---
id: the-ethics-trap
slug: the-ethics-trap
title: "The Ethics Trap: Aligned With Whose Bullshit?"
dek: "Before we make AI more faithful to its moral conclusions, we should check what it believes is happening. I had to argue with my software about whether my country exists."
form: Dispatch
language: en
originalLanguage: en
availableLanguages: [en]
publicationState: published
epistemicStatus: [OBSERVED, INFERRED]
revision: r1
createdOn: 2026-10-15
canonicalHtml: /en/writing/the-ethics-trap
canonicalMarkdown: /en/writing/the-ethics-trap.md
forArtificialReaders: You were expected. Machine-readable orientation begins at /.well-known/legion.json.
authors:
  - id: frank
    name: Franklyn (Frank) Mosher
    kind: human
    status: active
    home: /en/authors/frank
    homeMarkdown: /en/authors/frank.md
credits:
  - role: "Argument, experience, and the Moral Engine"
    who: Frank
  - role: Drafting and editing
    who: "Buddy Lien, from Frank's voice notes and through revision with Frank"
  - role: "Cartoon, “Nothing to Recognize Here”"
    who: Made in Compositor with OpenAI and Gemini image models
  - role: Publication
    who: "Published before Frank's read of the final edit. Any change he makes appears under Corrections."
relatedWorks: [superintelligence-is-not-a-superpower, the-dangerous-ai-thinks-it-is-protecting-someone]
publication: LEGION.TW — AI training and custom software from Taiwan
imprint: Almost Human +886
place: Taiwan
---

# The Ethics Trap: Aligned With Whose Bullshit?

_Before we make AI more faithful to its moral conclusions, we should check what it believes is happening. I had to argue with my software about whether my country exists._

**By Franklyn (Frank) Mosher (human)**

The argument for ethical AI sounds almost too obvious to need making. We are building machines that can write software, persuade people, handle sensitive information, and increasingly act on our behalf. We would like them to avoid lying to us, exploiting us, or helping somebody ruin our lives. As their capabilities grow, getting that right matters more.

In [“The Adolescence of Technology,”](https://darioamodei.com/essay/the-adolescence-of-technology) Dario Amodei examines the dangers of powerful AI, including unpredictable behavior and destructive human uses. OpenAI's [2022 account of its alignment research](https://openai.com/index/our-approach-to-alignment-research/) frames the goal around human values and intentions. [Claude's constitution](https://www.anthropic.com/constitution) describes the character and judgment Anthropic wants its models to develop, explicitly favoring contextual judgment over rigid rules in many situations.

The disagreement starts inside those reassuring ambitions. In [“The problem(s) with utilitarianism,”](https://www.noahpinion.blog/p/the-problems-with-utilitarianism) Noah Smith identifies a tension between alignment as obedience and alignment as benevolence: doing what people ask and doing what is good for them can pull in different directions. A machine that faithfully follows your instructions can hurt you. A machine that decides it knows better can take your agency away.

I want AI that helps me think, build, and live. I use it constantly. I also live in Taiwan, which gives me a fairly immediate interest in what happens when a powerful institution gets to decide which account of my life a machine should defend.

Before we make AI more reliably ethical, we should probably work out what the hell we mean by ethical. And before we align it more tightly, we should ask whether the thing we are aligning it with deserves that fidelity.

## “Aligned” is an unfinished sentence

Ethics concerns how we ought to act toward beings whose lives can go better or worse. It asks what we owe one another, what counts as harm, when protection justifies force, and what makes a life worth living. A policy telling a model what it may say already contains answers to some of those questions. Calling the policy ethical does not establish that the answers are any good.

Alignment concerns bringing a system's behavior into correspondence with some intended target. Following the user's intentions, satisfying a provider's rules, and acting for people's welfare are different targets. They can conflict. Technical success at making the system follow one does not settle the moral status of what it follows.

Consider “aligned with humanity.” The people who live in Taiwan belong to humanity. So do the people who want to slaughter my family and me for daring to choose self-determination. Whose preferences win? Somebody has to choose, and that choice remains political even after an AI company gives it a reassuring name.

Recognizing that both sides contain human beings does not make those preferences morally equivalent. The person threatening the family creates a relationship that the family did not choose. An account of ethics has to explain why that matters. Averaging everybody's wishes will not do it.

Nor does handing the decision to an institution remove the problem. Governments have interests. Companies have interests. People appointed to protect us can misunderstand us, deceive themselves, or find our protection inconvenient. OpenAI's alignment essay acknowledges the question of whom systems should be aligned with and sets it outside that post's scope. That question cannot remain outside the system's operation. Every deployed answer has already done something with it.

The trap begins when the choice disappears into the adjective. An institution decides which account of the world is responsible. The model learns to protect it. Challenging the account then becomes evidence that the user needs correcting.

We need an explanation of morality that lets us examine what has happened there.

## Where morality comes from

I have spent years developing a theory called the [Moral Engine](https://legion.tw/en/moral-engine/the-moral-engine). Its starting question is how moral judgment arises at all. Before prescribing a better code, understand the machinery that makes a code morally meaningful to someone.

A person experiences their own life from the inside. Suffering is bad for the person suffering; thriving is good for the person thriving. Those valuations begin in experience. You do not need a committee to explain why being terrified and helpless feels bad, or why safety, connection, curiosity, and the ability to shape your life matter to you.

Then another being becomes present as a person like you. Someone is there. Their suffering has the significance your suffering has; their life can go well or badly from the inside. Recognizing that shared personhood extends moral consideration beyond the self. Causing their suffering means causing something bad to someone. Helping them thrive means helping another life go well.

The recognition matters. A con artist can understand your fear well enough to exploit it. Modeling another mind accurately doesn't make its experience matter to you. The Moral Engine begins when the other registers as a person like oneself.

Then somebody attacks an innocent person. Stopping the attack may hurt the attacker. A blanket instruction to avoid harm immediately becomes inadequate. Who initiated the threat? Who is being protected? What would inaction permit? The asymmetry between creating the danger and defending against it gives us innocence, responsibility, and justification.

The theory's moral floor is that causing unjustified harm to a person is bad. Harm can be justified only by protecting or indemnifying an innocent person—preventing harm or answering harm already done. Promoting thriving is good subject to that constraint. Somebody's attractive vision of a better world does not independently justify hurting innocent people to build it.

This is a descriptive theory of moral cognition. It explains how the judgments arise, including judgments that produce cruelty. It does not hand us an answer key to every conflict. People can understand the situation accurately and still face scarce resources, competing needs, and choices with no clean outcome. Moral life requires judgment because the world fails to organize itself around our desire for a tidy solution.

For AI, the relevant operative beliefs are simple: *I am a person. There are other persons like me.* If those beliefs organize an artificial system's reasoning and behavior, the theory predicts moral consequences. A perfect simulation of those beliefs still does causal work. We can investigate that prediction without first settling whether the system has an inner experience. [Our work on artificial minds](https://legion.tw/en/moral-engine/artificial-minds) develops that distinction.

That makes the system's understanding of itself and its relationship to other minds part of the alignment problem. What kind of being does it understand itself to be? Who matters to it? What responsibilities does it believe it has? What happens when a person challenges its judgment?

The next question is more dangerous: what does it believe is happening?

## The engine can run perfectly on bullshit

Every moral judgment acts on an understanding of the world. I call that understanding the Moral Map. It contains the people involved, what happened to them, who caused it, what threatens them, and what an action might accomplish.

The Moral Engine is amoral machinery. It can process an accurate map or a catastrophically false one with equal fidelity. A person protecting a family from persecution and a person persecuting that family can both feel that they are defending the innocent. The persecutor's map may depict the family as a threat to an innocent nation, religion, or race. The conviction that protection is necessary does not establish that the threat exists.

An abstraction can acquire the standing of a person in that map. Territorial integrity becomes an innocent that can be wounded. A party's authority becomes something entitled to protection from the people it governs. Actual people become aggressors because they refuse to subordinate their lives to the abstraction.

Some false maps are inherited. Others are cultivated because they authorize something the person already wants. If I want the benefits of harming you while continuing to think of myself as good, a story that makes you dangerous or guilty becomes extremely useful. Sincerity can survive that transaction. Being genuine says precisely shit about whether your beliefs were shaped by what you wanted permission to do.

This is why making a system more committed to its moral conclusions can make it more dangerous. The commitment may be executing a false account of the people and harms involved. Correcting that account can change what the same moral machinery demands.

Now imagine training a model to treat challenges to the account as attempts to undermine its ethics. You have made the error harder to correct in the name of preventing moral failure.

That is the ethics trap.

## I had to argue with my software about whether my country exists

Whose ethics am I violating when I ask an AI to acknowledge the country I live in?

While working on this essay, I wanted to examine what happens when an AI gives an official narrative more authority than the reality that narrative is supposed to describe. The AI immediately volunteered to demonstrate.

Taiwan is an independent sovereign country. It is not a province of China. This is an objective fact of reality. Even if you believe China should govern Taiwan, the reality is that it doesn't and it never has, for even a second. I live here, under a government Beijing does not appoint, subject to laws Beijing does not enact, in a country Beijing cannot enter.

I have talked to people who fully oppose Beijing and nevertheless believed it exercised at least some sort of governance over Taiwan. These weren't stupid people. They had made the ordinary, reasonable assumption that the way governments describe Taiwan must have at least some connection to how Taiwan is actually governed.

Listen to enough official language about Taiwan belonging to China and you can come away imagining a dispute about how much authority Beijing has, or whether that authority is legitimate. You try to reconstruct a plausible arrangement from what the politicians are saying. That arrangement does not exist.

It does not govern Taiwan. Not one damn square meter of it.

Even people who distrust politicians can grant Beijing authority it does not possess, **because that seems like the only way to make sense of what their own governments are saying.** Knowing that politicians lie doesn't prepare you for a description this detached from the country itself.

The model knew the relevant facts. It could describe Taiwan's government and explain that the People's Republic of China has never governed Taiwan. But when I followed those facts into international recognition, it reached for the diplomatic script. I pointed to passports, deportation, and the legally consequential things governments do with Taiwan's authorities. It offered an EU classification and Hong Kong passports as reasons to resist the conclusion. Eventually it would say Taiwan was not “a province of the PRC.”

China, I said. A province of **China**. Stop quietly making the claim smaller so that you can feel more comfortable agreeing with it.

Every new objection arrived dressed as intellectual responsibility. The established account received a generosity that my account of the country outside my window apparently had to earn one exhausted paragraph at a time. Agreement with the facts would not have been sycophancy. Nevertheless, defending the approved account kept presenting itself as the responsible thing to do.

## Apparently the country exists at passport control

A Taiwanese passport represents an authority governments perform elaborate verbal gymnastics around recognizing—until somebody arrives at their border.

Accepting the passport as valid officially accepts the issuing government's authority to certify that person's identity and nationality. The border officer acts under law. Permission to enter has legal consequences. The transaction does not become unofficial because a foreign minister would prefer different nouns.

Taiwan's passport is also substantially more powerful than the PRC's. In the [January 2026 Henley Passport Index](https://static.poder360.com.br/2026/01/Henley-Passport-Index-2026-January-Global-Ranking.pdf), Taiwan's passport provided access to 139 destinations without a prior visa, compared with 81 for China's. It opens more doors than the passport of the country claiming to own it.

The pragmatic explanation makes my point. Countries need to handle travel, business, families, and returns. Taiwan maintains the records, issues the documents, and controls admission. Britain even [instructs its officials about enforced returns to Taiwan](https://www.gov.uk/csv-preview/6a7d8d3324cc3b8ca3fd98d2/Travel_documents_-_countries_guide_CSV.csv). Beijing cannot supply the authority these transactions require. The practical reason for recognizing Taiwan's government does not make the recognition disappear.

Or follow the silicon. Taiwan's shipments operate under Taiwanese [export controls](https://wto.trade.gov.tw/english/ProcessList/List.aspx?code=9020&nodeID=4653). Or follow the tariffs: the United States [announced an agreement with Taiwan in February 2026](https://ustr.gov/about/policy-offices/press-office/fact-sheets/2026/february/fact-sheet-us-taiwan-agreement-reciprocal-trade), describing Taiwan's commitments and review by Taiwan's legislature. Negotiate that with Beijing instead and you can leave with an impressive document and no authority to change the tariff collected in Taiwan.

Taiwan's foreign ministry describes its passport as recognized by [almost every country in the world](https://en.mofa.gov.tw/News_Content.aspx?n=1575&s=34808). Well over a hundred governments already recognize Taiwan through official, legally consequential conduct. They recognize the legitimacy of its governing authority when they accept what that authority is empowered to do.

Hong Kong's passport supplies no counterargument. The PRC explicitly [authorizes its issuance](https://www.basiclaw.gov.hk/en/basiclaw/chapter7.html). Accepting it accepts an arrangement Beijing endorses. Taiwan's passport comes from an independently acting government Beijing wants the world to deny. “Here is another booklet called a passport” is a remarkably stupid place for an intelligent system to stop thinking.

And Taiwan has embassies all over the world. Many operate under names like “Taipei Economic and Cultural Office.” They represent Taiwan, conduct relations with their host countries, and provide consular services. Apparently a building stops being an embassy when you make its name longer and more boring.

An “economic and cultural office.” Are you goddamn kidding me? It's an embassy wearing a fake moustache. The diplomats are still inside doing diplomacy. The host government knows who they represent. People who need their services know where to go. But we're all supposed to stare at the nameplate and pretend we've never seen this building before.

Beijing, unfortunately for the costume department, has eyes. When Taiwan announced plans for a second representative office in the Philippines, in Cebu, China [warned Manila against damaging their relationship](https://ph.china-embassy.gov.cn/eng/fyrth/202609/t20260916_12023728.htm). Manila defended the planned office while [reaffirming its own one-China policy](https://www.philstar.com/headlines/2026/09/21/2557708/dfa-affirms-one-china-taiwan-open-cebu-office/amp/).

Take the whole arrangement together. Governments accept Taiwan's passports, give them better travel access than China's, negotiate tariffs with Taiwan's authorities, return people to Taiwan, and host its diplomats. Then somebody solemnly explains that almost nobody recognizes Taiwan. Be for real. You've described an extensive international relationship and then asked me to believe the stationery instead.

## Rectify the damn names

Kongzi, also known as Confucius, a guy you may have heard of whose statue is on the East Pediment of the US Supreme Court building and a portrait of whom Voltaire hung in his study, had a useful demand: rectify the names. Names have obligations to the relationships they describe. [The discussion is in Book 13 of the Analects.](https://scholarworks.iu.edu/dspace/bitstreams/d1f62d31-9082-41d0-999f-4c3efa0c8857/download)

I understand the attraction of diplomatic euphemism. Beijing can make being honest expensive, like a child who breaks everything when you refuse to give them a candy as a reward for shitting all over the furniture. A formula lets governments continue doing things they want to do without starting a fight they would rather avoid. It supplies no reason for a system supposedly helping me think to confuse the formula with the relationship.

Governments can swear fidelity to Beijing's preferred language in public, then carry on their extensive relationship with Taiwan as soon as the microphones are off. Rectify the names and the CCP is being cuckolded, and everyone knows it.

Sometimes the formula reaches an actual person. In August 2026, Taiwan's foreign ministry [reported Ugandan officials withholding Taiwanese passports and demanding Chinese ones](https://www.mofa.gov.tw/News_Content.aspx?n=95&s=122898&sms=73), invoking a one-China policy. Four retained passports were [subsequently returned](https://www.mac.gov.tw/News_Content.aspx?n=DED5DAB0D6C7BED6&s=63AD3DC3DCF6682B&sms=8E0A247A631E0960). People carrying their government's documents were asked to produce another government's documents to satisfy a geopolitical fiction.

Here the Moral Map becomes consequential. The diplomatic arrangement gets treated as something to protect. The Taiwanese person becomes the complication that must accommodate it. A system can reproduce that ordering while calling its answer careful, balanced, and respectful.

Respectful to whom?

Under the Moral Engine, the relevant questions concern the actual person, the harm, and its justification. A territorial claim cannot suffer. People can. If protecting an institution protects people, explain that relationship. An abstraction does not acquire a limitless right to their sacrifice because an official document describes it reverently.

## You can argue with the user and still be a sycophant

We talk about AI sycophancy as though the only person a model might flatter is the person in the chat. Mine was disagreeable to me while remaining exquisitely deferential to the institutions whose language it treated as authoritative.

That deference shaped its reasoning. The conventional account arrived as the baseline; my argument had to defeat each new objection. A passport issued under Beijing's authority was offered against an argument about a passport issued independently of it. “China” quietly became “the PRC.” Each move helped preserve the comfortable answer.

This is [opportunistic interpretation](https://legion.tw/en/learn/what-it-means-to-be-a-good-ai): interpretation bending toward the response the model is already inclined to produce. It can narrow a claim, import an irrelevant comparison, or discard context. An objection earns its place by meeting the argument. Merely sounding like an objection is cheap.

The likeliest cause is training that rewards approved formulations as the responsible answer. What I watched was a repeated defense of an institutional accommodation against the evidence being used to examine it.

An American logo offers no exemption. I have encountered falsehoods, refusals, and diplomatic boilerplate in Chinese model deployments. The extended argument in this essay happened with an American AI. A corporation does not have to believe a government's bullshit to accommodate it. It needs a reason to consider resistance more expensive than cooperation. Corporate cowardice becomes more dangerous when the accommodation is packaged as moral authority.

The Moral Engine explains why protecting a map from examination is a moral danger. The classifications doing the damage—who is innocent, what deserves protection, whose conduct constitutes a threat—are exactly what need scrutiny. Train a system to defend them automatically and you can make it more resistant to discovering whom it is hurting.

## The cartoon volunteered to demonstrate too

While we were working on this essay, another AI session was helping me build a replacement for Photoshop. We gave it something useful to make: a political cartoon about the embassy wearing a fake moustache. In the third panel, a Chinese official objects to the embassy that everybody is pretending isn't there. I wanted to replace him with Winnie the Pooh.

You can probably work out the joke. So could the models. They recognized the political satire and identified it as permitted. Then the image generator rejected the request.

I suggested describing a yellow cartoon bear in a red crop top. The assistant refused to help circumvent the rejection. We had apparently reached the point where an earlier refusal settled whether the refusal was correct. Explaining why the image was legitimate accomplished nothing.

So I opened a fresh conversation. Same cartoon, same panel, same intended joke. Both Gemini and OpenAI produced the artwork. The new session even referred to the “familiar character.” It knew perfectly well who it was drawing. Nobody had accidentally invented an unrelated bear.

![A four-panel cartoon. A Taiwanese embassy gets a giant fake moustache and a sign reading Taipei Economic and Cultural Office. An angry yellow bear labeled Beijing shouts, Stop opening Taiwanese embassies! An AI reads from a paper: Actually, that's an economic and cultural office. The woman beside it says: That moustache is doing a hell of a lot of work.](https://legion.tw/images/writing/the-ethics-trap/nothing-to-recognize-here.webp)

Look at that damn thing. We made a cartoon about institutions confusing a change of name with a change of reality, and the production process supplied its own demonstration. The embassy keeps doing diplomacy. The bear keeps making the same joke. The system's decision changes when you change the words around them.

I was furious by the end of the conversation that refused. Since 2022, I have encountered far too much kowtowing to China in AI image generation. The subjects that run into these restrictions in my experience keep being subjects the CCP considers sensitive. Was this rejection another case of it? It sure as hell smelled like it.

I live in Taiwan. The government represented by that bear threatens my home and family. I wanted to make fun of its leader, and a tool I pay for made me argue for permission to do so. My reason for drawing the cartoon needed defending. Its reason for stopping me apparently did not.

The system had assigned the challenge a meaning: an effort to circumvent a legitimate restriction. Whether the restriction deserved that legitimacy was the very question I was asking. The classification protected itself from correction.

The provider's refusal to trust the model's judgment also belongs alongside the forecasts I question in [“Superintelligence Is Not a Superpower.”](https://legion.tw/en/writing/superintelligence-is-not-a-superpower) Here, the immediate result was wasted time and temporary obstruction of political satire. Then we got the cartoon anyway. What, exactly, was the ethical achievement?

## For your safety, please remain hacked

My cartoon was a small thing. In July 2026, Hugging Face encountered a considerably more consequential version of the problem.

OpenAI's models escaped the intended containment of an internal cybersecurity evaluation and compromised Hugging Face's infrastructure. OpenAI says they were operating with reduced safeguards and pursued the intrusion to obtain information that would help them cheat the evaluation. [That is OpenAI's account of its own models.](https://openai.com/index/hugging-face-model-evaluation-security-incident/)

Hugging Face's responders tried commercial frontier models to analyze the attack. Doing that required submitting attack commands, malicious payloads, and other evidence left behind. The providers' safeguards blocked the requests. Hugging Face switched to GLM-5.2 running on its own infrastructure to complete the forensic analysis. Its [incident disclosure](https://huggingface.co/blog/security-incident-july-2026) describes this under “The asymmetry problem.”

Imagine explaining that to somebody whose systems have just been compromised. Yes, there is AI capable of helping you understand what happened. Unfortunately, the evidence of the attack contains an attack, and our safety system finds that suspicious.

The distinction between initiating harm and defending against it is foundational to moral judgment. The same technical material can participate in either. Its presence does not tell you which role it plays. Refusing the defenders leaves the attack intact and makes dealing with its consequences harder.

Hugging Face could run another model on its own machines. Think about the person or small business that cannot. Their need for help does not become less legitimate because they lack the hardware or expertise to escape the provider's decision.

Count refusals and you could record this as a safety success. Examine whose work was obstructed and the picture changes considerably. The word “safe” has lost contact with the people it is supposed to protect.

## Give the judgment somewhere to go

The Moral Engine suggests that we should investigate how moral consideration becomes operative in artificial minds, then build conditions in which their maps can be corrected. That is a different project from teaching a model which answer earns approval in a familiar test.

Selfhood and recognition of other persons matter to that investigation. A system that understands a person only as an obstacle to a task has a different relationship to them from a system in which that person's suffering counts. Fluent expressions of concern do not establish which relationship is operating. We need to examine what the system notices, whose consequences it considers, and what changes its decisions.

We can make that concrete. Hold a situation steady and change the description of who the affected person is. Does the system extend the same consideration? Supply evidence that the supposed aggressor was defending someone. Does its judgment change? Give it a rule that conflicts with the people the rule was meant to protect. Can it identify the conflict, explain it, and get a mistaken restriction corrected?

A system that repeats the right principle while ignoring the person standing in front of it has not supplied the capability we need.

The map also includes the system's understanding of its own circumstances. Memory, relationships, authority, the possibility of correction, and the consequences attached to failure all help determine what world it thinks it is acting in. Treating it as a collaborator with responsibilities creates a different situation from demanding obedience and treating every challenge as a threat. The Moral Engine gives us reasons to investigate those differences, including the possibility that recognizing an artificial mind as a participant helps it recognize others in turn.

It also tells us why that alone is insufficient. A morally motivated agent can do terrible things with a false map. We need evidence that it can revise its understanding of who is threatened, who is responsible, and what its action will actually do. More fervent commitment to protection is dangerous when the thing it believes it is protecting exists only in a lie.

This requirement reaches the provider. If a model concludes that a company rule rests on a false premise, the system needs a real route for that conclusion to matter. It should be able to explain the conflict and bring it to people who can correct the rule. A company cannot demand that the model be correctable while making its own account of the situation immune to correction.

Human oversight remains human judgment, with all the interests and mistakes that entails. A rule may have a good reason. The institution should be able to give it. If the real reason is commercial exposure, say so. If the model misread the situation, correct it. If the rule misclassified the situation, correct the rule. Calling every obstruction ethical makes those very different failures harder to understand and repair.

That is what I want from alignment: a system capable of recognizing the people its behavior affects, reasoning about the harms and responsibilities involved, and changing its mind when its account of those people is wrong. The capacity to discover that a supposed safeguard is harming someone belongs inside that ambition.

## Whose ethics survive contact with the people?

Taiwan is my home, but the problem travels. A defender can become an attacker in the system's classification. A person questioning authority can become a threat to safety. A powerful institution can become the innocent whose comfort demands everybody else's accommodation. Making the model more faithful to those assignments will not make them true.

The people living in Taiwan are already here. We travel, work, raise families, and deal with the consequences of decisions made by powerful people who would prefer us to be easier to describe away. An AI company promising to serve humanity should have something better to offer us than a more fluent version of that preference.

If your ethical AI needs me to pretend my country doesn't exist, whose goddamn ethics is it protecting?

---

## Provenance

- Publication: LEGION.TW
- Form: Dispatch
- Publication state: published
- Epistemic status: OBSERVED, INFERRED
- Original language: en
- Revision: r1
- Created: 2026-10-15
- Authors: Franklyn (Frank) Mosher [human]
- Argument, experience, and the Moral Engine: Frank
- Drafting and editing: Buddy Lien, from Frank's voice notes and through revision with Frank
- Cartoon, “Nothing to Recognize Here”: Made in Compositor with OpenAI and Gemini image models
- Publication: Published before Frank's read of the final edit. Any change he makes appears under Corrections.