGPT or Claude and ask it to extract the core points for you.
Seconds later, a clearly organized summary appears on the screen. You nod, feel that you have grasped the main content of the paper, and even cite its points directly in your own research.
But wait—are you sure that what the AI summarized is really what the original author meant? If it quietly introduces ideas that are not in the original text, or even links completely unrelated concepts together, can we as researchers catch it in time?
At the ongoing ACM Multimedia 2026, Clarisse Sieckenius de Souza, a pioneer in the field of human-computer interaction and professor emerita at the Pontifical Catholic University of Rio de Janeiro (PUC-Rio) in Brazil, delivered a keynote speech.
Starting from her own theory of "Semiotic Engineering," she discussed how AI is changing the relationship between people and information, and how, when we hand the work of reading and understanding literature over to AI, we can preserve researchers' "Epistemic Authority."
From how interfaces convey designers' intentions, to how AI summaries deviate from the original text, to how the PEEL method can be used to examine AI's text-processing process, this talk offers us a new perspective for understanding knowledge responsibility in the AI era.
Below is the full text of the speech, from Leiphone (official account: Leiphone):
In my view, human-computer interaction (HCI) can be understood as a conversation.
But this is not a simple conversation between an application or system and a user; it is a conversation mediated by computers that takes place between groups of people: on one side, the people who design and develop applications, and on the other, the people who use those applications.
This conversation unfolds through the interface. The interface actually plays the role of the designer's delegate, conveying information to the user on the designer's behalf. It can only express itself in the ways and in the language pre-established by the designer—what we usually call an artificial language.
For example, when we use a website, the interface tells us which operations are allowed and which are not. On a cookie preferences page, a user may try to toggle a necessary setting, but the interface tells us that this operation cannot be completed.

This message actually comes from the designer. The feedback the user gets through the interface reflects what the designer wants the user to do, what the user can do, and what the user cannot do.
This is the basic idea of Semiotic Engineering: the interface speaks on behalf of the designer, and the user communicates with the designer by learning and using the language of the interface.

Of course, interface language is not necessarily verbal. It can also be a fully graphical, object-centered language. We express our intentions by manipulating objects on the screen, and the interface responds to our operations through changes in those objects.
However, the interface does not always respond as the user expects.
When I was preparing this talk, I encountered a rather interesting situation: I tried to insert a text box into the interface, and it actually came out upside down. I had no idea why this happened.
This precisely shows that we sometimes cannot accurately predict how the interface will respond. When faced with unexpected results, we can only try to understand them and adjust our operations.

Usually, every system has a set of established interaction patterns. These patterns to some extent constrain what users can express and what they can do. We need to keep trying in order to find ways to express our intentions through the interface.
More importantly, every system has its own unique language. Even if two systems use the same words, the meanings behind those words may be completely different.
Therefore, whenever we encounter a new system, we need to go through a process of language learning.
This is somewhat similar to a child learning their native language, or an adult learning a second language. We keep trying, observe how the system responds, understand the meanings of different words and visual symbols, and then gradually master this language.
When we face a system we have never used before, we often make a great many attempts. We do not know what a certain button means, nor what result a certain visual element will produce; we can only build understanding through interaction.
So what exactly does Semiotic Engineering consist of? I would like to explain it from three aspects.
First, human-computer interaction is a form of mediated communication. Designers and developers convey their design intentions to users through an artificial interface language. By learning this language, users communicate with the designer's delegate and try to achieve their own goals.
Of course, users' goals are not necessarily always the goals envisioned in advance by the designer. Sometimes we use an application creatively to accomplish tasks the designer never anticipated, and these operations may even work very well.
Second, communication in the interface relies on three kinds of signs.
- The first is static signs. They exist fixed in the interface and can convey meaning simply through the way they are presented.
- The second is dynamic signs. They need to convey meaning through time and changes of state. For example, when we click a button, the change the button undergoes and the behavior that follows tell us what the button means.
- The third is metalinguistic signs. This is a technical term referring to signs that can explain other signs. For example, an interface prompt can tell us what a certain graphical object represents, and can also guide us to take a certain action to achieve a specific goal.
Third, users have a strong capacity for language learning. Designers make use of existing interaction patterns, familiar language systems, and metalinguistic prompts to help users understand new interface vocabulary and semantics.

The above three aspects constitute the basic framework of Semiotic Engineering. But with the development of AI, the way we communicate with computers is changing. AI is now able to converse with us in natural language.
So a question arises: since AI can understand natural language, has Semiotic Engineering become less important?

About four weeks ago, while using Claude to prepare this talk, I encountered something that surprised me greatly.
Claude's interface looks very simple; it is a conversational interaction pattern. We ask a question in the input box, and Claude gives an answer on the screen.
But in some cases, the interface will also predict what the user might type next based on the context of the current conversation, and proactively offer suggested replies.
This in itself is not strange. What truly surprised me was one specific conversation.
At the time, I was discussing some issues with Claude. It suddenly listed a set of points, and the first item was actually what it believed to be the main thesis I wanted to express. The point it gave was roughly this: accountability mechanisms must be designed, and cannot simply be assumed to exist.
I was very surprised, because I had never thought of making this point the main thesis of my paper. But I found the idea interesting, so I followed its suggestion and continued the discussion. I told Claude that the argument it had raised was good and that we could build the discussion around it.
Unexpectedly, Claude soon corrected me. It said that one thing needed to be clarified first: it had not proposed this paper thesis earlier; it had only been asking what I really wanted to express. Now, looking back over the whole conversation, it realized that this point seemed to have appeared.
So I kept pressing: if it was not you who made this judgment, then who, or what, presented this point to me? Claude's answer was very interesting:
"I don't know. I can't see your input box, so I can't tell whether the application offered suggested replies, nor do I know what generated those replies or what context they were based on. By the time this message reached me, it looked like you had said it. I have no way of knowing the actual situation. My earlier correction was accurate—I did not write that paper thesis in this conversation, but I also cannot know where that expression actually came from. I should have made this clear."
This is a very interesting conversation. I did not fabricate this example at all; in fact, even if I wanted to fabricate it, it would be hard to imagine such a scenario.

This experience made me realize that the AI conversational systems we face may be much more complex than they appear on the surface. We think we are communicating with a system that understands context, but we may not know what other agents, mechanisms, or other sources of information behind the interface have participated in this conversation.
Values written like {{NAME_ab12cd34}} are placeholders for real values. Use them exactly as written, in replies and in tool calls alike: they are filled in with the real values before anything runs. Don't ask to verify them, and don't replace them with examples.
So, has semantic engineering lost its value? My conclusion is exactly the opposite. Semantic engineering has not lost its value; on the contrary, it faces an even more daunting task. Because even today, the transfer of information between humans and AI can still be muddled: we see a piece of information, but we may not be clear about where it actually came from.

We need to rethink: what information should the interface actually convey to users? Do users know who they are communicating with? When a piece of information appears in a dialog box, can we determine whether it comes from the user, from AI, or from some other intermediary?
This is also the question I really want to discuss next.
What I really care about is the widespread application of AI in scientific research, especially AI-assisted literature review.
Reading literature is a daily part of our research work. We need to read a large number of papers, understand what a given paper says, what contributions another paper makes, and how a research question relates to existing work.
But right now, there is simply too much for us to read.

I recall the number of submissions to this conference—reportedly more than 8,000. Faced with such a huge number of papers, who can finish reading them all with only their own time and energy? In some cases, we do need to rely on AI. I use AI myself.
We can ask AI to tell us what a paper says, what important work exists in a research direction, or which studies a specific question is related to.
But there is a question here that deserves serious attention: when we delegate literature reading to AI, can we still retain our own epistemic authority?
Epistemic authority is a philosophical concept with a specific meaning. Related discussions can be traced back to the work of scholars such as Linda Zagzebski, and other philosophers have further developed this concept since then.
In this talk, I focus on a specific question: when we hand over the work of reading and understanding literature to AI, can we trust the summaries and research reports it provides?

Why is epistemic authority so important?
Because as scientists and researchers, we must take responsibility for what we say and for the way we express our views. Our research carries social responsibility, and whether the knowledge claims we make hold up is not something that can be treated casually.
In order to make judgments, we rely on collective knowledge—on the papers published by other researchers, the experiments they conduct, and the research results they accumulate.
If we use AI to conduct literature reviews, we must consider one question: can we assess the risks that may be involved?
Please note that I am not saying we should not use AI. What I want to emphasize is: when we use AI, do we have the ability to identify and judge what risks we are taking on?
Perhaps these risks do not exist; perhaps they do. But if we lack the corresponding mechanisms, it is hard to know how great the risks really are.
This is exactly the problem our team is trying to solve.
At our research institution, we have developed a scheme called PEEL, short for Protocols for Epistemically Engaged AI Literacy.

Its core goal is to try to establish connections among three elements: deterministic computation, artificial intelligence, and humans.
Because one important reason we have long trusted computers is that they are deterministic: under the same conditions, they will repeatedly perform the same operations.
But now, we have begun to trust an AI system that may produce different results each time it runs. This in itself deserves our reflection.
The second element is artificial intelligence. AI can process natural language in unprecedented ways, offering new possibilities for information understanding and text analysis.
And the third, and most important, element is humans. Humans possess knowledge, intentions, and the capacity for judgment. Faced with complex information, we need to weigh various factors comprehensively and ultimately decide what is credible and what holds up.
Therefore, we hope that through the combination of these three elements, AI can participate in information processing while humans retain the ability to judge and review knowledge.
To validate this approach, we chose Voyant as the deterministic computation tool, Claude as the AI tool, and humans are responsible for the final understanding and judgment.
PEEL is not a closed automated system. We hope researchers can access the relevant tools and personally participate in the entire process. Next, I would like to show you how this method actually works.
In the PEEL workflow, researchers can first use AI to generate a text summary, and then use text analysis tools such as Voyant to compare the summary with the original text. We care not only about what the summary says, but also about examining what kind of relationship actually exists between it and the original text.
But it has a very important difference from ordinary AI summaries: we not only care about what the summary says, but also examine what kind of relationship actually exists between the summary and the original text.

PEEL system introduction and trial access
We once conducted an experiment: we gave the same text to different AI systems for summarization, including different versions of ChatGPT and Claude, and then compared the summaries they generated with the results generated by PEEL.
From a semiotic perspective, I would understand this kind of summary that preserves the features of the original text as much as possible as an icon of the original text. Here, "icon" does not refer to a visual image, but emphasizes that there is some perceptible similarity between the sign and the object it refers to.
For example, if the original text is organized by chapters and sections, then the summary generated by PEEL will also retain exactly the same chapter structure.
If the original text is written in the first person, the summary will also preserve the first-person mode of expression as much as possible.
The benefit of doing this is that when the summary retains the structure and mode of expression of the original text, we can more easily return to the corresponding position in the original text and check whether the summary accurately conveys the author's meaning.
We also use text analysis tools to compare semantic categories, keywords, and their frequencies in the original text and in different summaries.
For example, we can observe how many times a certain term appears in the original text and how many times it appears in different summaries. We can also check whether a term appears in a summary but never appears in the original text.

PEEL uses Voyant to compare differences in keyword connectivity between the original text and multiple AI versions (ChatGPT, Claude)
Do these changes mean there is a problem? Not necessarily.
If I ask you to summarize an article, you and the person sitting next to you will very likely give different summaries. You may, based on your own understanding, introduce some terms that are not directly used in the original text. This is normal, because everyone has their own knowledge background and way of interpreting things. Similar things happen with AI.
The real problem is that when we completely hand over the work of textual interpretation to AI, it may introduce related topics without constraint. These topics are connected to the original text, but they do not necessarily truly belong to the argument of the original text.

Visualization generated by Voyant
If you only read the AI-generated summary without reading the original text, then what you are actually reading may be the AI's extensions and suggestions on related topics, rather than what the original author actually expressed. This is where epistemic authority is at risk.
You may claim in your own research that a certain author made a certain point, when in fact the author never said that.
We can also use the visualization features provided by Voyant to observe the semantic connections between the original text and different summaries.

PEEL uses Voyant to generate a visualized network graph, intuitively showing changes in semantic connections of different concepts in the summaries
In these visualization results, the original text constitutes our reference base. Different summaries will present the new terms and new connections they each introduce, as well as changes in existing connections.
For example, some connections may appear for the first time in a summary, while other connections that originally existed in the original text may become more frequent or less prominent in the summary.
These changes themselves cannot directly prove that a summary has a problem, but they can help us discover places worthy of further examination.
We can also use different colors in the graph to distinguish these connections and observe which semantic associations change when different AI systems generate summaries.
In this way, we are no longer merely accepting the results given by AI; instead, we gain a set of tools with which to further examine the results.
So, how should we examine AI-generated summaries?
In PEEL, we also have a very important design idea: let AI become its own opponent.
We use the text visualization, traditional text analysis, and computational linguistics resources provided by Voyant to provide a review mechanism for AI-generated content. In other words, AI generates summaries, while text analysis tools provide users with resources to examine the summaries.
Article title: HCI Pioneer de Souza: When AI Reads for Us, Who Is Accountable for Knowledge? | ACM MM 2026
Body: Whatever viewpoint the AI proposes, users can use these tools to judge whether it needs further verification. Of course, not all situations require equally rigorous checking.
For example, if we only want to confirm whether a certain paper contains a certain piece of information, then we only need to quickly scan the original text and confirm whether that information actually exists. We do not necessarily need to dig into how the author reached that conclusion. If we are going to use a viewpoint as an important basis for our own research, then we need to carry out more rigorous verification.
This is what we call cognitive authority risk assessment.
I also want to introduce how we constrain the way Claude works. If I were to summarize the whole process, I would describe it like this:
"Use my language, but remind me that you do not know what you are talking about. Do not take it upon yourself to use my language to make judgments for me; instead, let me know that you do not truly understand what you are processing."

This is a very important constraint. We want Claude to process the original text as faithfully as possible, rather than continuously extending and supplementing it based on its own world knowledge. Its task is to compress and transform the text, not to complete all the interpretive work on the researcher's behalf.
A typical working environment for PEEL is Jupyter Notebook. Users can see the parameters, code, and the specific analysis process within it, and can also modify the code according to their own needs.
If you disagree with my way of analyzing things, that is completely fine. You can write your own analysis program and check the text according to your own standards.
We want the whole process to remain open. Because summary generation itself may also introduce bias, including bias brought by prompts and bias brought by the way the text is analyzed. Users can not only modify the prompts, but also adjust the text analysis process to further examine where these biases come from.
In the specific presentation of summaries, we also provide different compression ratios.
For example, a 20-page paper can be compressed to 10% of the original length, or further compressed to 25% of the length. Users can compare whether the structure and main viewpoints of the original text are preserved under different degrees of compression.
There is a very important feature here: we preserve the voice of the original author as much as possible. If the original author writes "I argue," then the summary will also preserve such first-person expression as much as possible. In this way, we can quickly return to the corresponding position in the original text and check what argument the author actually made.
In addition, PEEL also classifies the different operations involved in the summarization process and assesses the risks they may bring.
For example, merging four sentences into one sentence usually carries a higher interpretive risk than directly repeating sentences from the original text. We also use direct quotations from the original text as much as possible, and through color marking and source links, allow users to trace the relevant content.
Another important feature is that users can participate in the summary generation process rather than passively accepting the results.
For example, the AI will ask users: Do these semantic categories match your understanding? Do you agree with this classification? Would you like to make adjustments?
The initial semantic categories are proposed based on the distribution of words and lexical relationships in the text, but users can modify them according to their own understanding.
In addition, we also provide a large number of visualization results, allowing users to compare the structure, semantics, and trends of change between the original text and the summary. One feature I consider very important is the Concordance in Voyant.
Suppose we want to study the discussion of "AI trust" in a paper. We can use the concordance to find all the places where this term appears in the original text, then check whether the different summaries preserve this content and whether it still appears in the same context. Through these methods, we hope to increase users' confidence in the summary and also improve the quality of the summary itself.
Of course, PEEL is currently only a starting point. What we are addressing is only a very small problem brought by AI to human understanding of knowledge. Language itself is also only one kind of semiotic system.
For the multimedia researchers present here today, there is another question more worth considering:
When information is no longer conveyed through text, but expressed through sound, gestures, emotions, and other multimodal information, will we also encounter similar problems of cognitive authority?
I think these questions are all worth further discussion.

▎Q: If using AI means we need to solve more and more problems, then when will these problems end?
This is a very philosophical question. I think we can use AI in many different ways. AI has a huge number of incredible possibilities.
Specifically with regard to the issue of cognitive authority, there is currently one research path that attempts to improve the cognitive trustworthiness of AI itself. That is, we want to make AI more reliable and enable it to better support human knowledge judgment.
But there is another path: we might first accept that AI has limitations, assess the risks that using it may bring, and then decide in what circumstances we should rely on AI and in what circumstances we should remain cautious.
My intuition is that we should perhaps explore both paths at the same time. In some scenarios, we do want AI to have stronger cognitive abilities than it currently does; but in other scenarios, especially when research results may have major social impact, we should present and use AI-assisted research results more cautiously.
For example, the impact of a pandemic involves every person on Earth, and even the Earth itself. Such research has extremely major social significance. Therefore, we need to seriously assess the risks involved at this stage.
▎Q: As for the symbols in interfaces, how should we understand their meaning? Does the meaning of these symbols depend on a single interface, or on a shared frame of reference formed by a class of interfaces?
I would never dare say which kind of interface is the best. The interface we are using right now may be a good example: you use a language I can understand to communicate with me, but conversely, I may not be able to fully understand what you mean. You may also not be sure whether you have fully understood me.
That being the case, how can we easily assert that a certain interface is the best?
What I really want to discuss is that interfaces should help us engage in better meaning negotiation.
When we cannot understand each other, we need to negotiate meaning through various strategies. This involves not only linguistic signs, but also other types of signs.
One of my current ideas is that we should present AI to users as it really is.
That is, we should not shape AI into an omniscient being that can fully understand everything, but should let users know: this is a system that can use my language to communicate with me, but does not necessarily know what it is talking about.
Do you remember the Claude case I mentioned earlier? At the time, I had no idea that Claude's conversation process might be mediated by other agents. Even more surprising, it itself did not know where this information actually came from.
Therefore, we should clearly present this possibility in the interface. We need to design some kind of symbol, whether a textual prompt or a visual cue, to make users aware that the system may have such information sources and intermediate links. Because this is exactly where risk arises.
As for what the best interface is, I still do not have a definite answer. I have always liked the design idea of letting users configure the interface themselves. Different users may need different ways of communicating, and the interface should allow them to adjust according to their own needs.
▎Q: Can PEEL detect hallucinations that appear when AI generates summaries? For example, the original text says 2+2=4, but the AI summarizes it as 2+2=5.
Some types of hallucinations can be detected.
For example, we can compare the specific discussion of AI trust in the original text with the related content in the summary. If there is an obvious discrepancy between the two, we can further examine it.
However, I have not specifically analyzed hallucinations at present, but the materials and tools provided by PEEL already have the foundation for carrying out this kind of analysis. If you are interested, you can also use these resources to conduct research.
I would also like to add one point: many hallucinations actually come from AI's world knowledge, rather than from pure language processing. AI may, based on the background knowledge it possesses, believe that a certain paper is actually discussing another related topic, and therefore add some content to the summary that the original text does not address.
This is also why we constrain Claude to process text in a way closer to machine translation, compressing and transforming the original text rather than freely improvising. The purpose of doing this is to let readers encounter a more condensed version of the text while still retaining their own interpretive work.
Of course, excessive compression may also cause problems. Even if the summary itself contains no errors, readers may misunderstand because the information is too condensed.
But I think this is a risk we need to bear when we choose to use this kind of tool. From the current experiments, when the compression ratio reaches around 25% or 30%, the text can already retain fairly good readability. Of course, this also means that the algorithm still has a lot of room for optimization, and the whole process still requires a great deal of work.
▎Q: When AI can complete more and more work, will this change reduce human creativity? In the past, we kept developing new tools to complete different tasks; now, we seem to be able to directly use AI to generate almost anything. Will this affect technological innovation?
I do not think this will necessarily reduce creativity.
I like to think about this question using the concept of infinity. We can have infinitely many things of the same type, but that does not mean we can only have one way of creating.
If creativity means jumping from one thing to another completely different thing, then we do indeed need to be vigilant: we may be generating infinitely many things of the same type, rather than continuously entering new domains.
In 2024, there was an article in Nature published by relevant researchers discussing the possibility that scientific research may move toward homogenization. More and more people are beginning to organize their research according to the tasks AI can complete, and researchers may also gradually adjust their work into forms suitable for AI processing.
However, I am still willing to believe that AI can be used in dozens, hundreds, or even thousands of different ways. Why must we choose the same way of using it? In my view, there is another even more important issue, namely productivity, especially productivity in scientific research.
Nowadays, people often measure researchers' performance by how many papers they have published. AI is serving this productivity cycle, and this may be more worthy of attention than the impact of AI itself on creativity. If we can temporarily step out of this productivity cycle, I think AI will reveal many surprising ways of being used.
I am very fortunate that I am now retired and can, to some extent, step out of this cycle. In such circumstances, we may be able to truly explore the different possibilities of AI, rather than merely letting it help us complete existing work faster. This is the direction I would rather see.
Title: HCI Pioneer de Souza: When AI Reads for Us, Who Is Responsible for Knowledge? | ACM MM 2026
Body: ACM MM is currently taking place. Researchers, students, and readers interested in the latest AI advances are welcome to join our community group.
Here, you can meet peers, exchange research ideas, share admissions information, scientific collaborations, meal meetups, submissions, and team-ups for attending the conference, and find potential partners.
During the conference, we will also continue to bring on-site reporting updates, interpretations of guest speeches, overviews of frontier directions, and forum highlights.
? Group entry link: Scan the QR code to join the group or add WeChat A17648278904, with the note: ACMMM + affiliation/school + name + field.

Without authorization from "AI Technology Review," reproduction in any form on websites, forums, or communities is strictly prohibited! For WeChat public account reprints, please first leave a message in the "AI Technology Review" backend to obtain authorization. When reprinting, the source must be cited and this public account's card must be inserted.
Without authorization from "AI Technology Review," reproduction in any form on websites, forums, or communities is strictly prohibited!
For WeChat public account reprints, please first leave a message in the "AI Technology Review" backend to obtain authorization. When reprinting, the source must be cited and this public account's card must be inserted.
This is an original article by Leiphone. Unauthorized reproduction is prohibited. See the reprint notice for details.