Responsible AI Principles, Copyright in AI Output and Korea's AI Basic Act: What to Keep in Mind When Building and Using AI
The last post covered attacks on LLMs and where to stop them. This one covers the problems that show up even without an attacker. When an AI is unfair or cannot explain itself, which principle has it broken? Who owns what an AI produces, and what does Korean law require, and of whom? This post covers Microsoft's responsible AI principles, the U.S. Copyright Office's position on AI output, and the transparency duties and fines in Korea's AI Basic Act. It does not cover the detailed obligations for high-impact AI, Korea's copyright position, or other countries' regulations. I got every practice question right, so there are no misses to note this time either.
1. Six principles of responsible AI
Many companies and institutions have published principles for building and using AI, and one of the most often cited sets is Microsoft's six. The names alone are easy to mix up, so I grouped them by what goes wrong when each one is broken. The right-hand column is made-up examples for illustration.
| Principle | Microsoft's one line | What breaking it looks like |
|---|---|---|
| Fairness | AI systems should treat all people fairly | A résumé-screening AI keeps rating applicants from one region lower. Biased training data is often the cause. |
| Reliability and safety | AI systems should perform reliably and safely | After a process recipe changes, an anomaly detection model quietly starts getting things wrong. |
| Privacy and security | AI systems should be secure and respect privacy | A support chatbot mixes another customer's details into its answer. |
| Inclusiveness | AI systems should empower everyone, regardless of background | A voice assistant only understands standard pronunciation, so some workers cannot use it. |
| Transparency | AI systems should be understandable | Engineers cannot tell why the model flagged this wafer as defective. Nobody can tell whether a text was written by AI or a person. |
| Accountability | People should be accountable for AI systems | An AI decision stops a production line and nobody takes responsibility, only "the AI decided". |
Two pairs are easy to confuse.
- Fairness and inclusiveness: fairness is about the AI judging people differently; inclusiveness is about some people being unable to use it at all.
- Transparency and accountability: transparency asks "can we tell why it decided that?"; accountability asks "who answers for it when it goes wrong?"
When picking a principle for a scenario, first say in one sentence what went wrong, then match it against the right-hand column above.
2. Copyright in AI output
In January 2025, the U.S. Copyright Office set out its view on AI output in Part 2 of its AI report.
- Providing prompts alone does not earn copyright protection, however detailed they are.
- Where a human-authored work is perceptible in the output, that part can be protected.
- Creative arrangements or modifications a person makes to the output can be protected too.
- Using AI to assist the creative process, or including AI-generated material in a larger human-made work, does not take the whole work out of protection.
- Existing copyright principles are enough, and the Office did not find a case for new protection specifically for AI output.
Splitting an output by what a person did, here is which parts can be protected.
| Part of the output | Protection |
|---|---|
| A very detailed prompt | Providing prompts alone is not protected |
| Text I wrote myself, fed in as input and still present in the output | Can be protected |
| How I selected and arranged several AI results, and text I edited myself | Can be protected |
| The rest that the AI produced | Not protected |
Only the parts a person contributed are protected, and mixing in AI-made material does not strip protection from the human parts. This is the U.S. position; I did not check Korea's this time.
3. Korea's AI Basic Act
Its full name is the Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust, and it took effect on January 22, 2026 (Korea Policy Briefing). This post looks only at the transparency duties (Article 31) and fines (Article 43).
The duties fall on AI business operators. According to a legal commentary, these split into developers, who build and provide AI, and deployers, who use that AI to provide products or services (Bae, Kim & Lee commentary). A company that plugs an outside LLM into a service it offers is a deployer. Individual users who only use a service carry no notice or labeling duties.
Article 31 requires three things.
| Article 31 | When | What |
|---|---|---|
| Para. 1, advance notice | When providing products or services using high-impact or generative AI | Tell users in advance that it runs on AI |
| Para. 2, output labeling | When putting out generative AI output | Label it as made by generative AI |
| Para. 3, virtual output | When providing virtual sound, images or video hard to tell from the real thing | Give notice or a label users can clearly recognize. For artistic or creative works, it can be done in a way that does not hinder enjoyment |
- The exception in paragraph 3 is not an exemption. The label still has to be there; only its form can be adjusted. The article does not use the word "deepfake"; it says virtual output "hard to tell from the real thing".
- The specific labeling methods are set by the enforcement decree. According to the government, outputs such as animation or webtoons can use a visible mark, an invisible digital watermark, or a notice.
- High-impact AI is AI used in areas that can significantly affect people's life, physical safety or fundamental rights. According to the government, systems where a person makes the final decision are treated as controllable and fall outside the high-impact category.
The fine under Article 43 is up to 30 million won. It applies to failing to give the advance notice in paragraph 1, failing to appoint a domestic representative, and ignoring a suspension or corrective order. Breaching the output labeling duty in paragraph 2 is not on that list. The government said it would hold off on fines for at least a year after the Act took effect.
4. What stays with the user, law or no law
The legal duties sit with operators, but the judgment of the person using AI still matters on its own. In Anthropic's AI Fluency 4D framework, Diligence is "ensuring you interact with AI responsibly" (Anthropic Academy), and the scoring criteria for Programmers' AI assessment include responsible use that avoids entering sensitive material or giving unethical instructions (VentureSquare). Gathered as behaviors from the earlier posts in this series:
| Behavior | Where it came from |
|---|---|
| Use only company-approved tools, and remove confidential and personal data before pasting | Data leaving through prompts in the data post, sensitive information disclosure in the security post |
| Do not hand in AI output as is; check its sources and numbers | Hallucination in the LLM limits post, the evaluation post |
| Check whether the output disadvantages a particular group | Bias in the data post, fairness in section 1 here |
| Responsibility for the result lies with the person who used the AI, not the AI | Accountability in section 1 here |
| Disclose AI assistance where a rule asks for it | Transparency in section 1 here |
Practice: pick the principle or rule
- A résumé-screening AI keeps rating applicants from one region lower. Which principle is most directly broken?
Answer
Fairness. The AI judges people differently. Inclusiveness is about some people being unable to use it at all.
- Engineers cannot tell why an equipment anomaly detection AI flagged this wafer as defective. Which principle is most relevant?
Answer
Transparency. Nobody can tell why it decided that.
- An AI decision stopped a production line, and everyone just says "the AI decided"; nobody takes responsibility. Which principle is broken?
Answer
Accountability. Someone has to be responsible for an AI system.
- Under the U.S. Copyright Office's position, is an image generated from nothing but a very detailed prompt protected?
Answer
No. Providing prompts alone is not enough, because a person did not determine the expression in the output.
- Someone heavily edited an AI-written draft and arranged it alongside text they wrote themselves to make a report. Which description is right?
Answer
The parts the person edited and arranged, and the parts they wrote, can be protected. The AI-made parts themselves are not.
- Who carries the advance notice and output labeling duties in Article 31 of the AI Basic Act?
Answer
AI business operators. Individual users carry no notice or labeling duties.
- When did the AI Basic Act take effect?
Answer
January 22, 2026. The government said it would hold off on fines for at least a year.
- A video generation service provides virtual video that is hard to tell from the real thing. What does the law require?
Answer
Notice or a label that users can clearly recognize. For artistic works it can be done in a way that does not hinder enjoyment, but the labeling duty is not waived.
Recap
- Microsoft's responsible AI principles are fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability.
- Fairness is about judging people differently, inclusiveness about some people being unable to use it. Transparency asks "why", accountability asks "who".
- Under the U.S. Copyright Office's position, prompts alone are not protected; only expression a person put in and creative arrangement or editing can be.
- Korea's AI Basic Act took effect on January 22, 2026, and its notice and labeling duties fall on AI business operators, not users.
- Article 31 requires advance notice, output labeling, and clear notice or labeling for output hard to tell from the real thing. Fines go up to 30 million won and apply to missing advance notice.
- Separately from the law, the person using AI is still responsible for using approved tools, removing sensitive data, checking the output and owning the result.
