Can I use a real person's voice or face in my AI character?
Not without that person's permission, in writing, covering the use you have in mind. If the person is a celebrity you have no contract with, assume the answer is no. This is the highest-risk thing you can do with an AI character, and the risk doesn't come from copyright. It comes from laws that protect a person's identity: their face, their voice, their name.
That's why the usual AI-creator reasoning doesn't help here. "I made it myself", "the model generated it", "I changed it a bit": none of these answer the question, because the question isn't who made the file. It's whose identity is in it. This post goes through what the official texts actually say in the US and in Taiwan, the patterns that cause trouble, and what a registration with us can and can't do about any of it.
This is general information, not legal advice. If a real person's face or voice is the point of your project, talk to a lawyer where you publish.
Copyright doesn't cover a face or a voice
The US Copyright Office says this directly in Part 1 of its report on copyright and AI, Digital Replicas, published in July 2024. On page 17: "Copyright does not, however, protect an individual's identity in itself, even when incorporated into a work of authorship. A replica of their image or voice alone would not constitute copyright infringement." A footnote on the same page quotes the Ninth Circuit in Midler v. Ford Motor Co. (1988): "A voice is not copyrightable." (Part 1 report, PDF)
Two things follow. First, a copyright-style defence like "I didn't copy any of their photos" doesn't settle anything. You can generate a face from scratch and still use someone's likeness. Second, the person you're imitating doesn't need to own a copyright to come after you. Their claim is about identity, and it runs on different rules.
Taiwan draws the same line. A May 21, 2026 copyright briefing from Taiwan's Intellectual Property Office puts it in two lines: the photo is a matter of copyright, owned by the photographer; the person in it is a matter of 肖像權 (portrait rights), which "屬民法人格權範疇，與照片的著作權無涉", meaning it belongs to Civil Code personality rights and has nothing to do with the photo's copyright (TIPO briefing, PDF).
If you want the copyright side of AI work, how much you have to do for an AI work to be protected covers it. This post is about the other half.
In the US: the right of publicity, state by state
There's no single US federal law for this. Most of it is state law, usually called the right of publicity, and the states differ on what's covered, for how long, and what counts as a use.
California. The Copyright Office's report (page 12, footnote 75) describes California Civil Code §§ 3344 and 3344.1 as "protecting against the unconsented commercial use of a person's name, voice, signature, photograph, or likeness, affording a 70-year postmortem term". So in California, a dead celebrity's estate can still have a say for decades.
Tennessee. In March 2024 Tennessee passed the ELVIS Act (Ensuring Likeness, Voice, and Image Security Act of 2024, Public Chapter No. 588, effective July 1, 2024). We read the enrolled text on the Tennessee Secretary of State's site. Three parts matter for AI characters:
- It defines "voice" as "a sound in a medium that is readily identifiable and attributable to a particular individual, regardless of whether the sound contains the actual voice or a simulation of the voice of the individual." A clone counts. So does a model that just sounds like them.
- It makes a person liable who "publishes, performs, distributes, transmits, or otherwise makes available to the public an individual's voice or likeness" knowing the use wasn't authorized. That isn't limited to advertising. Posting is enough.
- It reaches tools, too: distributing an "algorithm, software, tool, or other technology" whose "primary purpose or function" is producing a particular, identifiable person's photograph, voice or likeness, without authorization. A voice model or a LoRA trained to reproduce one specific real person is close to what that sentence describes.
The Act keeps exceptions, "to the extent such use is protected by the First Amendment", for news, "comment, criticism, scholarship, satire, or parody", and uses that are "fleeting or incidental", among others. Whether your piece fits one is exactly the kind of question a lawyer should answer, not a blog post.
Federal. The Copyright Office's July 2024 report recommended that Congress "establish a federal right that protects all individuals during their lifetimes from the knowing distribution of unauthorized digital replicas" (page 57). That's a recommendation in a report, not a law. Check the current state of federal legislation before you rely on anything either way.
In Taiwan: personality rights in the Civil Code
Taiwan's Civil Code doesn't have a statute named after voice or likeness. It works through general personality-rights articles, which we read on the national law database on November 10, 2026:
- Article 18: when a personality right is infringed, you can ask the court to remove the infringement, and to prevent it when it's threatened. Damages for it need a specific legal basis (民法 第 18 條).
- Article 19: a person whose right to their name is infringed can ask for removal and damages (第 19 條). Relevant if you put a real person's name on the character, not just their face.
- Article 184: the general rule that intentionally or negligently infringing someone's rights makes you liable for damages (第 184 條).
- Article 195: non-pecuniary damages for infringing body, health, reputation, liberty, credit, privacy, chastity, or "其他人格法益而情節重大者", other personality interests where the case is serious (第 195 條).
Portrait isn't named in Article 195's list, and voice isn't named anywhere in these articles. TIPO's briefing places portrait rights within Civil Code personality rights. We didn't find an official source on voice specifically, so we won't guess how a court would treat a cloned voice.
There's also a criminal provision for the worst case. Criminal Code Article 319-4 punishes making, with intent to distribute, "以電腦合成或其他科技方法製作關於他人不實之性影像", false sexual images of another person made by computer synthesis or other technology, where it is liable to harm them, with up to five years in prison; distributing them is punished the same way, and doing it for profit raises the maximum to seven years (刑法 第 319-4 條).
Patterns that cause trouble
None of these is a legal conclusion. They're the places where creators most often walk into the question without noticing.
| What you're doing | Why it's a problem |
|---|---|
| Training a LoRA or voice model on one real person, then sharing it | Tennessee's Act names tools whose primary purpose is producing a particular person's likeness or voice |
| A character that's "obviously" a famous actor or singer, under a new name | Changing the name doesn't change whose face it is. The Copyright Office's report (page 13, footnote 79) cites a New York case where a look-alike in an ad infringed Jacqueline Onassis's right of publicity |
| A voice that "just sounds like" a known singer, no clone involved | Tennessee's definition covers a simulation, "regardless of whether the sound contains the actual voice" |
| Your friend's or partner's face, because they said "sure" once | Permission that isn't written down, or doesn't cover selling, is hard to prove later |
| A deceased star | California's protection runs 70 years after death, per the report's description; Tennessee's Act names heirs and executors as the ones who give consent |
| Anything sexual, with any real person | Taiwan's Criminal Code Article 319-4, and in most places far worse than a civil claim |
The safe pattern is the boring one: a face and a voice that don't belong to anyone. What makes an AI character recognisable is about building a character people know at a glance, without borrowing someone else's identity to do it.
If you do have permission
Sometimes the real person is in on it: your own face, a voice actor you hired, a model who signed up for this. Then the job is to make the permission provable later.
- Get it in writing, signed, dated. A chat message saying "sure, go ahead" is weak if the relationship goes sour.
- Name the specific things. Face, voice, name, or all three. Cloning or training a model, or only recording. AI-generated variations or only the original takes.
- Name the uses. Personal posts, commercial releases, ads, selling or licensing the character to someone else. A character that might be sold later needs that in the permission, or the buyer gets something they can't use. Selling vs licensing a character explains why the difference matters.
- Set the term and the exit. How long, and what happens if they withdraw.
- For a minor, the parent or guardian signs. Tennessee's text says so explicitly, and it's the sensible default everywhere.
Keep that document yourself. It's the thing that answers the question, not anything a platform stores.
What OpenStela does, and doesn't do
Our Terms, Section 5, list the likeness of a real person without their permission first among the things you must not use the service for. That includes public figures, celebrities, and deceased persons where local law protects them. The instructions our MCP server gives AI assistants say the same thing: never register a real person.
Here's what we don't do. We don't scan faces or voices. In September 2026 we removed all visual comparison from the platform, face matching included; the Terms say the service does "no marker detection, palette comparison or likeness scoring". We also don't check the rights you declare. When you register a work, you pick My own work or I hold a licence, and the public record shows that as your declaration, nothing more.
So a registration is not a clearance. A record shows that a specific file existed on a date, under the name you gave, with the contribution you wrote down. What a character ID proves, and what it doesn't has the full version. If the file uses someone's face without consent, the record only fixes the date you put it out there.
If you see a public character, work or post on OpenStela that uses a real person without permission, sign in, open the ⋯ menu and choose Report post. One of the reasons is Illegal content, and every report goes to the platform team. If it's your own face or voice, add that in the note.
我的 AI 角色可以用真人的聲音或臉嗎？
不行——除非你拿到本人的同意，而且是白紙黑字、涵蓋你打算怎麼用的那種。如果對方是你根本沒簽過約的藝人、網紅、明星，就直接當成答案是「不行」。用真人的臉或聲音做 AI 角色，是 AI 創作裡風險最高的一件事；而這個風險不是來自著作權，是來自保護一個人「身分」的法律：他的臉、他的聲音、他的名字。
所以 AI 創作者平常的那套說法，在這裡都不管用。「這是我自己算出來的」「是模型生成的」「我有改過」——這些都回答不了問題，因為問題不是「檔案是誰做的」，而是「檔案裡是誰的身分」。下面依序整理美國和台灣的官方條文怎麼寫、最容易出事的幾種做法，以及在 OpenStela 登記一筆紀錄，能做到什麼、做不到什麼。
本文為一般資訊，不是個案法律意見。如果你的企劃本身就建立在某個真人的臉或聲音上，請先找你發表地的律師談。
臉和聲音，不歸著作權管
美國著作權局在《著作權與人工智慧》報告第一部《數位複製品》（Digital Replicas，2024 年 7 月發布）第 17 頁寫得很清楚：「Copyright does not, however, protect an individual's identity in itself, even when incorporated into a work of authorship. A replica of their image or voice alone would not constitute copyright infringement.」意思是：著作權不保護一個人的身分本身，就算它出現在著作裡也一樣；單純複製某人的形象或聲音，不構成著作權侵害。同頁註腳引了美國第九巡迴上訴法院 1988 年 Midler v. Ford Motor Co. 案：「A voice is not copyrightable.」（聲音不能受著作權保護。）（第一部報告 PDF）
這帶出兩件事。第一，「我沒有抄他任何一張照片」這種著作權式的辯解，在這裡不算數——你可以從零生成一張臉，照樣用到別人的肖像。第二，被你模仿的人，不需要擁有任何著作權就能找你；他主張的是身分，走的是另一套規則。
台灣也是同樣的切法。經濟部智慧財產局 2026 年 5 月 21 日（115 年 5 月 21 日）的著作權宣導簡報，用兩行講完：照片是著作權問題，屬攝影著作、要經拍攝者同意；照片裡的人是肖像權問題，「屬民法人格權範疇，與照片的著作權無涉」（智慧局簡報 PDF）。
AI 作品的著作權那一半，請看〈AI 作品，我要做多少才會受保護？〉。這篇講的是另一半。
美國：公開權，各州各自規定
美國沒有一部統一的聯邦法管這件事。主要是州法，通常叫「公開權」（right of publicity），各州在保護什麼、保護多久、怎樣算使用上都不一樣。
加州。 著作權局報告第 12 頁註 75 描述加州民法典第 3344 條與第 3344.1 條：「protecting against the unconsented commercial use of a person's name, voice, signature, photograph, or likeness, affording a 70-year postmortem term」——禁止未經同意商業使用他人的姓名、聲音、簽名、照片或肖像，並在本人死後保護 70 年。也就是說，在加州，過世明星的遺產管理方幾十年後仍有發言權。
田納西州。 田納西 2024 年 3 月通過 ELVIS Act（Ensuring Likeness, Voice, and Image Security Act of 2024，Public Chapter No. 588，2024 年 7 月 1 日施行）。我們讀的是田納西州務卿網站上的公布版本。對 AI 角色來說，有三處最要緊：
- 「聲音」的定義是：媒介中可辨識、可歸屬於特定個人的聲音，「regardless of whether the sound contains the actual voice or a simulation of the voice of the individual」——不論是真正的聲音，還是模擬出來的。聲音複製算，只是「聽起來像」的模型也算。
- 明知未經授權，仍「publishes, performs, distributes, transmits, or otherwise makes available to the public」（發表、演出、散布、傳輸或以其他方式公開提供）他人的聲音或肖像，就要負民事責任。不限於廣告，公開發出去就算。
- 連工具也管：散布主要目的或功能在於產生特定可辨識個人的照片、聲音或肖像的「algorithm, software, tool, or other technology」（演算法、軟體、工具或其他技術），而未經授權者，同樣要負責。專門拿某一位真人訓練、用來重現他的聲音模型或 LoRA，很接近這句話描述的東西。
這部法保留了例外，「to the extent such use is protected by the First Amendment」（在受美國憲法第一修正案保護的範圍內），包括新聞、「comment, criticism, scholarship, satire, or parody」（評論、批評、學術、諷刺或戲仿）、「fleeting or incidental」（短暫或附帶出現）等。你的作品算不算，正是該問律師、而不是看部落格的問題。
聯邦層級。 著作權局 2024 年 7 月的報告建議國會「establish a federal right that protects all individuals during their lifetimes from the knowing distribution of unauthorized digital replicas」（建立聯邦層級的權利，在每個人在世期間，保護他不被明知而散布未經授權的數位複製品）（第 57 頁）。那是報告裡的建議，不是法律。聯邦立法現在走到哪一步，要用之前請自己再查。
台灣：民法的人格權
台灣民法沒有一條叫「聲音權」或「肖像權」的條文，而是透過人格權的一般規定處理。以下條文我們在 2026 年 11 月 10 日於全國法規資料庫讀過（資料庫顯示民法最新修正公布日為 115 年 8 月 17 日）：
- 第 18 條：「人格權受侵害時，得請求法院除去其侵害；有受侵害之虞時，得請求防止之。」第二項規定，請求損害賠償或慰撫金，以法律有特別規定者為限（民法 第 18 條）。
- 第 19 條：「姓名權受侵害者，得請求法院除去其侵害，並得請求損害賠償。」角色不只用了臉、還掛上真人名字時要注意（第 19 條）。
- 第 184 條：因故意或過失，不法侵害他人之權利者，負損害賠償責任（第 184 條）。
- 第 195 條：不法侵害他人之身體、健康、名譽、自由、信用、隱私、貞操，「或不法侵害其他人格法益而情節重大者」，被害人雖非財產上之損害，亦得請求賠償相當之金額（第 195 條）。
第 195 條的列舉裡沒有「肖像」兩個字，這幾條也都沒提到「聲音」。智慧局簡報把肖像權放在民法人格權的範疇裡；至於聲音，我們沒有找到專門談它的官方來源，所以不替法院猜一段被複製的聲音會怎麼被看待。
最糟的那一種情況，另外有刑責。刑法第 319-4 條：意圖散布、播送、交付、公然陳列或以他法供人觀覽，「以電腦合成或其他科技方法製作關於他人不實之性影像，足以生損害於他人者」，處五年以下有期徒刑、拘役或科或併科五十萬元以下罰金；散布這類影像者亦同；意圖營利者，處七年以下有期徒刑（刑法 第 319-4 條；資料庫顯示刑法最新修正日 115 年 7 月 22 日）。
最容易出事的幾種做法
下表都不是法律結論，只是創作者最常在沒察覺的情況下踩進去的地方。
| 你在做的事 | 為什麼有問題 |
|---|---|
| 拿某一位真人的照片或錄音練 LoRA、聲音模型，然後分享出去 | 田納西 ELVIS Act 直接點名「主要目的在產生特定個人肖像或聲音」的工具 |
| 一看就是某位男神女神的角色，只是換了名字 | 換名字不會換掉那張臉是誰的。著作權局報告第 13 頁註 79 提到紐約一案：廣告裡用了賈桂琳・歐納西斯的相似替身，被認定侵害她的公開權 |
| 沒有複製，只是「聲線很像」某位歌手 | 田納西的定義連模擬聲音都算進去，不論是不是本人的聲音 |
| 用朋友、另一半的臉，因為他有一次說「好啊」 | 沒有寫下來、或沒講到能拿去賣的同意，日後很難證明 |
| 已經過世的巨星 | 依報告描述，加州保護到死後 70 年；田納西法條明定由遺囑執行人、繼承人等人同意 |
| 任何跟性有關的內容，只要牽涉真人 | 台灣刑法第 319-4 條；在多數地方，這都遠比民事求償嚴重 |
最安全的做法也最無聊：一張不屬於任何人的臉、一副不屬於任何人的嗓子。想讓角色一眼被認出來、又不借別人的身分，可以看〈AI 角色靠什麼被認出來？登錄簿上三隻公開角色的設計筆記〉。
如果你真的拿到同意
有時候真人本來就是合作的一方：你自己的臉、你請來的配音員、簽約拍素材的模特兒。這時要做的，是讓這份同意日後拿得出來。
- 書面、簽名、寫日期。 LINE 上一句「可以啊你用」，關係鬧僵時很薄弱。
- 寫清楚是哪些東西。 臉、聲音、名字，還是三樣都有；可不可以複製聲音、拿來訓練模型，還是只能用原始錄音；可不可以讓 AI 生出新的變化。
- 寫清楚用途。 個人貼文、商業發行、廣告、把角色賣給別人或授權出去。角色將來可能出售的話，同意書裡就要涵蓋這點，否則買家拿到的是一個不能用的東西。讓與和授權差在哪，見〈賣斷和授權角色，法律上差在哪？〉。
- 寫清楚期間與退出。 用多久、對方撤回同意時怎麼辦。
- 未成年人由法定代理人簽。 田納西條文明文這樣寫，在哪裡都是合理的預設。
這份文件你自己收好。能回答「有沒有同意」的是它，不是任何平台存的東西。
OpenStela 會做什麼、不會做什麼
我們的服務條款第五條「使用規範」，把「未經同意之真實人物肖像」列為第一項，包含公眾人物、名人，以及當地法律仍予保護的已故者。我們的 MCP 伺服器給 AI 助理的使用說明也寫了同一句：不要替使用者登記真人。
我們不做的事也要講清楚。平台不掃描臉、不比對聲音：2026 年 9 月，我們把所有視覺比對都拿掉了，人臉比對也在內，條款寫明本平台「不進行特徵偵測、色票比對或相似度評分」。我們也不檢查你聲明的權利。登記作品時，你選「這是我的作品」或「我已取得授權」，公開紀錄上顯示的就是「登記人聲明」，僅此而已。
所以，登記不是放行。一筆紀錄能說明的是：某個檔案在某一天、以你給的名稱存在，附上你自己寫的貢獻說明。完整的說法見〈角色編號證明什麼，不證明什麼〉。如果檔案裡未經同意用了別人的臉，這筆紀錄唯一固定下來的，是你把它公開的那一天。
如果你在 OpenStela 上看到公開的角色、作品或貼文未經同意使用真人，登入後點右上角「⋯」選單裡的「檢舉貼文」，原因可以選「違法內容」，檢舉會送到平台團隊。如果被用的是你自己的臉或聲音，請在備註裡寫明。
Questions常見問題
Is it fine if it's parody or satire?如果是諷刺或戲仿，就可以嗎？
Sometimes, and it depends on where you are. Tennessee's Act keeps a First Amendment exception that names satire and parody, but whether a particular piece qualifies is decided case by case. A character you plan to sell or license isn't usually what people mean by parody. Get advice before you rely on it.
有時候可以，要看你在哪裡。田納西 ELVIS Act 保留了第一修正案的例外，明列諷刺與戲仿，但個別作品算不算，是逐案判斷。一隻你打算拿去賣或授權的角色，通常不是大家說的「戲仿」。要倚靠這點之前，先問過律師。
The person died years ago. Is their likeness free to use?那個人已經過世很多年，肖像就能自由使用了嗎？
Not necessarily. The Copyright Office's report describes California's protection as lasting 70 years after death, and Tennessee's Act names executors, heirs and devisees as the ones who give consent for a deceased person. Other places differ. Don't assume death ends it.
不一定。著作權局報告描述加州的保護到死後 70 年；田納西的條文則明定已故者由遺囑執行人、繼承人或受遺贈人同意。其他地方又各有不同。別以為人走了，權利就跟著結束。
Can I use my own face or voice?用我自己的臉或聲音可以嗎？
Yes, it's your identity, so the identity question mostly goes away. Two things remain: the AI tool's terms for uploading and cloning, and the usual copyright question about AI output. If you might sell the character later, decide now what the buyer gets to do with your face. Write it into the deal.
可以，那是你自己的身分，這部分的問題大致就消失了。剩下兩件事：AI 工具對上傳與聲音複製的條款，以及 AI 產出物本來就有的著作權問題。如果角色以後可能賣掉，現在就想好買家可以拿你的臉做什麼，並寫進交易裡。
I hired a voice actor. Does paying them cover an AI clone?我付錢請了配音員，這樣就能做 AI 聲音複製嗎？
Only if your contract says so. A fee for recording a session isn't automatically permission to train a model on the voice or to generate new lines. Put cloning, model training, commercial use and transfer in writing, and keep the signed copy.
只有合約裡寫到才算。付錢錄一次音，不等於同意你拿他的聲音訓練模型、生成他沒講過的台詞。把聲音複製、模型訓練、商業使用、能不能轉讓都寫進書面，簽好的那份自己留著。