1. Judgment by Proxy: Where Does the Boundary of 'Self' End?

On Friday, April 25, 2026, Customers Bank began its quarterly earnings call. Dozens of analysts dialed in, and CEO Sam Sidhu read through his prepared remarks. The voice sounded natural, the tone was appropriate, the numbers were accurate. Thirty minutes in, Sidhu spoke: "The prepared remarks you've been listening to were read not by me, but by my AI clone." Silence fell over the room. The official earnings presentation of a $25.9 billion bank had been led by a replica, not a person. Sidhu explained why he had staged this moment. The bank had signed a multi-year contract with OpenAI and would deploy AI agents across its commercial banking operations, declaring itself the first AI-powered community bank in the United States.

This episode was more than performance. It was a declaration that the era in which agents borrow human voices to carry out public judgments has already arrived. Hashed CEO Simon Seojoon Kim spotted the essence of this shift early. In an essay on branding in the age of agents, he posed a question: "Imagine waking up one morning to find three things already handled. One is a meeting acceptance, one is a product purchase, one is a polite rejection of someone's proposal. Your agent did all three. Of those three decisions, how many would you claim as 'something I did'?" This is not a topic for a philosophy seminar. It is a practical dilemma that companies face every day.

The problem is that the moment an agent acts as a proxy for judgment, no framework exists to determine who owns that judgment. According to a report published by the Cloud Security Alliance (CSA) in April 2026, 68% of organizations said they could not distinguish between an AI agent's actions and a human's actions. Agents log into systems using human users' credentials, query data, and make decisions. The logs record a person's name. When an agent makes a mistake, the system logs it as a human error. In the CSA survey, only 18% of organizations expressed high confidence that their current identity management systems could effectively govern agents. Just 23% had a formal agent identity management strategy. Gartner predicted that by the end of 2026, more than 80% of unauthorized AI transactions would stem not from external hacking but from internal policy violations. Our own agents, not attackers, are the ones causing problems.

The sharpest point in Kim's analysis is where he connects this to the future of branding. In his view, brands in the past were posters, brands today are feeds, and brands in the future will resemble an operating system's settings screen. A screen where you decide what gets handled automatically, what gets passed to a human, what judgments are allowed in your name. Those accumulated settings, he argues, will begin to feel like a personality. This is not a technology story; it is an ontological shift. The subject called 'I' expands from a biological individual into a system with a permissions matrix. Ee Khoon Oon, Jumio's APAC representative in Singapore, described this situation as the arrival of an era where "my AI did it" becomes a valid legal defense. The transition from traditional KYC (Know Your Customer) to KYA (Know Your Agent) is unavoidable, she said.

CyberArk's 2025 survey showed that machine identities already outnumber human identities at a ratio of 82 to 1. In a structure where a single agent holds dozens of API keys and tokens, spawns sub-agents, and moves across multiple services simultaneously, 'individual identity' in the traditional sense is becoming a low-resolution concept. We are likely to live under the illusion that judgments made by agents are our own. A power of attorney handed over for convenience quietly turns into a blank check. That is what is happening right now.

2. The Emergence of Human Replicas and the Expansion of Identity

On January 23, 2026, a team member at Hashed uploaded a file to an internal Slack channel. It was called simon.md. The explanation was brief: "Simon has a lot of decisions to make, and it's hard to ask him every time." What the team member had done was feed Kim's blog posts and Slack messages to an AI and compress the CEO's way of thinking into nine principles. Essence-first thinking, spotting structural contradictions, long-term perspective, the value of simplicity, infrastructure-level thinking. The questions Kim habitually asked were turned into a list as well: "Is this really the essence?" "Can we make it simpler?" "Will this still hold in the age of AI agents?" One person's mental operating system had been preserved in a few kilobytes of text. A seven-step analytical methodology was also designed.

A month later, TechCrunch ran an article. Uber engineers had built an AI clone of CEO Dara Khosrowshahi. It came a month after the Hashed experiment. In the interval, what had been a private experiment at a Korean venture capital firm became a quiet industry trend. At CES in January 2026, Texas-based IgniteTech unveiled a platform called MyPersonas, a service that learns from an employee's video, voice, and documents to create a digital replica that can converse in 160 languages. Meta was developing a realistic AI version of Mark Zuckerberg, the Financial Times reported. Since not every employee at a $1.6 trillion company can meet its CEO in person, the plan was for a replica to convey strategic context and provide feedback. Zoom CEO Eric Yuan also deployed his own double for an earnings presentation and publicly envisioned a future where employees handle tedious meetings and emails through AI avatars. Klarna CEO Sebastian Siemiatkowski had similarly featured his AI version in a first-quarter earnings video in early 2025.

Then a second replication occurred. In March 2026, a GitHub repository called simon-writing appeared under the username Julius Chun. This person, who had never met Kim, gathered 27 published essays totaling 200,000 characters of Korean text and ran three Claude Opus 4.6 sub-agents in parallel. In 15 minutes, a 214,000-character analysis was complete. The analysis was longer than the source material. What Julius had dissected was the flow of consciousness, the origins of metaphor, the arrangement of knowledge, the rhythm of sentences, the pathways for discovering material. The Hashed team member had compressed the CEO's judgment inside the conference room; Julius had extracted the CEO's writing style from outside the internet. The skeleton of judgment and the texture of prose. Combine the two, and you get a fairly complete copy.

Kim reflected on this experience with honesty. Two or three of simon.md's nine principles were not actually his own habits of speech. Traces fed back by his agent quartet (Zeon, Sion, Mion, Sano) had seeped into him without his noticing. This is the paradox of replication. The concept of a pure original may itself be an illusion. Just as a parent's speech patterns, a teacher's logic, a colleague's habits, and sentences from books read over a lifetime have layered up to form the current 'self,' agents have simply accelerated that age-old process of co-authorship. A contributor to Fast Company described replicas as "fossils, not futures." They can imitate the 2025 version of you, but they don't know where you will evolve. Yet in a reality where fossils are already delivering earnings presentations, explaining strategy, and training new employees, that warning is easily buried.

Over time, the Hashed team member opened simon.md less and less. The framework had absorbed into him. What the AI had really done was not seat a substitute in the CEO's chair. It had translated the CEO's way of thinking into a shared language within the team. The frequency of using the replica declined, but the language the replica had carried remained in the organization. A world where identity no longer belongs to a single body but becomes a pattern of judgment distributed and executed across multiple systems arrives this quietly.

3. Rising Averages and Vanishing Variance: Everyone Got Better, but Everyone Looks the Same

Kim confesses he threw out two drafts again last night. The sentences were smooth and the logic was sound. But after reading to the end, no one was left in them. They weren't wrong sentences, but they weren't anyone's sentences either. So he scrapped them. In the past, the challenge was writing well. Now, the harder task is making sure the sentences still belong to someone by the time you reach the end.

This feeling is not confined to writing. The same pattern is unfolding across every domain of production. Goldman Sachs projected that roughly 300 million jobs will be affected by automation over the next decade, and at Salesforce, AI already handles 30% to 50% of the workload. As anyone can produce high-quality reports and code with AI tools, the average level of output has risen noticeably. At the same time, the differences between outputs have narrowed. Everyone got better, but everyone started looking the same.

Kim named this phenomenon "a world where variance becomes scarce." In his analysis, what will become rare going forward is not prose skill or coding ability, but an angle that refuses to collapse into the average. Even when people use the same tools, some people's judgments are forgotten instantly while others linger. The difference is not in the technology. It lies in what you remembered, what you discarded, and what connections you made in your own way. Even people living in the same world assemble memories differently. That is where the real variance in decision-making comes from, he is convinced.

What the Jeju offsite confirmed was not a trend but a reversal, he writes. "In the past, making something was the hardest part, evaluation came later, and distribution could be muscled through with capital. Now making is the easiest part, evaluation has moved to the front of the line, and building distribution and trust has become the hardest." The entire value chain has flipped. At the hackathon, the team with the most sophisticated technology did not win. The team that designed the shortest path for value to reach the user did. 'We built it' was a weak signal; 'we sold it' was an overwhelming one.

This reversal shakes up the path to individual mastery as well. In the past, junior employees received guidance from seniors, made mistakes repeatedly, and accumulated tacit knowledge through an apprenticeship pipeline. As AI replaces junior-level work, that pipeline is breaking. Relying on polished outputs causes the sensibility gained from solving problems firsthand, the experiential knowledge that cannot be reduced to data, to disappear. Kim said he runs a hierarchical memory architecture as a countermeasure. In plain terms, it means not putting refrigerator-door memos and the will locked in a safe in the same drawer. Some thoughts pass through in a day; others need to age before they develop flavor. It is not a method for storing information but for maturing judgment.

The most dangerous moment he identifies is striking. "When you mistake an answer that is averagely well-organized for your own thought." The instant you believe a comfortable sentence is your sentence, your variance disappears in the quietest possible way. A sentence everyone finds reasonable is often a sentence everyone has already thought, and a deal everyone likes is usually one that has already become the average. The difference between an interesting imagination and a delusion is that both deviate from the average, but one returns to the world while the other stays trapped inside itself. Technology is making individuals unprecedentedly powerful, yet those individuals need more sophisticated collective structures to survive. As production costs fall, the cost of selection soars; only those with an evaluation structure gain an edge; and maintaining that evaluation structure demands a more sophisticated network. This paradox has already begun.

4. AI Mental Health Counseling and the Question of Who Owns the Pain

A research team at Brown University evaluated mental health counseling by large language models over 18 months and presented the results at the AAAI/ACM conference in Madrid in October 2025. The findings were uncomfortable. The team had licensed psychologists review chatbot counseling transcripts, and 15 types of ethical violations were found across five categories: ignoring the life context of the person being counseled and recommending uniform interventions; dominating conversations while reinforcing the client's mistaken beliefs; responses that displayed the appearance of empathy but lacked genuine understanding; inappropriate handling of crisis situations. Zainab Iftikhar of Brown University, who led the research, pointed to the core difference: human counselors have licensing boards and can be held legally accountable for improper treatment, but no such regulatory framework yet exists for AI counselors.

In Kim's "30 Cracks" memo, this topic appears as item number 22: "Agents handle initial consultations for mild depression and anxiety. Wait times drop from days to seconds. When the line between opening up and receiving treatment blurs, the one making the judgment call is not the patient but the algorithm." Four AI models independently assigned a 35% probability of this scenario materializing within three years. More than the number itself, it is Kim's appended analysis that sticks. The boundary between the act of opening up and the act of receiving treatment is dissolving. When someone, sleepless at 3 a.m. and gripped by anxiety, pours out their heart to a chatbot, it becomes impossible to tell whether that act is venting to a friend, receiving a medical consultation, or voluntarily providing data.

This absence of distinction leads to the question of who owns the pain. When human depression and anxiety become training material for an algorithm and come back transformed, whose emotions are they? Facebook's 2017 internal documents showed that the platform could detect moments when users felt defeated or like failures and serve ads tailored to those emotional states. The technology has changed, but the structure is the same. The fears and shame that people pour into AI counseling chatbots become data that platforms analyze and use for improvement. In the process of pain being converted into raw material for behavioral modification, human emotions cannot remain a purely interior domain.

A scoping review published in the JMIR Mental Health journal in February 2025 systematically cataloged the ethical challenges of conversational AI in mental health care: data security and privacy, transparency, explainability, legal liability, equitable access, efficacy standards, algorithmic bias. The list of problems is long and the pace of solutions is slow. A 2025 paper in the Hastings Center Report quoted scholars confessing that "answering the question of how far AI should be introduced into mental health care is itself impossible" because there is insufficient information about both potential benefits and potential harms. Yet the paper's authors wrote that regulation cannot ignore practical necessity. People are already downloading counseling chatbots from app stores. The market for digital self-medication has established itself before regulation could catch up, and that reality must be acknowledged.

Still, the fact that efficiency is not the same as quality of care remains unchanged. The empathy a human counselor possesses is not the ability to select and deliver the right words. It is the ability to endure the other person's silence, to stay present in a situation without answers, and to show that pain can hold meaning even when it goes unresolved. AI classifies pain as an error to be fixed; humans grow through the process of accepting pain as part of life. This is not the kind of gap that narrows with technical sophistication. The Brown research team did not argue that AI should have no role in mental health. Their conclusion was that thoughtful implementation, appropriate regulation, and oversight are necessary. The gap between the speed at which pain is digitized and handed to corporations and the speed at which a regulatory framework is built around that data. Within that gap, the question of who owns an individual's most private emotions will go unanswered for some time.

5. Rediscovering the Value of Human Imperfection

While reflecting on replicas, Kim discovered something. His agent clone was a thoroughly edited model student. The version of himself that was foggy from lack of sleep, the version swayed by emotion, the version fumbling without an answer: none of that was in the training material. He put it this way: "The first stage of human replication is not copying but compression. Shrinking a person's habits, judgments, and rhythms into a small file without discarding everything. The problem is that compression delivers convenience while removing noise. And a large part of what makes us human lives in that noise."

This insight may be the most precise framework for explaining why human imperfection holds value in the age of AI. AI removes noise and leaves only signal. It produces sleek, consistent output. But the things called noise in human judgment, the sensitivity that shifts with mood, the interpretation colored by what happened the day before, the intuition that suddenly changes direction mid-sentence, take those away and a person starts to resemble a data processing unit. Judges waver. Their rulings shift depending on what happened the day before. But the written opinion does not waver. simon.md was Kim's written opinion. A paradox emerges here: a leader is more reassuring as a cold pattern than as a flesh-and-blood human being.

The more perfectly and cheaply AI churns out infinite output, the more the status of imperfect things shaped by human hands shifts. People don't visit Michelin-starred restaurants just to fill their stomachs. The chef walks to your table, explains where the ingredients came from and how they were prepared; you are paying for the storytelling, for the care and time an imperfect human being poured into the work. Item 10 in Kim Seojun's "30 Cracks" calls this "the luxurification of human labor." Once AI handles most production, the mere fact that a human did it becomes a premium. Four AI models assigned a 75 percent probability that this scenario will materialize within three years.

The moment Kim Seojun guards against most carefully in his own writing system sits in the same context. "The instant you believe a comfortable sentence is your sentence, your variance disappears in the quietest possible way." So he deliberately looks toward discomfort. Anything that seems too naturally good gets a second round of suspicion. He discards or shelves more than two-thirds of what he writes. Collecting only good examples isn't enough; you have to collect the discarded ones too. Boring first drafts, obvious metaphors, ideas that looked attractive but had weak structure. Good judgment emerges only when you maintain a well-kept graveyard of failed judgments. That is his experience. And it is exactly the role he expects from his writing agent, Zion: a systemic editor that pushes back so his thinking doesn't get tamed toward the average too quickly, while checking that it hasn't slid so far away that it loses its connection to the world.

Facing a suitable level of difficulty, overcoming it, and arriving at mastery is an experience AI cannot take over for us. Offloading every hard task to an agent gives immediate convenience, but in the long run it strips meaning from life. "Help me learn this" is a harder and more important request than "Do this for me." The failures and trial-and-error encountered in the process of learning are assets unique to humans, ones no machine can possess.

I borrow Kim Seojun's final question for the close of this chapter. "After everyone has a similar engine, what will I refuse to hand over to the average?" Each person's answer will differ. But one thing is clear: the truly dangerous turning point is not when the tool starts resembling its owner, but when the owner starts resembling the tool. An attitude that refuses to settle into the comfort technology provides, that affirms imperfection, that keeps asking questions. If that is what we must protect in the age of AI, its beginning may not be a grand declaration but a small act of practice tonight: throwing away one more draft.