Are my prompts copyrightable?
Maybe, but probably not the way you're hoping. There are two questions here. Can the words you typed be protected, the way a paragraph or a poem can? And does a good prompt make the image, track or clip the model gave back yours? The U.S. Copyright Office answers them separately, and the answers differ.
Short version: a prompt with enough original expression of its own may be protected as text. A keyword string like "cinematic lighting, 8k, volumetric fog" almost certainly isn't. And even a protected prompt doesn't make you the author of the output. Below is where each answer comes from and what to keep. This is general information, not legal advice.
Two questions hiding in one
The Copyright Office's Registration Guidance for works containing AI material (published in the Federal Register on March 16, 2023) puts the split in one footnote:
"While some prompts may be sufficiently creative to be protected by copyright, that does not mean that material generated from a copyrightable prompt is itself copyrightable." (footnote 27)
The Office's Copyright and Artificial Intelligence, Part 2: Copyrightability, published January 29, 2025, comes back to it. Summarizing public comments, the report notes that "prompts themselves, if sufficiently creative, may be copyrightable," and then: "The copyright status of the output generated, however, is a separate question" (p. 13).
So keep two boxes: your text, and the model's output. Nothing moves between them automatically.
The prompt as text: when it can be protected
A prompt is words. Words can be a literary work if they carry enough original expression. U.S. law protects "original works of authorship", but under 17 U.S.C. § 102(b) protection never extends to "any idea, procedure, process, system, method of operation, concept, principle, or discovery, regardless of the form in which it is described."
Two things in that rule cut against most prompts.
Short phrases. The Copyright Office's Circular 33, Works Not Protected by Copyright (revised March 2021) says words and short phrases "are uncopyrightable because they contain an insufficient amount of authorship," and that the Office "will not register individual words or brief combinations of words, even if the word or short phrase is novel, distinctive, or lends itself to a play on words." A string of style tags is a list of short phrases.
Instructions and methods. The same circular says "a simple set of directions" is uncopyrightable. Its example is a recipe: a list of ingredients and steps can't be registered, while a recipe that "creatively explains or depicts" how to do something may be, as text. Many prompts read like recipes: subject, lens, lighting, aspect ratio, negative terms. The technique a prompt encodes, the "add this weight to that token" trick that took you a week to find, sits on the method side of § 102(b).
What might clear the bar is a prompt that reads like writing: a few hundred words of scene description in your own voice, a short story, a set of lyrics. Those are judged like any other text. The Office doesn't publish a word count, and registration decisions are case by case.
Taiwan draws the same line in its statute. Article 10-1 of the Copyright Act (last amended June 15, 2022) says, in the Ministry of Justice's English translation, that protection "shall only extend to the expression of the work in question, and shall not extend to the work's underlying ideas, procedures, production processes, systems, methods of operation, concepts, principles, or discoveries." Article 9 excludes "slogans and common symbols, terms, formulas" and similar items. Article 5 lists literary works as a category. In a letter on lyrics, 電子郵件1140310 (March 10, 2025), Taiwan's IP Office restated the general test: a work needs originality (not copied) and a minimum of creativity, and must not fall under Article 9. We didn't find an official Taiwanese source that addresses prompts as text directly, so we won't say more than that.
What protection would actually cover
Suppose your prompt is protectable. What you'd have is control over copying that text: someone pasting your 400 words into their own tool, or reposting them in a prompt pack. You would not have:
- the idea or the look you were going for,
- the technique ("use this sampler, this weight, this negative prompt"),
- anyone else's prompt that reaches a similar result in their own words.
What makes a prompt useful is usually the method, and the method is the part the law leaves open.
The output: why a good prompt doesn't carry over
"The Office concludes that, given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output. Prompts essentially function as instructions that convey unprotectible ideas" (Part 2, p. 18).
The report's demonstration is a detailed prompt for a "bespectacled cat in a robe reading the Sunday newspaper and smoking a pipe". The image followed some instructions, ignored others, and "where no instructions were provided, the AI system filled in the gaps" (p. 19). The Office "re-ran the prompt above and received a much different image" (p. 20, n. 107). Its reasoning: prompts "may reflect a user's mental conception or idea, but they do not control the way that idea is expressed" (p. 19).
Iterating doesn't help. "By revising and submitting prompts multiple times, the user is 're-rolling' the dice, causing the system to generate more outputs from which to select, but not altering the degree of control over the process" (p. 20). Choosing the best of forty isn't authorship either: "providing instructions to a machine and selecting an output does not equate to authorship" (p. 21).
The conclusion repeats it: "Prompts alone, however, at this stage are unlikely to satisfy those requirements" (p. 41). Note "at this stage". The report says "there may come a time when prompts can sufficiently control expressive elements" (p. 21). This is the current position, not a permanent rule.
Taiwan's IP Office answered a nearly identical question. Asked whether images generated by typing instructions into ChatGPT are protected, 電子郵件1140522c (May 22, 2025) said it depends on whether there was actual human creative input: AI used as an assisting tool with real human input can produce a protected result, while content completed entirely by the AI's computation, without human creative input, is not protected. Courts decide disputes.
Where your words do reach the output
There's one case where your text crosses over: when it shows up in the output as written. The report calls these "expressive inputs" and says they "are different from those that merely communicate desired outcomes" (p. 22). Its example: uploading a first-person story and asking the system to convert it to third person. Your story is still there.
For creators, the everyday version is lyrics. If you write the lyrics and paste them into a music generator, the words sung in the track are yours. The style description in the other box ("dreamy synthpop, female vocal, 110 bpm") is a prompt in the ordinary sense. Taiwan's IP Office said in the same March 2025 letter that lyrics you write yourself are protected when you finish them, with no registration needed, while a song completed entirely by the AI's computation is not.
The same goes for a script or dialogue you feed to a video model: your words are yours, the pictures and sound are a separate question. For the full ladder of human contribution (sketches, edits, arrangement), see How much do I have to do for an AI work to be protected?
What to keep, and what to keep private
Protectable or not, your prompts are evidence of what you did. Cheap habits:
- Keep a prompt log with dates. The prompt, the tool and model version, the seed if there is one, and which output you kept. If anyone asks later what your contribution was, this is where the honest answer comes from.
- Separate your text from your settings. Lyrics, scripts and long written passages in one file; style tags and parameters in another. The first might be a work in its own right. The second is a method.
- Keep the raw output and your edited version. If you repaint, cut or re-record, the difference between the two files is your contribution, and that's the part with the strongest claim.
- Decide what to publish. If a prompt is part of your edge, copyright probably won't stop someone reusing the technique once it's public. Keep it in your own files.
- Read your tool's current terms. Rights in inputs and outputs, and commercial use, depend on your plan and change over time. Check the generator's own terms page.
Where OpenStela fits
OpenStela can't register a prompt on its own. Uploaded works are audio, image or video files (mp3, m4a, wav, jpg, png, webp, mp4, webm), so a text file isn't accepted. What you can register is the output and what you said about it.
When you register a work at app.openstela.io/works, ticking This work was made with AI tools makes Your contribution required (up to 600 characters). You also name the tools. Those fields are frozen with the file's SHA-256 fingerprint into an authorship record, and the record is anchored to a public chain in that night's batch. After that, nobody can edit them, including you.
The contribution field is public on a public work. If you don't want your prompt out there, describe what you did in verbs ("wrote the lyrics; wrote a 300-word scene prompt; generated 25 takes; re-timed the chorus") and keep the prompt in your log. We don't check or score that text; it's your statement.
What you get is a dated, verifiable record: this file existed on this date, under this name, with this statement of what a person contributed. It is evidence if a dispute comes up. It doesn't decide whether you hold copyright in the prompt or the output; that's for a court or the Copyright Office. For a series of outputs from one prompt session, see How to date a Midjourney series. On what a record shows and doesn't: What an ID proves.
Registering works is free. Uploads share 300 MB per account, link-type works take almost none, and there's a daily cap on new registrations (see pricing).
This is general information, not legal advice. If a real dispute comes up, talk to a lawyer where you live.
我的提示詞有著作權嗎？
可能有，但多半不是你期待的那種。這題其實藏了兩個問題，大家常常混在一起問：第一個是提示詞這段文字本身——你打的那些字，能不能像一段文章、一首詩那樣受保護？第二個是模型吐回來的東西——咒語下得好，生出來的圖、歌、影片就是你的嗎？美國著作權局把這兩題分開回答，答案也不一樣。
先講結論：一段本身就有足夠原創表達的提示詞，有可能作為文字受保護；像「cinematic lighting, 8k, volumetric fog」這種關鍵字串，幾乎不可能。而且就算提示詞受保護，也不會讓你因此成為產出的著作人。下面逐一說明根據在哪裡、對你存下來的那些咒語代表什麼、以及該留什麼。本文為一般資訊，不是個案法律意見。
一個問題，其實是兩個
美國著作權局 2023 年 3 月 16 日刊登於聯邦公報的〈AI 生成內容著作登記指引〉（Registration Guidance），用一個註腳就把兩題拆開了：
「有些提示詞可能具備足夠的創作性而受著作權保護，但這不代表由受保護的提示詞所生成的內容本身就受著作權保護。」（"While some prompts may be sufficiently creative to be protected by copyright, that does not mean that material generated from a copyrightable prompt is itself copyrightable."，註 27）
2025 年 1 月 29 日發布的報告〈Copyright and Artificial Intelligence, Part 2: Copyrightability〉又提了一次。整理公眾意見時，報告寫道：提示詞本身若有足夠創作性，可能受著作權保護（"prompts themselves, if sufficiently creative, may be copyrightable"）；但「生成內容的著作權地位，是另一個問題」（"The copyright status of the output generated, however, is a separate question"，第 13 頁）。
所以腦中要放兩個抽屜：一個裝你寫的文字，一個裝模型的產出。一個抽屜裡的東西，不會自動跑到另一個抽屜。
提示詞當作文字：什麼時候可能受保護
提示詞就是文字，文字只要有足夠的原創表達，就可能是語文著作。但不論美國或台灣，著作權都只保護表達、不保護背後的想法和方法。美國著作權法 17 U.S.C. § 102(b) 明定，保護不及於任何「思想、程序、製程、系統、操作方法、概念、原理或發現」，不論它以什麼形式被描述。
這條規則有兩個地方，對大多數提示詞不利。
短語。 美國著作權局〈Circular 33：不受著作權保護的作品〉（2021 年 3 月修訂）說，單字與短語「因創作量不足而不受著作權保護」，即使新穎、獨特或是雙關語，著作權局也不會登記個別單字或簡短的字詞組合。一串風格標籤，就是一串短語。
指示與方法。 同一份文件說，「單純的一組操作步驟」不受保護。它舉的例子是食譜：材料清單加上步驟不能登記；但如果是以有創意的方式解說怎麼做、為什麼這樣做，那段文字可能受保護。很多提示詞讀起來就像食譜：主體、鏡頭、光線、比例、負面詞。至於你花一個禮拜才摸出來的「這個詞加權重、那個參數調多少」，屬於 § 102(b) 說的方法那一邊。
比較有機會的，是讀起來像寫作的提示詞：用你自己的語氣寫的幾百字場景描寫、一篇交給影片模型的短篇故事、一份歌詞。這些就和其他文字一樣判斷。著作權局沒有公布字數門檻，登記與否逐案決定。
台灣的條文劃在同一條線上。著作權法（最近修正 111 年 6 月 15 日）第 10 條之 1：著作權的保護「僅及於該著作之表達，而不及於其所表達之思想、程序、製程、系統、操作方法、概念、原理、發現」。第 9 條把「標語及通用之符號、名詞、公式」等排除在外；第 5 條則把語文著作列為著作類別之一。智慧局在一封談歌詞的函釋 電子郵件1140310（114 年 3 月 10 日）重申了一般標準：要有「原創性」（不是抄來的）與「創作性」（最起碼的創作高度），而且不屬於第 9 條不受保護的標的。我們沒有找到智慧局直接談「提示詞文字本身」的官方說明，所以這部分只寫到這裡。
就算受保護，保護的是什麼
假設你的提示詞真的受保護，你能管的是這段文字被照抄：有人把你那四百字整段貼進自己的工具，或截圖收進「咒語包」轉賣。你管不到的是：
- 你想要的那個點子、那種畫面感，
- 技巧本身（用哪個取樣器、哪個權重、哪組負面詞），
- 別人用自己的話寫、但做出類似結果的提示詞。
提示詞真正好用的地方通常是方法，而方法正好是法律留白的那一塊。
產出：咒語下得再好，也不會順帶變成你的
這是多數創作者真正在意的問題。美國著作權局 2025 年的答案很直接：
「本局認為，以目前一般可取得的技術，單靠提示詞無法提供足夠的人為掌控，使 AI 系統的使用者成為產出的著作人。提示詞本質上是傳達不受保護之想法的指示。」（"prompts alone do not provide sufficient human control to make users of an AI system the authors of the output. Prompts essentially function as instructions that convey unprotectible ideas"，第 18 頁）
報告自己做了示範：一段很詳細的提示詞，要一隻「戴眼鏡、穿浴袍、讀週日報紙、抽菸斗的貓」。生出來的圖照做了一部分、漏了一部分，「沒有指示的地方，由 AI 系統自己補上」（第 19 頁）。著作權局再跑一次同一段提示詞，「得到一張很不一樣的圖」（第 20 頁註 107）。結論是：提示詞也許反映了使用者心中的構想，但它控制不了這個構想如何被表達（第 19 頁）。
反覆改也一樣。報告形容反覆修改、送出提示詞是「重新擲骰子」（"re-rolling" the dice），只是讓系統多吐幾張給你挑，並沒有提高你對過程的掌控（第 20 頁）。從四十張裡挑一張也不算創作：「對機器下指示、再挑選一個產出，不等於創作」（"providing instructions to a machine and selecting an output does not equate to authorship"，第 21 頁）。
結論章再說一次：單靠提示詞，「現階段」不太可能符合要求（"Prompts alone, however, at this stage are unlikely to satisfy those requirements"，第 41 頁）。注意「現階段」三個字——報告也說，或許有一天提示詞能足夠掌控表達元素（第 21 頁）。這是目前的立場，不是永久規則。
台灣這邊，智慧局被問過幾乎一樣的題目。有人問「利用 ChatGPT 輸入指令後生成的圖畫」有沒有著作權，電子郵件1140522c（114 年 5 月 22 日）回答：要看創作過程中有無人類實際的創意投入。把 AI 當輔助工具（例如繪圖軟體）、而有人類實際創意投入，成果仍可受保護；若完全由 AI 的演算功能獨立完成、沒有人類精神文明的投入，則無法享有著作權。個案有爭議時，由司法機關依具體事實調查證據認定。
你的文字真的「進到」產出裡的時候
只有一種情況，你的文字會越過那條線：它原封不動出現在產出裡。美國報告稱之為「表達性輸入」（expressive inputs），並說它和「只是傳達想要什麼結果」的提示詞不同（第 22 頁）。報告的例子是：上傳一篇第一人稱寫的故事，請系統改成第三人稱——你的故事還在裡面。
對創作者來說，最常見的就是歌詞。你自己寫好歌詞、貼進音樂生成工具的歌詞欄，歌裡唱出來的那些字是你的；另一格的風格描述（「夢幻合成器流行、女聲、110 bpm」）才是一般意義的提示詞。智慧局在上面那封 114 年 3 月的函釋就說：自行創作的歌詞，於完成時享有著作權，無須登記；而完全由 AI 演算功能獨立完成的歌曲，無法享有著作權。
劇本、台詞也一樣：你寫好一段段丟給影片模型，文字是你的，畫面和聲音是另一個問題。人的貢獻還有哪些層次（自己的草稿、修改、編排），完整整理在〈AI 作品，我要做多少才會受保護？〉。
該留什麼、什麼該自己收好
不管提示詞受不受保護，它都是「你做了什麼」的證據。幾個成本很低的習慣：
- 留一份有日期的提示詞紀錄。 提示詞、工具與模型版本、有的話記 seed、最後留了哪一張。日後有人問你的貢獻是什麼，誠實的答案就從這裡來。
- 把「你寫的文字」和「你的設定」分開存。 歌詞、劇本、長篇描寫放一個檔；風格標籤和參數放另一個。前者可能本身就是一件作品，後者是方法。
- 原始產出和修改後的版本都留。 如果你重畫、剪輯、重錄，兩個檔之間的差別就是你的貢獻，也是主張最站得住的部分。
- 想清楚要不要公開。 如果某段咒語是你的看家本領，一旦公開，著作權大概擋不住別人拿技巧去用。放在自己的檔案裡就好。
- 看你所用工具目前的條款。 輸入與產出的權利歸誰、能不能商用，依方案而定、也會改版，請直接看該工具官方的條款頁。
OpenStela 能幫上的那一步
先講做不到的：OpenStela 沒辦法單獨登記一段提示詞。上傳型作品只收音訊、圖片、影片檔（mp3、m4a、wav、jpg、png、webp、mp4、webm），文字檔不收。能登記的，是產出的檔案，以及你對它的說明。
在 app.openstela.io/works 登記作品時，勾選「這件作品有用 AI 工具生成」，「你的貢獻」就變成必填（上限 600 字元），也要填用了哪些工具。這些欄位會和檔案的 SHA-256 指紋一起凍進著作紀錄，當晚批次上鏈；之後誰都改不了，包括你自己。
要提醒一件事：公開的作品，貢獻欄也是公開的。不想讓咒語外流，就用動詞寫你做了什麼（「自己寫歌詞；寫了一段約三百字的場景提示詞；生成 25 版；副歌重新剪拍點」），提示詞本身留在自己的紀錄裡。那段文字我們不審核、也不評分，它是你自己的聲明。
你拿到的是一筆可查驗的時間紀錄：這個檔案在這一天、以這個名稱存在，附上當時聲明的人的貢獻。發生爭議時，這是多一個證據。它不決定你對提示詞或產出有沒有著作權——那由法院或著作權主管機關判斷。同一輪咒語生出一整組圖，可以看〈我用 Midjourney 做了一整個系列，怎麼一次替全部留下日期？〉；一筆紀錄證明什麼、不證明什麼，看〈角色編號證明什麼，不證明什麼〉。
作品登記免費；上傳型作品每個帳號共用 300 MB 空間，連結型幾乎不佔空間，每天新登記的件數有上限（見價格）。
本文為一般資訊，不是個案法律意見。真的遇到爭議，請找你所在地的律師。
Questions常見問題
Can I copyright a short Midjourney prompt?一句很短的 Midjourney 咒語，可以主張著作權嗎？
Unlikely. The Copyright Office's Circular 33 says words and short phrases lack enough authorship to register, even when they're novel. A list of style tags and parameters is also closer to a set of instructions, and under 17 U.S.C. § 102(b) methods and ideas aren't protected however they're written.
不太可能。美國著作權局 Circular 33 說，單字與短語創作量不足，即使新穎也不予登記。一串風格標籤和參數，也比較接近一組操作指示；依美國 § 102(b) 與台灣著作權法第 10 條之 1，想法與方法不論怎麼寫都不受保護。
If my prompt is protectable, do I own the image it made?如果我的提示詞受保護，生出來的圖就是我的嗎？
No. The Copyright Office's 2023 guidance says material generated from a copyrightable prompt isn't copyrightable for that reason, and its 2025 report says prompts alone don't give enough control to make you the author of the output. Your rights in the output come from what you contributed beyond the prompt, such as your own input that shows in the result, or your edits.
不是。美國著作權局 2023 年的指引說，由受保護提示詞生成的內容，不會因此就受保護；2025 年的報告也說，單靠提示詞的掌控不足以讓你成為產出的著作人。你對產出能主張的，來自提示詞以外的貢獻，例如在成品裡看得出來的你自己的輸入，或你的修改。
Someone used my prompt and got a similar image. Is that infringement?別人用了我的咒語，生出很像的圖，這樣算侵權嗎？
Depends what they took. If they copied your prompt text word for word and that text is protectable, the copying of the text is the question. Reaching a similar look with their own words reuses an idea or technique, which neither U.S. § 102(b) nor Taiwan's Article 10-1 protects. Whether a specific case is infringement is decided by a court.
要看他拿了什麼。如果他一字不漏照抄你的提示詞，而那段文字本身受保護，問題在於「照抄文字」這件事。如果他用自己的話做出類似的畫面，用的是想法或技巧，美國 § 102(b) 和台灣第 10 條之 1 都不保護這些。具體個案算不算侵權，由法院認定。
Can I register my prompt on OpenStela?可以把提示詞登記到 OpenStela 嗎？
Not as a work. Uploads are audio, image and video files only. You can register the output and describe your contribution in the record, but that text is public on a public work, so keep any prompt you want private in your own files.
不能當作一件作品登記，上傳只收音訊、圖片、影片檔。你可以登記產出、在紀錄裡說明自己的貢獻；但公開作品的貢獻欄是公開的，想保密的咒語請留在自己的檔案裡。