我担心陌生人会把我的小说称为 AI 作品,所以我为文字构建了一个 GitHub
I'm Scared a Stranger Will Call My Novel AI, So I Built GitHub for Words

原始链接: https://dylan.blog/2026/08/03/im-scared-a-stranger-will.html

在这个作家愈发频繁被指控使用人工智能的时代,证明人类创作已成为一种令人倍感压力的必要之举。由于文风“习惯”已不再是可靠的证据,作家兼开发者艾里斯·杨(Eris Young)开发了 **VellumProof**,旨在为写作过程提供“凭证”。 受 GitHub 版本控制和热图的启发,VellumProof 会追踪手稿的演变,记录逐字修改内容和时间戳,而非仅仅记录代码行。通过可视化每日字数统计,该工具为作者的创作进度创建了一份可核实、透明的历史记录。尽管作者承认没有任何系统能完美防止恶意使用者伪造数据,但该工具的目标是将证明负担从主观的“感觉”转移到客观的记录成果上。 归根结底,VellumProof 是创作者的“测谎仪”,让作家能够提供其作品的创作延时过程。它是对被无端指控使用人工智能这一恐惧的积极回应,并建立在这样一种信念之上:大多数作家都是诚实的,且理应拥有验证其创作成果的方式。

最近的一场 Hacker News 讨论探讨了在生成式 AI 时代证明人类作者身份的困难。该话题源于一位作者的博文,他建立了一个“文字版 GitHub”来记录自己的写作过程,以此作为防御被指控使用 AI 生成内容的手段。 评论者们对于这种验证机制的可行性和必要性存在严重分歧。一些人认为,随着受众对 AI 生成内容感到疲劳,证明作品来源对于建立消费者信任至关重要。另一些人则反驳称,任何“证明”系统(如提交历史)都可以被 AI 伪造,或者通过标准的编辑流程绕过。 许多参与者指出,在没有实时监控的情况下,外部验证从根本上是不可能的,并指出即使是历史文学作品也面临着作者身份的质疑。辩论中有很大一部分集中在一个讽刺现象上:一些试图证明自己是人类的作者,反而因为写作风格在潜移默化中采用了“类 AI”的模式,而被指控使用 AI。归根结底,这一共识突显了一种普遍的焦虑:随着 AI 变得越来越复杂,人类与机器创造力之间的界限正变得越来越模糊,使创作者和消费者都陷入了相互猜疑的状态。
相关文章

原文

A courtroom scene where a nervous novelist at the defendant’s stand holds up an enormous printed green contribution-graph as an alibi — but instead of GitHub squares, each tile is a tiny quill pen — while an easel displays a placard reading “EXHIBIT A: 6,000 WORDS, TUESDAY, BY HAND,” the prosecutor is a smug laptop in a powdered wig, the judge is a giant magnifying glass with a monocle, a single ghostly em-dash floats above the witness stand like a suspect being identified in a lineup, dramatic gothic-meets-90s-VGA lighting, one sweat-pixel rolling down the novelist’s brow

As I’ve mentioned before, we now live in a world where creative people have to prove they created the thing they created.

Sit with that for a second, because it’s genuinely deranged. I have been writing fiction since 2014 — unpublished, because getting published is the hard part and I am bad at the hard part — and the thing that keeps me up at night isn’t the rejection. It’s that some stranger is going to read my perfectly normal writing tics, feel that they’re AI, and torpedo the whole dream over a vibe.

The part that makes me want to chew glass is that none of this would be a problem if people were just honest about what they used to make a thing. But that’s apparently too much to ask, because cheaters gonna cheat.

So instead, the rest of us have to carry receipts.

First I had to meet the enemy

The first thing I did — because I had to — was see if I could teach an LLM to write fiction. Know your enemy, etc. This was a whole project on its own, because every AI has tics. And it’s not that the tics don’t exist in real writing (I’m looking at you, em-dash, my beloved). It’s that the machine reaches for them constantly.

Which is exactly the dilemma. Some human writers lean on the em-dash too. So how do you prove that you put the words on the page and not a robot with a thesaurus? You can’t do it from the words alone. The words are the crime scene, not the alibi.

A brief detour into nerd territory

Let me tell you about GitHub, in case you’re not a nerd (I’m a baby nerd, so I’ll keep it gentle).

GitHub is where coders keep their projects. It saves versions, it lets you work on new features without breaking the live thing, and — here’s the part I love — it draws you a little heat map showing how much code you committed on each day. That map makes it stupidly easy to tell a human from a machine. Someone who commits 50,000 lines in an afternoon did not type those by hand. Someone doing a couple hundred? Probably real. Simple.

The other magic trick is that it keeps all your changes. All of them. Which means writers' oldest nightmare — losing the work — basically stops being possible. If every version is saved, there’s nothing to lose.

I’ve been keeping all my projects on GitHub, writing included. But GitHub has one problem for people like me: it counts lines. And writers don’t think in lines. We think in words. We love words. Give us a word and we’ll turn it over for an hour.

So I just needed to take the beautiful change-tracking guts of GitHub and rebuild them for people who count in words.

Here’s how it works, and I promise it’s less scary than GitHub: at the end of a writing day, you upload your manuscript. VellumProof tucks that version away, and the next time you upload, it compares the two and records exactly what changed — words added, words cut — and stamps it with the date. It’s git commits, minus the part where you have to learn the command line to feel like a real boy.

You pick which files in a project get tracked, and you can see what’s changed between any two of them. Send your book to an editor? Upload their pass and see precisely what they touched and how much. That alone is worth it.

But my favorite part is that it’s proof.

There’s a graph that tracks how many words got added per upload. And that graph is a little bit of evidence of your own hand in the thing. The most I have ever written in a single day is around 6,000 words. It was a great day. It is also the absolute ceiling of my ability. So if my history suddenly shows a 10,000-word Tuesday, there’s a very good chance I did not write those 10,000 words — and anyone looking can see the same thing I can.

Now, a wrinkle: your first upload of an old project is going to look insane. If I dump in Calliope and the Rogue Biologic — 50-some-thousand words — on day one, the graph will scream that I wrote a whole novel in an afternoon. That’s what the notes are for. I tag that commit “Initial upload of CRB, drafted during NaNoWriMo 2014 (this is realistically the fifth draft),” and now anyone reading the history understands: day one was me importing a book, and everything after is the real, grinding, word-by-word edit trail.

The dream — the actual dream — is that when you submit something and somebody squints and says “this feels like AI,” you can hand them a timelapse of yourself writing it. Not a vibe. A record. Receipts.

The obvious objection

Yes. Cheaters are going to cheat. I have personally thought of five ways to game this thing before breakfast, and there is no version of it that blocks all of them. You could paste in AI text a little at a time to fake a human curve. You could do a lot of things.

But here’s where I’ve landed: I could spend forever building a fortress against the worst people, or I could build a good tool for the honest majority and trust that most people aren’t the villain of this story. I’m choosing the second one. I’d rather bet on the human side of things and be occasionally wrong than treat every writer like a suspect — which is, you’ll notice, the exact thing I’m mad about in the first place.

Anyway. I built a lie detector for my own novel because I’m scared, and building things is how I metabolize being scared. If that’s not the most writer thing I’ve ever done, I don’t know what is.

Keep your receipts, be honest about your tools, and go write something nobody asked for.

联系我们 contact @ memedata.com