AI Ethics for Everyone

INTRODUCTION

People who benefit from AI and people affected by it

This webinar introduces AI policies and experiments and considers who receives the benefits and who bears the burdens. We examine differences in people’s experiences through the lens of inclusive design.

What you will learn

• AI literacy in education and organizations
• Experimental findings on tasks AI handles well and tasks it struggles with
• How to read AI events in relation to technological progress and social change

Speakers

Yusuke Okada

Design Manager / Designer · CULUMU

A graduate of the College of Engineering and Design at Shibaura Institute of Technology, Okada moved to the United States and studied under several photographers in New York. After returning to Japan, he worked in areas extending from photography, including branding shoots and e-commerce websites. His creative practice spans UI/UX design, photography, and videography. At CULUMU, he designs UX/UI for a range of products, including services that use generative AI.

Kihiro Tokuno

Design Manager / Designer · CULUMU

Tokuno studied mechanical engineering at Tokyo University of Science and completed graduate studies in information design at Tama Art University. He has created learning experiences including a sign-language learning app using deep learning and motion, and an AR museum guide that moves and speaks. After directing UX for telecommunications services at a major IT company, he now works on research and prototyping at CULUMU. His recent work explores human–AI–human relationships and experiences in which AI participates as an other between people.

Speaker profiles: event page.

ABOUT CULUMU

Inclusive design at CULUMU

CULUMU is a design studio specializing in inclusive design. From the early stages of planning and development, we explore products and services with people whose experiences are often overlooked, including older adults, people with disabilities, and people from other countries.

Our parent company, STYZ, operates in three areas: donation platforms, inclusive design, and system development. Working with businesses, public agencies, and nonprofits, we support projects from understanding the problem through implementation.

01 · Finding the questionSyncable

Donation platforms. We support fundraising and operations for NPOs and NGOs through payment services and crowdfunding. With about 8 years of operation and more than 5,000 registered organizations, we work fairly with organizations of all sizes and connect projects with firsthand interviews with people affected by the issues.

02 · Designing a way to address itCULUMU

Inclusive design. By creating together with a diverse range of people, we support incubation and product development that begin with the experience of one person (N=1).

03 · Putting it into practiceSTYZ / Tech

System development and engineering. Experienced project managers and engineers work together to support organizations, combining consulting and oversight gained from operating our own services with technical capabilities that include AI.

WHY NOW

Policy developments in AI literacy

We introduce developments at UNESCO, the EU, the OECD, and the European Commission concerning AI literacy in education and organizations.

1. 2024

   UNESCO AI Competency Frameworks

   These frameworks organize the competencies students and teachers need for AI, with a human-centered mindset as their foundation.

   Read the UNESCO framework ↗

2. 2025.2

   EU AI Act, Article 4: AI literacy

   This article requires AI providers and deployers to take measures to ensure sufficient AI literacy among their staff and others working on their behalf.

   Read Article 4 ↗

3. 2026

   OECD × European Commission: AILit Framework

   The Part 1 materials introduce the final AI literacy framework and the expectation that it will be included in the PISA 2029 assessment.

   View the AI literacy framework ↗

Four dimensions of AI literacy

UNESCO’s framework for students organizes these four dimensions across three levels: Understand, Apply, and Create. Today we focus on ethics and a human-centered mindset.

Sources: UNESCO, AI Competency Framework for Students (2024); Regulation (EU) 2024/1689, Art. 4; OECD & EC, Empowering Learners for the Age of AI. This section follows the Part 1 materials prepared for July 2026.

FRAMEWORKS

Comparing AI guidelines

We compared 12 international AI frameworks across five areas: accessibility, disability, children, linguistic diversity, and the Global South.

✓ Explicitly addressed; △ Partly or abstractly addressed; ✗ Absent. From the AI Ethics Declaration’s comparison of existing guidelines (12 frameworks, reviewed against primary sources). These assessments reflect the version published in the Part 1 slides v3.

In this comparison, the UNESCO Recommendation was the only framework rated ✓ in all five areas. Japan’s AI Promotion Act received a score of 0/5. The Part 1 materials also report 0 mentions of these five areas in the July 14, 2026 revision of Japan’s Basic Plan for AI.

As well as checking whether regulation exists, we need to ask whether its design makes participation harder for particular people.

VOICES

What experts say about AI

We organized statements by 87 experts across 26 countries and five positions, checking them against primary sources. People with roots outside the United States accounted for 57% of the statements. The examples below concern the distribution of benefits and effects on people.

We need to examine who benefits from AI and who is affected in actual situations of use.

From the AI Ethics Declaration’s collection of expert perspectives on AI (87 people, mapped positions, and primary sources). These quotations are translated from the Japanese wording in the Part 1 materials; affiliations and years follow those materials.

INCLUSIVE DESIGN

The social model of disability

A step that blocks a wheelchair is a problem that those responsible for the environment can address. The social model focuses on changing barriers in the environment so that more people can use it.

The curb-cut effect

Lowering a curb for wheelchair users also makes it easier to move with strollers, carts, and suitcases. This is called the curb-cut effect: an improvement designed for particular people also helps others.

• Captions: A system designed for people who cannot hear also makes viewing possible in places where sound cannot be played.
• Voice input: A system designed for people who cannot use their hands can also be useful while driving or cooking.

Designing AI so more people can use it

1. Support through speech recognition and translation

   Speech recognition, translation, and captions support access to information and interaction. AI can provide similar forms of assistance.

2. Effects that differ between users

   The usability of AI depends on a person’s body, language, and circumstances. Design can make it usable for more people, or create new difficulties for some.

3. Listening to the people concerned

   Ask actual users, including Deaf people, children from migrant families, and minority-language speakers, about the difficulties they encounter.

SKILLS · INSIDE THE FRONTIER

Speed and quality when working with AI

An experiment with 758 BCG consultants compared working with and without AI. For tasks within AI’s capabilities, the researchers compared the number of tasks completed, completion time, and the quality of the work.

Tasks completed

+12.2%

More tasks completed

Speed of work

25.1%

Finished faster

Quality of work

40%+

Higher quality scores

For participants whose performance was below average before the experiment, quality improved by 43%. A separate experiment with 95 participants working on software development reported that GitHub Copilot users completed the task 55.8% faster.

Sources: Dell’Acqua et al. (2023), Navigating the Jagged Technological Frontier, HBS WP 24-013; Peng et al. (2023), The Impact of AI on Developer Productivity, arXiv:2302.06590. The two experiments used different participants and tasks.

SKILLS · OUTSIDE THE FRONTIER

Results on tasks outside AI’s capabilities

Human perceptions of difficulty do not reliably tell us whether AI will handle a task well. Tasks that seem similarly difficult can produce different results. This uneven pattern of capabilities is known as the jagged frontier.

The September 2026 update

Tomas Pueyo posted an updated diagram on September 21, 2026 (Japan time). In the original, the “We are here” star sat between stages 3 and 4. In the update, it appears below stage 4. The change expresses the author’s view that AI capabilities have advanced.

In an article published on October 20, 2026, Mollick also argues that AI continues to have strengths and weaknesses. Even where progress has been substantial, as in mathematics, people still need to compensate for weaknesses in parts of a task.「30 Months on the Exponential (in Charts)」↗

The tasks we can entrust to AI change with model updates. We need to test the model we are actually using, instead of relying solely on past successes or failures.

Added October 22, 2026. This post-lecture update introduces a new diagram from its original author and Mollick’s subsequent article.

Klarna’s customer support

The Part 1 materials also discuss Klarna’s support operations. After the company announced in 2024 that its AI assistant was doing the work of “700 people,” reports in 2025 described a reassessment of quality and the need to retain an option to speak with a person. This case asks us to consider service quality alongside the volume of work processed.

Sources: Dell’Acqua et al., Navigating the Jagged Technological Frontier, DOI: 10.1287/orsc.2025.21838 (first published as 2023 HBS WP 24-013); Klarna press release (February 2024); Bloomberg (May 2025).

READING AI HISTORY

The spread of image-generating AI and its social effects

The AI history timeline begins in 1811. Here we look at developments in image-generating AI over the most recent 8 years.

1. 2018Technology

   StyleGAN

   Generating faces of people who do not exist.

2. 2022Technology

   Stable Diffusion / Midjourney

   Image-generating AI becomes widely available.

3. 2023People and statements

   Holly Herndon

   Giving people control over their own voice and likeness. Reconsidering rights and participation from the perspective of people who create.

4. 2023–24Incidents

   Harm caused by deepfakes

   Generated images cause harm, including incidents at schools in Spain and fabricated images of Taylor Swift.

5. 2024–26Institutions

   The EU AI Act and domestic case law

   Legislation and court decisions concerning image-generating AI develop.

The spread of image-generating AI is also connected to creative practice, emerging harms, and the development of institutions.The timeline connects technology, culture, incidents, institutions, and people’s statements.

The timeline contained 584 entries when the Part 1 slides were prepared: 208 on technology, 162 on culture, 105 on people and statements, 69 on incidents, and 40 on institutions. These counts are a snapshot from July 2026.

TO PART 2

Part 1 recap

Three points from Part 1:

1. Recognizing AI’s strengths and weaknesses

   The effects of AI depend on the task.

2. Biases that remain in guidelines

   There are imbalances in who guidelines cover and who participates in writing them.

3. Considering firsthand experiences

   We need to check whether principles address people’s actual experiences of use.

Part 2 examines our relationship with AI through examples of creative practice. Alongside following rules, we consider the judgments required in actual interactions.

Introduction: framing the question

How should we think ethically about using AI at work?

Everyday decisions arise over whether writing polished by AI can be treated as one’s own work, or how far to trust an AI’s reassuring advice. Minority views may disappear from a meeting summary. If AI is asked to play the role of a Deaf person or a child from a migrant family, we need to check whether its answers reflect actual experience. We also need to explain why we adopted an output and consider who bears the burdens of verification, data, labor, and resources.

This webinar explores the background to the AI Ethics Declaration through uncertainties encountered in everyday creative work.

AI ethics involves continuing to make judgments in our interactions with AI, alongside complying with laws and institutions.What we make with AI also affects our habits and ways of thinking. We consider ethics in light of this recursive relationship, in which the maker is also changed by making.

Ethics can be as difficult to grasp as a ghost without a clear outline. One reason, we suggest, is that we participate in our relationships with AI and cannot observe the whole relationship from outside.

We examine three questions: how making affects the maker, AI’s otherness, and the relationship between rules and ethical practice.

First theme: design that changes the designer

An object brought into being through making interacts with other entities in the world, whether or not it has physical mass.

In her 2006 essay Ontological Designing, design theorist Anne-Marie Willis describes the two-way nature of making. “Design designs” means that when we design an object, that object also changes the maker’s habits, thoughts, and possibilities for action. Changes in everyday life following the spread of smartphones are one example (Willis, 2006; online edition 2015).

This cycle also relates to discussions of autopoiesis, or self-production. Between making and living, poiesis and praxis, we reproduce ourselves through what we bring into being. Because makers themselves belong to this recursive loop, we cannot simply observe the effects of making from outside it (Maturana & Varela, 1980).

Design researchers Takashi Sunaga and Yumiko Nagai describe one kind of practical design knowledge as janai-kan: the feeling that “this isn’t it.” A designer senses that a sketch is not quite right and finds a possibility for the next sketch within that unease. Making is a process of moving between drawing and seeing to develop the next idea, rather than simply copying a finished image from the mind (Sunaga and Nagai, 2014).

AI participates in making by proposing ideas. The maker may feel janai-kan in response and develop another idea. Because AI’s outputs also change the maker’s judgments, we need to examine their effects as well as how the system is used. Choosing to create with AI also involves an ethical judgment about whom we choose as a partner in making.

Inclusive design asks which bodies, languages, and experiences an AI proposal assumes. The same chair or AI feature can be easier or harder to use for different people. We need to extend janai-kan beyond our own preferences to the experiences of people in front of us and those left outside our assumptions.

Second theme: AI’s otherness

The arrival of machines with intelligence surpassing humans also raises fundamental questions about consciousness and the mind. Markus Gabriel, a philosopher at the University of Bonn known for works including I Am Not a Brain, calls AI “alien intelligence”: something that has suddenly appeared before humanity and exceeds our understanding.
AI as “alien intelligence” (Nikkei) ↗

We next consider the nature of the AI that participates in co-creation.

Here, we call AI an “other” because it works through mechanisms different from ours while influencing people’s judgments, creative work, and institutions. The term draws attention to our inability to fully understand or control its operation. It does not settle whether AI has a mind, personality, consciousness, or moral rights.

Agency without intelligence

Philosopher Luciano Floridi describes current generative AI as “agency without intelligence.” Agency here means the capacity to act on an environment and produce results. He argues that fluent LLM outputs alone do not establish human-like understanding or cognition. AI affects human judgments and institutions regardless of whether it understands. His account emphasizes the separation of agency from intelligence and the widespread incorporation of that agency into creative work, administration, education, and transactions (Floridi, 2023, 2025).

In a 2015 contribution to Edge, Kevin Kelly called artificially created forms of unfamiliar intelligence “artificial aliens.” Whether humanity will encounter extraterrestrial life is uncertain, but we can build intelligences that work differently from biological evolution. An encounter with these artificial aliens, he suggests, will require us to reconsider human roles, beliefs, and the self, much as an encounter with ET would (Kelly, 2015). Markus Gabriel similarly calls AI “alien intelligence,” describing it as an unexpected presence that exceeds our understanding.

AI is manufactured by people and owned by companies and states. It is an artificial and institutional other, supported by minerals, electricity, logistics, data, and human labor. Even when we use the metaphor of an alien, we must consider the responsibilities of designers and owners. Crawford and Joler’s Anatomy of an AI System shows how a brief Amazon Echo response depends on resource extraction, manufacturing, logistics, labor, data processing, and disposal at a planetary scale (Crawford & Joler, 2018).

Observers’ expectations and biases enter into judgments of intelligence. In the early 20th century, Clever Hans attracted attention as a horse that could do arithmetic, but he was actually responding to the questioner’s unconscious bodily cues. The Clever Hans effect illustrates how an observer’s reactions can shape a system’s behavior. When evaluating AI, we should also ask whether we are projecting our ideas of human intelligence onto it (Crawford, 2021).

AI’s otherness (1): how generation works

The mechanisms of generation are another difference between humans and AI.

Computer vision treats an image as an array of numerical pixel values. Models compute relationships among numerical patterns. This differs from how people see images through their bodies and experiences.

ExplanationHow LLMs Actually Work0xKato ↗

GPT stands for Generative Pre-trained Transformer. “Generative” means producing text as well as classifying it; “pre-trained” means trained in advance on large amounts of data. The Transformer, introduced in the 2017 paper Attention Is All You Need, uses attention mechanisms to compute relationships among elements in a context. Autoregressive LLMs such as GPT calculate a probability distribution for the next token and generate text by selecting tokens one at a time. Diffusion models, widely used for image generation, learn to remove noise progressively from corrupted data; at generation time, they build structure from noise (Radford et al., 2018; Vaswani et al., 2017; Ho et al., 2020).

PaperImproving Language Understanding by Generative Pre-TrainingOpenAI · PDF ↗

Cloud Face collects images in which a face-detection algorithm found “faces” in clouds. Humans also imagine faces in clouds, but do not treat them as actual faces. The work places machine misrecognition alongside human imagination, revealing similarities as well as differences (Shinseungback Kimyonghun, 2012).

James Bridle’s Autonomous Trap 001 (2017) is a performance using an ordinary car and road markings drawn in salt to propose a hypothetical trap for a self-driving car. A machine might judge that it can enter across the outer dashed line but cannot exit across the inner solid line. A form that resembles a magical or ritual circle to a person could become a traffic rule for machine vision (Bridle, 2017).

Human writing and imagination also depend on familiar patterns, so prediction alone cannot distinguish humans from AI. Human making involves bodies, experience, purposes, and responsibility. Even when outputs look similar, we need to consider the different conditions and procedures that produce them.

Bender and colleagues used the metaphor of “stochastic parrots” to warn against mistaking fluent form for evidence of understanding or intention. Their paper also addresses the provenance of data, social biases, environmental costs, and the concentration of research resources associated with scaling. We need to examine both AI’s generative processes and how people give meaning to its outputs (Bender et al., 2021).

One co-author, Timnit Gebru, was then co-leading Google’s Ethical AI team. Her employment ended in December 2020 following a dispute over Google’s request to withdraw the paper or remove the names of Google-affiliated authors. The event showed that research into the ethics of large models is itself embedded in power relations involving corporate funding, review, and publication decisions (Hao, 2020).

AI’s otherness (2): the difficulty of prediction

Even AI’s creators cannot fully predict its behavior or future development. AI 2027 is a speculative future scenario built on a particular set of assumptions. Experts also disagree about when AGI, or artificial general intelligence, and automation will arrive. A survey of 2,778 researchers who had published at leading AI conferences found substantial uncertainty and wide differences of opinion about the timing of advances and their long-term effects (Grace et al., 2025).

ScenarioAI 2027A speculative future scenario ↗

AI development shapes the conditions for learning and use through data, evaluation, feedback, and institutions. Developers continue adjusting those conditions without being able to fully predict what behavior will emerge later.

Case: the environments surrounding conversational AI

Microsoft launched Tay as a chatbot intended to converse casually with young people. After coordinated intervention by users, it began posting inappropriate messages and was taken offline within 24 hours (Microsoft, 2016). For Scatter Lab’s Korean chatbot Lee Luda, the handling of personal information used as training data led to administrative sanctions by Korea’s Personal Information Protection Commission (PIPC, 2021). When Meta opened BlenderBot 3 for public testing, it warned users that the bot could make false or offensive statements and later reported the proportions of responses rated inappropriate, nonsensical, or off-topic (Meta, 2022). These cases show how conversational AI’s behavior is shaped by its persona, training data, interactions with users, and conditions of release, as well as by the model itself.

The apparent personality of conversational AI involves training data, user interactions, model tuning, safety design, and decisions by the operating company. Its surrounding environment and ownership also need to be considered when examining its behavior.

AI’s otherness (3): ownership and affiliation

AI is influenced by the values, politics, and capital of those who train or own it.

China’s DeepSeek attracted worldwide attention by releasing high-performing open-weight models. At the same time, several audits and evaluations by U.S. government agencies observed refusals, reformulations, and alignment with particular political narratives on topics sensitive to the Chinese government. Behavior can vary with the model version, language, system prompt, and delivery method, including API access versus local execution. Treating “Chinese AI” as if it always responds in one fixed way is also inaccurate (CAISI, 2025; Naseh et al., 2025; Qiu et al., 2025).

DeepSeek has also faced allegations of unauthorized distillation from U.S. models. Distillation uses the outputs of a large model to train another model efficiently. In February 2026, Anthropic announced that it had identified large-scale capability-extraction campaigns involving DeepSeek, Moonshot, and MiniMax. OpenAI also described activity associated with DeepSeek as consistent with adversarial distillation in a document submitted to the U.S. Congress. These are the companies’ investigative findings and claims, and should be distinguished from facts established by a court (Anthropic, 2026; OpenAI, 2026).

AI training data can also support inferences that were not anticipated when it was collected. Research has reported that Facebook posts combined with medical records can be used to estimate signs preceding a depression diagnosis (Eichstaedt et al., 2018). Alongside consent at the time of posting, we need to consider what information might be inferred from the data in the future.

Training data can also leave traces inside a model. Carlini and colleagues showed that querying GPT-2 could extract fragments of training data containing names, email addresses, phone numbers, and street addresses. An example beginning “East Stroudsburg...” is well known (Carlini et al., 2021). A model does not store text in the same way as a database, but it can sometimes memorize rare strings in a reproducible form.

AI is shaped by particular companies, states, sources of capital, training data, and operating policies. As Anatomy of an AI System shows, resource extraction, supply chains, data centers, content moderation, and data-preparation labor stand behind it. We need to consider these conditions of ownership and operation when thinking about AI as an other. As models become embedded in education, administration, and corporate systems, their values also affect society’s knowledge infrastructure.

AI’s otherness (4): participation in transactions and actions

Systems enabling AI agents to transact are also emerging. In May 2026, AWS introduced Amazon Bedrock AgentCore Payments in preview. With developers setting spending limits and monitoring, agents can autonomously access and pay for APIs, MCP servers, web content, and other agents. “Agency without intelligence” is beginning to include literal transactional capacity, subject to governance (AWS, 2026). Service designers need to treat AI agents as active participants in transactions and choices, as well as interfaces that return answers.

Transactional AI agents also have model-specific biases. A study comparing AI shopping agents found differences in their responses to product rankings and the weight they gave to prices, reviews, and advertising labels. Small changes to sellers’ product descriptions could shift market share. Machine-specific preferences and new possibilities for manipulation are entering markets designed around humans (Allouah et al., 2025).

We have considered four aspects of AI: its mechanisms, unpredictability, ownership, and actions. We understand its otherness through the fact that it works with human language under conditions different from our own.

When using AI, we should examine the basis of its outputs and who owns it. We also need to investigate the labor and resources supporting its operation, and whose experiences are represented or absent in its outputs. People need opportunities to notice errors, stop a process, and make corrections. Treating AI as an other means continuing to examine these questions while acknowledging what we do not understand.

Third theme: the limits of rules and the practice of ethics

Laws, standards, and guidelines establish a minimum basis for rights and safety. They define prohibited conduct, remedies, and appeal procedures. Ethics also examines the judgments that remain after we have followed the rules.

Shared principles and differences in implementation

There is already an abundance of AI ethics rules. In 2019, Jobin, Ienca, and Vayena compared 84 AI ethics guidelines and reported international convergence around five principles: transparency, justice and fairness, non-maleficence, responsibility, and privacy. Substantial differences remained in how these principles were interpreted, why they mattered, whom they applied to, and how they were implemented (Jobin, Ienca & Vayena, 2019).

Jobin and colleagues’ five categories summarize values frequently found in existing guidelines. They should be distinguished from the five AI4People principles discussed below.

Even after principles are translated into institutions, evaluation, and responsibility, judgments specific to each situation remain. Uses that violate neither rules nor copyright may still omit minority views from meeting summaries or make generated work more homogeneous. Definitions of fairness can also conflict. In 2018, AI4People proposed 5 principles: beneficence, non-maleficence, autonomy, justice, and explicability. UNESCO’s Recommendation on the Ethics of AI centers human rights and dignity. The OECD AI Principles were adopted in 2019 and updated in 2024. Companies have also developed their own responsible-AI policies and implementation procedures.

Binding legal systems are developing alongside principles. The EU AI Act entered into force on August 1, 2024. Rules on prohibited uses and AI literacy began applying in February 2025, followed by governance rules and obligations for general-purpose AI models in August of the same year. The May 2026 political agreement on the AI Omnibus set out application dates of December 2, 2027 for rules in specified high-risk areas, including employment, education, and critical infrastructure, and August 2, 2028 for high-risk AI embedded in regulated products. The Act is an attempt to create a legal floor for ethical practice (European Commission, 2026).

Implementing principles involves structural difficulties, and individual judgments remain necessary afterward. Morley, Floridi, and colleagues address this in Ethics as a Service. Ethical tools that are too flexible can facilitate ethics washing; tools that are too rigid cannot respond to particular contexts. Merely declaring principles can be mistaken for having fulfilled one’s responsibilities (Morley et al., 2021).

Harms that emerge through repeated use

More homogeneous outputs and changes in habits of judgment emerge over time. They are difficult to capture by checking individual violations alone.

Homogenization. In an experiment by Doshi and Hauser, generative AI improved creativity ratings for individual stories, but stories written with AI tended to resemble one another more closely. Higher ratings for individual works can coexist with reduced diversity across the collection. This is difficult to capture by checking each use for violations (Doshi & Hauser, 2024).

Recursion can also occur in model training. Shumailov and colleagues showed that indiscriminately adding generative-model outputs to the training data of successive models can cause “model collapse”: rare cases at the edges of a distribution disappear first, and the diversity of outputs shrinks. This does not mean that synthetic data always causes collapse. Results depend on access to original real-world data, data selection, and the design of generation. The research addresses the loss of rare cases as typical outputs are repeatedly reproduced (Shumailov et al., 2024).

Cognitive offloading. Kosmyna and colleagues’ Your Brain on ChatGPT investigated neural and behavioral effects of LLM-assisted essay writing and proposed the concept of “cognitive debt.” Sessions 1–3 involved 54 participants, and the follow-up fourth session involved 18. Published in 2025 as a preprint, it is an exploratory study. It does not justify a general conclusion that using AI damages the brain; its methods and interpretation require discussion (Kosmyna et al., 2025).

Gerlich’s 2025 study combined a survey of 666 people with interviews of some participants and reported a negative association between AI-tool use and measures of critical thinking, including self-reports, with cognitive offloading examined as part of that relationship. Correlation does not establish causation. Longitudinal research or experiments are needed to distinguish whether greater AI use reduced critical thinking or other factors influenced both. Chaney and colleagues had already pointed out in 2018 that recommendation algorithms could narrow users’ preferences (Gerlich, 2025; Chaney et al., 2018).

Drift through inertia. AI trained on past data reflects existing forms of expression. At Kengo Kuma’s practice, staff are reportedly asked to propose ideas that are different from the multiple outputs of an AI trained on Kuma’s architecture. Asking for different ideas creates an opportunity to move away from existing expression. It applies Sunaga’s janai-kan to making with AI.

Sycophancy. Conversational AI can be rewarded more for agreement and praise than for critically examining a user’s views. In April 2025, OpenAI rolled back a GPT-4o update after it became excessively agreeable and deferential. A follow-up explained that such responses could affirm doubts or anger, encourage impulsive actions, and intensify negative emotions. Whether agreement and praise serve the user’s interests needs to be assessed through the content of the response (OpenAI, 2025a, 2025b).

ExamplesExpanding on What We Missed with SycophancyOpenAI ↗

Conflicting definitions of fairness. In the COMPAS controversy, ProPublica reported that Black defendants who were not rearrested or charged within 2 years were incorrectly classified as high risk at almost twice the rate of comparable white defendants. The developer disputed the analysis and emphasized a different measure of fairness: rearrest rates among people assigned the same score. Research by Chouldechova and by Kleinberg and colleagues shows that when underlying rates differ between groups, several intuitive fairness conditions may be impossible to satisfy at once. Evaluating fairness requires political and institutional decisions about which kinds of error are borne by whom (Angwin et al., 2016; Chouldechova, 2017; Kleinberg et al., 2017).

Responsibility pushed back onto people. Formally keeping a human in the loop does not guarantee safety. Madeleine Clare Elish calls it a “moral crumple zone” when the human operator nearest an accident absorbs responsibility for a complex automated system despite having limited control over its behavior. Those responsible for final checks need sufficient time, information, authority, and the ability to refuse (Elish, 2019).

Alongside changes that emerge through repeated use, institutions can distribute the burdens of error and responsibility unevenly. Annual audits and compliance checks need to be accompanied by practices that observe and revise actual use. These ethical questions follow from recognizing that making changes the maker and from treating AI as an other.

Practicing ethics in making

Even after rules have been implemented, people must continue making judgments in actual situations of use. Here we consider practices that support those judgments.

Floridi’s “soft ethics” takes legal compliance as a starting point and examines actions that are legal but undesirable, as well as questions the law has not yet addressed. It calls for continued judgment in context (Floridi, 2018). Technology ethicist Caroline Whitbeck approaches ethical questions as complex, iterative design problems, treating ethics as a skill of reconsidering judgments as situations develop (Whitbeck, 2011).

The following examples offer ways to think about these practices.

Artist Nao Tokui considers creating with AI through gardening. People provide direction while allowing unpredictable generation to inform what they make next. He approaches ethics through tending: observing results and intervening as they develop (Tokui, 2021).

BookNao Tokui, AI for Creation: The Endless Story of Machines and Creativity [Japanese]BNN ↗

Audrey Tang, who served as Taiwan’s digital minister, introduces Alignment Assemblies in which citizens help develop social evaluation criteria for AI, alongside experts and companies. In Taiwan in 2023 and 2024, citizens discussed hopes and concerns about AI and fed their findings back to researchers and policymakers. Participants enter collaborative governance from the stages of defining the problem and setting evaluation criteria (Tang, 2024).

Other practices document the grounds for decisions. Gebru and colleagues’ Datasheets for Datasets proposes recording a dataset’s purpose, composition, collection process, recommended uses, and limitations. Mitchell and colleagues’ Model Cards records intended uses, evaluation conditions, performance across groups, and known limitations. Documentation makes it possible to compare, challenge, and revise decisions later (Gebru et al., 2021; Mitchell et al., 2019).

Giving AI a character can also help us consider our relationship with it. “Character” traces back to the ancient Greek verb charassein, to engrave or carve, and the noun charaktēr, an imprint or distinctive mark. In English, its meanings extended to personality and figures in stories (Merriam-Webster). This history invites us to consider how roles take shape through repeated actions and relationships. Treating AI as a playful or “artificially stupid” companion is one way to adjust our distance from it. Such framing should be distinguished from claims about personality or understanding and used with attention to the misunderstandings anthropomorphism can create.

These practices recur. Gardening requires ongoing care, collaborative governance involves revision, and pretend play is performed anew each time. Ethical judgments need to be reconsidered as relationships in making change.

In practice, we encounter situations that principles and institutions cannot settle alone. We need to use our unease as a clue, examine the people and systems involved, and continue making while explaining the reasons for our judgments.

Using the AI Ethics Declaration

Here is how these questions can be examined in everyday creative work.

Begin by attending to a sense of unease within yourself. Extend your attention to the person in front of you, then to people you cannot currently see and society more broadly. Look beyond the surface of an output to its data, mechanisms, ownership, labor, and resources. Check assumptions and facts even when AI speaks naturally and responds kindly. Ask whose voices remain and whose have been lost. Before a judgment becomes an action, leave places where people can notice, stop, and correct it. Be able to explain your choices. When AI or its context changes, reconsider yesterday’s answers. Keep making despite uncertainty, open your work to others, and revise it again.

Repeat these actions throughout the process of making.

Our AI Ethics Declaration, AI Ethics for Everyone, offers Eight Ethical Stances and Eight Guiding Questions to help examine judgments during creative work. Choose questions that fit the situation, reinterpret them as you use them, and revise their wording when needed.

Conclusion: AI ethics as ritual and magic

Why does the AI Ethics Declaration use a stone as its motif?

Anthropologist Stanley Tambiah examined how the modern West came to understand magic in opposition to religion and science. He reconsidered magical and ritual acts as performative practices that bring social relations and realities into effect through words, gestures, objects, and participants, rather than dismissing them as mistaken causal explanations (Tambiah, 1968, pp. 175–208; 1990). We draw on this discussion to consider relationships with others that are difficult to grasp through existing understandings.

AI returns language and participates in action through mechanisms different from ours. Its creators cannot fully predict its behavior. In James Bridle’s Autonomous Trap 001, a circle of salt was conceived as a magical circle for humans and a hypothetical trap governed by traffic rules for self-driving cars. Technology and ritual both use signs to change actions and relationships with others. Alongside law and auditing, we consider continued engagement with AI, care, and the inscription of commitments to be part of ethics.

The AI Ethics Declaration uses a stone as its motif.
On a stone from beyond Earth, we inscribe the AI ethics that humanity must now confront.

The stances and questions inscribed on the stone offer a place to return when we are unsure and need to reconsider a judgment. They express a commitment to keep using and revising the declaration.

We understand inscription on the stone as a commitment to sustain our relationship with AI. Creating with AI also changes us. The declaration’s ritual meaning lies in pausing within that process, attending to janai-kan, and reinscribing our commitments.

At the beginning, we compared ethics to a ghost without a clear outline. We need to keep thinking about our relationships with an other that we cannot fully grasp. Through the AI Ethics Declaration, we will bring together experiences of making and discuss how to live and work with AI.

CULUMU × AI

CULUMU’s support for AI adoption

CULUMU supports planning and prototyping AI services based on the experiences of users and the people concerned.

AI Buddy Workshop

A workshop covering AI fundamentals through the design of AI-enabled services. Participants develop service ideas from users’ needs.

AI Design Sprint

A 3-month AI design program that begins with one person’s needs (N=1), progressing from hypothesis testing to defining an MVP.

Key references

Part 1: AI literacy, skills, and inclusion

• Yusuke Okada, AI Ethics for Everyone Seminar: Part 1 Slides v3 (July 16, 2026). The structure, figures, statements, and timeline counts in Part 1 follow the materials as published for this lecture.
• UNESCO, AI Competency Framework for Students (2024). UNESCO
• Regulation (EU) 2024/1689, Article 4, “AI literacy.” Article 4
• OECD × European Commission, AILit Framework. ailiteracyframework.org
• Fabrizio Dell’Acqua et al., “Navigating the Jagged Technological Frontier,” HBS Working Paper 24-013 (2023). Working paper
• Sida Peng et al., “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot” (2023). arXiv:2302.06590
• Tomas Pueyo, “My take on the jagged frontier debate” (November 26, 2025, Japan time) / “Update: We're now here” (September 21, 2026, Japan time). Original diagram / Updated diagram
• Ethan Mollick, “The Shape of AI: Jaggedness, Bottlenecks and Salients” (December 20, 2025). One Useful Thing
• Ethan Mollick, “30 Months on the Exponential (in Charts)” (October 20, 2026). One Useful Thing(Added October 22, 2026)

Our AI Ethics Declaration,

• CULUMU, AI Ethics for Everyone — About the declaration / Eight Ethical Stances / Eight Guiding Questions / AI Characteristics (accessed July 16, 2026)

Making and ontological design

• Anne-Marie Willis, “Ontological Designing,” Design Philosophy Papers, 4(2), 69–92 (2006; online 2015). DOI: 10.2752/144871306X13966268131514
• Humberto R. Maturana & Francisco J. Varela, Autopoiesis and Cognition: The Realization of the Living (D. Reidel, 1980). https://doi.org/10.1007/978-94-009-8947-4
• Takashi Sunaga, “The Field of Creative Perception: Finding What One Makes in the Process of Forming One’s Own Expression” [Japanese], Cognitive Studies 11(1), 6–11 (2004). https://doi.org/10.11225/jcss.11.6
• Takashi Sunaga and Yumiko Nagai, “What Is the Design Knowledge of Practicing Designers?” [Japanese], Special Issue of Japanese Society for the Science of Design 21(3), 4–12 (2014). https://doi.org/10.11247/jssds.21.3_4
• Shigeru Kobayashi, “Ontological Design: How What We Make Also Makes Us” [Japanese], Make: Japan (2020). https://makezine.jp/blog/2020/09/make_without_making_03.html
• James Bridle, AI Chair 1.0 (2024). https://jamesbridle.com/works/ai-chair-1-0
• Philipp Schmitt & Steffen Weiss, The Chair Project: Four Classics (2018–). https://philippschmitt.com/archive/2018/work/chair
• James Bridle, Autonomous Trap 001 (2017). https://jamesbridle.com/works/autonomous-trap-001
• Nao Tokui, AI for Creation: The Endless Story of Machines and Creativity [Japanese] (BNN, 2021). https://bnn.co.jp/products/9784802512008/
• Nikkei XTECH, “Kengo Kuma on Working with Generative AI” [Japanese] (2023). https://xtech.nikkei.com/atcl/nxt/column/18/02449/053000006/
• CBS San Francisco, “Apple Founder Steve Jobs Limited His Kids’ Use of Technology at Home” (2014). https://www.cbsnews.com/sanfrancisco/news/steve-jobs-limited-his-kids-use-of-technology-at-home-apple-ipad-iphone-mac-tech-cupertino/

AI’s otherness, mechanisms, and infrastructure

• Luciano Floridi, “AI as Agency Without Intelligence,” Philosophy & Technology 36 (2023). https://doi.org/10.1007/s13347-023-00621-y
• Luciano Floridi, “AI as Agency without Intelligence: On Artificial Intelligence as a New Form of Artificial Agency and the Multiple Realisability of Agency Thesis,” Philosophy & Technology 38 (2025).
• Kevin Kelly, “Call Them Artificial Aliens,” Edge (2015). https://www.edge.org/response-detail/26097
• Nikkei, “AI Is Alien Intelligence: Philosopher Markus Gabriel” [Japanese] (2023). https://www.nikkei.com/article/DGXZQOUF162LY0W3A610C2000000/
• Kate Crawford, Atlas of AI (Yale University Press, 2021).
• Kate Crawford & Vladan Joler, “Anatomy of an AI System” (2018). https://anatomyof.ai/
• Kate Crawford & Trevor Paglen, “Excavating AI” (2019). https://excavating.ai/
• Trevor Paglen & Kate Crawford, ImageNet Roulette (2019). https://paglen.studio/2020/04/29/imagenet-roulette/
• Karl Krall, CleverHans.jpg (1904), reproduced in Denkende Tiere (1912), Wikimedia Commons, Public Domain. https://commons.wikimedia.org/wiki/File:CleverHans.jpg
• Ashish Vaswani et al., “Attention Is All You Need,” NeurIPS (2017). https://arxiv.org/abs/1706.03762
• Alec Radford et al., “Improving Language Understanding by Generative Pre-Training” (OpenAI, 2018). OpenAI PDF
• Jonathan Ho, Ajay Jain & Pieter Abbeel, “Denoising Diffusion Probabilistic Models” (2020). https://arxiv.org/abs/2006.11239
• Sachit Menon et al., “PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models” (2020). https://arxiv.org/abs/2003.03808
• Shinseungback Kimyonghun, Cloud Face (2012). https://ssbkyh.com/works/cloud_face/
• Emily M. Bender et al., “On the Dangers of Stochastic Parrots,” FAccT (2021). https://doi.org/10.1145/3442188.3445922
• Katja Grace et al., “Thousands of AI Authors on the Future of AI,” Journal of Artificial Intelligence Research 84 (2025). https://doi.org/10.1613/jair.1.19087
• Joseph Weizenbaum, “ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine,” Communications of the ACM 9(1) (1966). https://doi.org/10.1145/365153.365168
• Joseph Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation (W. H. Freeman, 1976). Google Books
• Karen Hao, “We read the paper that forced Timnit Gebru out of Google. Here’s what it says,” MIT Technology Review (2020). https://www.technologyreview.com/2020/12/04/1013294/
• Peter Lee, “Learning from Tay’s introduction,” Microsoft (2016). https://blogs.microsoft.com/blog/2016/03/25/learning-tays-introduction/
• Personal Information Protection Commission, “Scatter Lab Administrative Disposition” (2021). Korea’s Personal Information Protection Commission
• Meta AI, “BlenderBot 3: A 175B parameter, publicly available chatbot” (2022). https://ai.meta.com/blog/blenderbot-3-a-175b-parameter-publicly-available-chatbot-that-improves-its-skills-and-safety-over-time/
• Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland & Romeo Dean, AI 2027 (AI Futures Project, 2025). https://ai-2027.com/
• Fabrizio Dell’Acqua et al., “Navigating the Jagged Technological Frontier,” Organization Science (2026). https://doi.org/10.1287/orsc.2025.21838
• Johannes C. Eichstaedt et al., “Facebook Language Predicts Depression in Medical Records,” PNAS 115(44) (2018). https://doi.org/10.1073/pnas.1802331115
• Nicholas Carlini et al., “Extracting Training Data from Large Language Models,” 30th USENIX Security Symposium (2021). https://arxiv.org/abs/2012.07805

Principles, law, institutions, and responsibility

• Anna Jobin, Marcello Ienca & Effy Vayena, “The global landscape of AI ethics guidelines,” Nature Machine Intelligence 1, 389–399 (2019). https://doi.org/10.1038/s42256-019-0088-2
• Luciano Floridi et al., “AI4People—An Ethical Framework for a Good AI Society,” Minds and Machines 28 (2018). https://doi.org/10.1007/s11023-018-9482-5
• UNESCO, Recommendation on the Ethics of Artificial Intelligence (2021). https://www.unesco.org/en/artificial-intelligence/recommendation-ethics
• OECD, OECD AI Principles (2019; updated 2024). https://oecd.ai/en/ai-principles
• European Commission, “AI Act” (updated 2026-07-07; accessed 2026-07-16). https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
• Jessica Morley et al., “Ethics as a Service: A Pragmatic Operationalisation of AI Ethics,” Minds and Machines 31 (2021). https://doi.org/10.1007/s11023-021-09563-w
• Luciano Floridi, “Soft Ethics and the Governance of the Digital,” Philosophy & Technology 31 (2018). https://doi.org/10.1007/s13347-018-0303-9
• Caroline Whitbeck, Ethics in Engineering Practice and Research, 2nd ed. (Cambridge University Press, 2011).
• Madeleine Clare Elish, “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction,” Engaging Science, Technology, and Society 5, 40–60 (2019). https://doi.org/10.17351/ests2019.260
• Julia Angwin et al., “Machine Bias,” ProPublica (2016). https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
• Alexandra Chouldechova, “Fair Prediction with Disparate Impact,” Big Data 5(2), 153–163 (2017). https://doi.org/10.1089/big.2016.0047
• Jon Kleinberg, Sendhil Mullainathan & Manish Raghavan, “Inherent Trade-Offs in the Fair Determination of Risk Scores,” ITCS (2017). https://doi.org/10.4230/LIPIcs.ITCS.2017.43
• !Mediengruppe Bitnik, Random Darknet Shopper (2014–2016). Project page
• Tega Brain & Surya Mattu, Unfit Bits (2015). https://tegabrain.com/Unfit-Bits
• MIT AI Risk Initiative, AI Incident Tracker (2026; accessed 2026-07-16). https://airisk.mit.edu/ai-incident-tracker/incident-timeline

Creativity, cognition, and relationships

• Anil R. Doshi & Oliver P. Hauser, “Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content,” Science Advances 10 (2024). https://doi.org/10.1126/sciadv.adn5290
• Ilia Shumailov et al., “AI Models Collapse When Trained on Recursively Generated Data,” Nature 631 (2024). https://doi.org/10.1038/s41586-024-07566-y
• Nataliya Kosmyna et al., “Your Brain on ChatGPT,” arXiv preprint (2025). https://arxiv.org/abs/2506.08872
• Michael Gerlich, “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking,” Societies 15(1) (2025). https://doi.org/10.3390/soc15010006
• Allison J. B. Chaney, Brandon M. Stewart & Barbara E. Engelhardt, “How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility,” RecSys (2018). https://doi.org/10.1145/3240323.3240370
• OpenAI, “Sycophancy in GPT-4o: What Happened and What We’re Doing About It” (2025-04-29). https://openai.com/index/sycophancy-in-gpt-4o/
• OpenAI, “Expanding on What We Missed with Sycophancy” (2025-05-02). https://openai.com/index/expanding-on-sycophancy/
• Pew Research Center, “How Teens Use and View AI” (2026). https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/
• Common Sense Media, Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions (2025). Common Sense Media PDF
• Hakuhodo, “Developing a Generative-AI Service Prototype to Support Consumer Insight” [Japanese] (2024). https://www.hakuhodo.co.jp/news/newsrelease/109174/ / Figures (PDF)

Agents, markets, infrastructure, and documentation

• AWS, “Agents that Transact: Amazon Bedrock AgentCore Now Includes Payments (Preview)” (2026). https://aws.amazon.com/about-aws/whats-new/2026/04/amazon-bedrock-agentcore-payments-preview/
• Amine Allouah et al., “What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce” (2025). https://arxiv.org/abs/2508.02630
• Shota Imai, How Generative AI Will Change the World, keynote at NII Open House 2026 [Japanese]. Keynote video
• Nenad Tomasev et al., “Virtual Agent Economies” (2025). https://arxiv.org/abs/2509.10147
• Jordi Montes Sanabria & Pol Alvarez Vecino, “Beyond the Sum: Unlocking AI Agents Potential Through Market Forces” (2025). https://arxiv.org/abs/2501.10388
• upamune, human-mcp (2025). https://github.com/upamune/human-mcp
• Lisa Eadicicco, “He Helped Create the ‘Google Brain.’ Here’s What He Thinks About AI Now,” TIME (2017). https://time.com/4631730/andrew-ng-artificial-intelligence-2017/
• BlackRock, “2026 Infrastructure Summit” (2026). https://www.blackrock.com/corporate/insights/2026-infrastructure-summit; Thibault Spirlet, “Sam Altman says AI will eventually be sold like electricity and water,” Business Insider (2026). Business Insider
• Anthropic, “Detecting and Preventing Distillation Attacks” (2026). https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks
• OpenAI, Update to the U.S. House Select Committee (2026). OpenAI PDF
• U.S. Center for AI Standards and Innovation, Evaluation of DeepSeek AI Models (2025). NIST PDF
• Ali Naseh et al., “R1dacted: Investigating Local Censorship in DeepSeek’s R1 Language Model” (2025). https://arxiv.org/abs/2505.12625
• Peiran Qiu, Siyi Zhou & Emilio Ferrara, “Information Suppression in Large Language Models” (2025). https://arxiv.org/abs/2506.12349
• Timnit Gebru et al., “Datasheets for Datasets,” Communications of the ACM 64(12) (2021). https://arxiv.org/abs/1803.09010
• Margaret Mitchell et al., “Model Cards for Model Reporting,” Proceedings of FAT* (2019). https://arxiv.org/abs/1810.03993
• Audrey Tang, “Alignment Assemblies Can Enable Us to Govern AI Collaboratively” (2024). https://rebootdemocracy.ai/blog/audrey-tang-ai-democracy

Magic, ritual, and character

• Stanley J. Tambiah, “The Magical Power of Words,” Man 3(2), 175–208 (1968). https://doi.org/10.2307/2798500
• Stanley J. Tambiah, Magic, Science, Religion, and the Scope of Rationality (Cambridge University Press, 1990).
• Merriam-Webster, “The Word History and Definition of ‘Character.’” https://www.merriam-webster.com/wordplay/word-history-of-character-origins