在正确的提示下，大型语言模型可以产生高质量的写作——并让我们质疑人类创造力的极限。
尼古林·图克森 (Nicholine Tuxen) 绘制的《躺在床上读书的女子肖像》(图片来源：Wikimedia Commons)
关键要点
- 与普遍看法相反，生成式人工智能能够创作出高质量的散文。
- 虽然一些作家反对人工智能可以独自创作艺术的想法，但其他人却将其用作创意写作的工具。
- 人工智能很可能会改变文学，就像过去的创新改变其他艺术一样。
2024 年，奇幻小说作家乔治·RR·马丁 (George RR Martin)正式花了 12 年时间创作《凛冬的寒风》 ，这是 HBO 热播剧《权力的游戏》系列的第六部作品，这部作品备受期待。由于看不到发布日期，一位精通技术的粉丝决定自己写这个故事，或者更确切地说，他让 ChatGPT 为他写。在要求为每个章节制定大纲，然后将这些大纲变成散文的过程中，人工智能写出了一本 683,276 字的巨作，质量令人惊讶。
令人惊讶的是，尽管这个“粉丝制作”版的《凛冬的寒风》未能达到马丁作品的水准，但它确实包含了意想不到的曲折，使他的奇幻史诗如此成功。马丁是这本书的读者之一，他得知这本书的存在后，不仅对粉丝采取了法律行动，还对 ChatGPT 的开发商 OpenAI 采取了法律行动。
马丁加入了约翰·格里森和乔纳森·弗兰岑等其他畅销书作家的行列，他与大型语言模型的持续斗争——这些模型以他们的作品和无数其他来源为训练基础——提出了一个关于机器学习创造力的重要问题。最近几个月使用过 ChatGPT 等人工智能程序的人都知道，它们非常擅长总结教科书和撰写样板求职信，但它们是否也能创作出让我们感动并触动我们灵魂的写作呢？
训练数据
记者 Vauhini Vara 认为答案是肯定的。在一篇题为《病毒式 AI 作家的自白》的文章中，她讲述了 OpenAI 如何在 2020 年授予她 GPT-3 的早期使用权。她将创意写作主要定义为“等待合适的词语”，她认为生成式 AI 可以即时访问英语词典中的每个名词和形容词，这可能是一个有用的工具，尤其是对于她难以写出的主题。
其中一个主题——实际上是主题——是 Vara 姐姐的去世，她在高中时被诊断出患有一种罕见的癌症。最初，GPT-3 基于这个前提编写故事的尝试并不令人满意。AI 会将她姐姐在现实世界中的命运换成奇迹般的康复，或者在承认她死亡的草稿中，将 Vara 变成一名为慈善事业而赛跑的长跑运动员。直到她转向更具协作性的写作过程，GPT-3 从她自己的输入中学习并适应，它才成功创作出一行真正打动她的文字：
我们从克拉克海滩开车回家，在红灯前停了下来，她握住我的手。她握住的就是这只手：我用来写字的手，我用来写这篇文章的手。
如果人工智能可以与我们的灵魂对话，这说明了什么？
尽管 ChatGPT 被编程为以一种平淡的、解释性的方式书写，但只要它拥有足够的训练数据，它就能学会模仿其他声音。
“我认为人们有一种误解，认为像 ChatGPT 这样的大型语言模型不擅长用抒情、文学散文风格写作，”作者 Sean Michaels 在被引荐阅读 Vara 的文章后告诉 Big Think。“事实上，它们可以轻松而且相当出色地做到这一点，就像所有图像生成软件都可以像韦斯·安德森或大卫·林奇的风格那样制作照片一样。”
迈克尔斯的小说《你还记得出生吗？》讲述了一家科技公司找到一位诗人，邀请他与他们的人工智能共同创作一本诗集的故事。迈克尔斯为 ChatGPT 创建了一个定制版本，称之为“Moorebot”。通过训练，Moorebot 学会了以现实世界诗人玛丽安·摩尔的风格写作，进一步模糊了人类与机器之间的界限。
人工智能：创造力与联系
尽管一些作家出于对就业能力的担忧而反对人工智能的崛起，但另一些人将这场辩论置于一场关于艺术意义和目的的更古老、更广泛的对话中。奥斯卡获奖编剧查理·考夫曼的写作风格非常独特，评论家们创造了“考夫曼风格”这个词来形容他的写作风格。考夫曼可能最直言不讳地表达了他个人对机器学习的厌恶。考夫曼将创造力的标准定义（创造新颖和原创的东西）扩展到了艺术，即任何一种在人与人之间建立情感联系的表达形式。
“说出你是谁，”他在英国电影学院奖的演讲中说道，这预示了2023 年美国演员工会罢工期间的言论。“在你的生活中和工作中，真心实意地说出来。告诉别人——那些迷失的人、那些尚未出生的人、那些 500 年后才会出生的人。你的写作将记录你的时间。这无可避免。但如果你诚实，你会帮助那个人在他们的世界中不那么孤独。”
即使人工智能能够写出一些感人或深刻的东西（瓦拉和迈克尔斯声称它可以），考夫曼也不会承认这是艺术，因为它无法在读者和作者之间建立联系。“如果我读了一首诗，”他最近在一份关于人工智能的声明中说道，“这首诗让我感动，我就会爱上写这首诗的人。但我不可能爱上一个计算机程序。我不能，因为它什么都不是。”
好莱坞编剧担心人工智能会取代他们。（图片来源：David James Henry / Wikipedia）
考夫曼的观点应该与迈克尔斯的观点相提并论，迈克尔斯通过对人工智能的研究得出了不同的结论：人类的创造力并不像我们想象的那么特殊。
他解释道：“想象一下，我写道：‘月亮看起来像一片柠檬。’你读了之后会想：‘这是一个有趣的图像。’但是当人工智能这样做时，你会想：‘嗯，这很有趣吗？因为它看起来完全是随机的。’人类有能力从事物中找出意义，并在没有联系或没有意图的地方找到联系。我发现这很有启发性，不仅因为人工智能可以想出我认为美丽、奇怪或有趣的句子，而且它让我怀疑人工智能是否成功地成为了一名作家，或者我是否成功地成为了一名读者。”
至于人工智能是否具有“真正的”创造力：“人类艺术的确借鉴了记忆、身体、梦想以及抽象、超现实思想之间任意但有意义的联系。这些东西都不是 ChatGPT 所拥有的。至少不是字面上的。但话说回来，小说中关于假装的身体、假装的梦想、假装的生活经历的方式有些奇怪。如果你读到一篇关于 17 世纪荷兰商人的现代故事，它并不是基于任何人作为 17 世纪商人的真实生活经历。同样，因为这些大型语言模型没有我们所拥有的东西，这并不意味着它们的工作没有我们的任何品质。这不是那么容易的。”
写作的未来
迈克尔斯表示，大型语言模型的使用增加有望“导致风格在一定程度上趋同，由于我们使用相同的算法，写作将变得缺乏多样性，变得更加单一”。
巴黎西岱大学博士生弗洛朗·文雄 (Florent Vinchon) 研究生成式人工智能的创造性技能，他认为，尽管现在下结论还为时过早，但机器学习“对写作的影响可能与摄影对绘画的影响相同”，该领域的其他专家也持相同观点。正如相机将画家从现实主义推向抽象主义，催生出关注我们如何感知而不是感知什么的艺术运动一样，人工智能可以激励新一代作家尝试他们的写作。
绘画曾经是这个样子的。（来源：Art UK / Wikipedia）
科技也会让写作变得更加抽象吗？（图片来源：Hennie Niemann / Wikimedia Commons）
由于 ChatGPT 之类的程序允许任何人创作出合格的散文，未来的小说和短篇小说可能会故意不那么精致，充斥着语法错误和语言怪癖，这些都表明了人类大脑的局限性和个人观点的独特性。人工智能不仅不会对创意写作或整个人类创造力的生存构成威胁，反而可能会像之前的创新技术一样成为变革的力量和灵感的源泉。
ChatGPT is great at summarizing books. But will AI ever write a true work of literature?
With the right prompts, large language models can produce quality writing — and make us question the limits of human creativity.
上传失败，网络异常。
重试
上传失败，网络异常。
重试
Portrait of a Woman Reading in Bed by Nicholine Tuxen (Credit: Wikimedia Commons)
KEY TAKEAWAYS
Contrary to popular belief, generative AI is capable of producing quality prose.
While some writers resist the idea that AI can produce art alone, others use it as a tool for creative writing.
It’s probable that AI will change literature the same way past innovations changed other arts.
Tim Brinkhof
Copy a link to the article entitled http://ChatGPT%20is%20great%20at%20summarizing%20books.%20But%20will%20AI%20ever%20write%20a%20true%20work%20of%20literature?
Share ChatGPT is great at summarizing books. But will AI ever write a true work of literature? on Facebook
Share ChatGPT is great at summarizing books. But will AI ever write a true work of literature? on Twitter (X)
Share ChatGPT is great at summarizing books. But will AI ever write a true work of literature? on LinkedIn
Sign up for the Smarter Faster newsletter
A weekly newsletter featuring the biggest ideas from the smartest people
Fields marked with an * are required
It’s 2024 and fantasy author George R.R. Martinhas officially spent 12 years working on The Winds of Winter, the long-awaited sixth installment in the series that inspired the HBO hit Game of Thrones. With no release date in sight, one tech-savvy fan decided to write the story himself, or rather, he asked ChatGPT to write it for him. Prompted to develop outlines for each chapter, and then turn those outlines into prose, the AI churned out a 683,276-word tome of surprising quality.
FEATURED VIDEOS
3 powerful mind states: Flow state, good anxiety, and Zen Buddhism
00:03
25:36
Surprising because although this “fan-made” version of Winds of Winter failed to live up to the standards of Martin’s work, it did contain the unexpected twists and turns that made his fantasy epic so successful. Among the book’s flabbergasted readers was Martin, who upon learning of its existence not only took legal action against the fan but also ChatGPT’s developer, OpenAI.
Joining the ranks of other bestselling authors such as John Grisham and Jonathan Franzen, Martin’s ongoing war with large language models — which were trained on their writing, among countless other sources — raises an important question about the creative capabilities of machine learning. Anyone who has used AI programs like ChatGPT in recent months knows they are great at summarizing textbooks and writing boilerplate cover letters, but can they also produce the kind of writing that moves us and speaks to our souls?
Training data
Journalist Vauhini Vara thinks the answer is yes. In an article titled “Confessions of a Viral AI Writer,” she relates how in 2020 OpenAI granted her early access to GPT-3. Defining creative writing mainly as “waiting for the right word,” she suspected that generative AI, with its instantaneous access to every noun and adjective in the English dictionary, could be a helpful tool, especially for topics she struggled to write about.
One of these topics — the topic, really — was the death of Vara’s older sister, who was diagnosed with a rare form of cancer while still in high school. Initially, GPT-3’s attempts at writing a story based on this premise were unsatisfactory. The AI would swap her sister’s real-world fate with a miraculous recovery or, in drafts where her death was acknowledged, turn Vara into a long-distance runner racing for charity. It wasn’t until she shifted to a more collaborative writing process, with GPT-3 learning from and adapting to her own inputs, that it managed to produce a line of text that genuinely touched her:
We were driving home from Clarke Beach, and we were stopped at a red light, and she took my hand and held it. This is the hand she held: the hand I write with, the hand I am writing this with.
上传失败，网络异常。
重试
If AI can speak to our soul, what does that say about us? (Credit: Corey Coyle / Wikipedia)
Although ChatGPT has been programmed to write in a bland, explanatory way, it can learn to mimic other voices provided it has enough training data.
TOP STORIES
重试
00:01
02:00
Read More
“I think there is a misconception that large language models like ChatGPT are not very good at writing in a lyrical, literary prose style,” author Sean Michaels tells Big Think after being pointed to Vara’s article. “In fact, they can do it easily and quite well, just like all the image-generating software can do things like making photos in the styles of Wes Anderson or David Lynch.”
For his novel Do You Remember Being Born, in which a poet is approached by a tech company to co-author a collection of poems with their AI, Michaels created a custom version of ChatGPT he called “Moorebot,” which through training learned to write in the style of real-world poet Marianne Moore, further blurring the distinction between human and machine.
AI: creativity and connection
Although some writers oppose the rise of AI out of concern for their employability, others have embedded the debate in an older, broader dialogue on the meaning and purpose of art. The Oscar-winning screenwriter Charlie Kaufman, who writes in a style so idiosyncratic critics coined the phrase “Kaufmanesque” to describe it, has been perhaps most vocal about his personal distaste of machine learning. Looking beyond the standard definition of creativity — creating something new and original — Kaufman describes art as any form of expression that establishes an emotional connection between people.
“Say who you are,” he said in a BAFTA lecture that foreshadowed comments made during the 2023 SAG strikes. “Really say it, in your life and in your work. Tell someone out there — someone who is lost, someone not yet born, someone who won’t be born for 500 years. Your writing will be a record of your time. It can’t help but be. But if you’re honest, you will help that person be less lonely in their world.”
Even if AI were capable of writing something emotional or profound, which Vara and Michaels claim it can, Kaufman would refuse to recognize this as art because it cannot establish a link between reader and author. “If I read a poem,” he recently said in a statement on AI, “and that poem moves me, I am in love with the person who wrote it. I can’t be in love with a computer program. I can’t, because it isn’t anything.”
上传失败，网络异常。
重试
Hollywood screenwriters fear AI will replace them. (Credit: David James Henry / Wikipedia)
Kaufman’s sentiment should be assessed alongside that of Michaels who, through his work with AI, has reached a different conclusion: Human creativity is less special than we want to believe. He explains:
“Imagine I write, ‘The moon looked like a slice of lemon.’ You read it and think, ‘That’s an interesting image.’ But when AI does it, you think, ‘Well, is that interesting? Because it just seems completely random.’ Human beings have this capacity to make meaning from things and find connections where there are none or where there is no intention behind them. I found it provocative not just that AI could come up with lines I found beautiful or strange or interesting, but also that it got me wondering if the AI was doing something successfully as a writer or if I was doing something successfully as a reader.”
As for whether or not AI is capable of ‘true’ creativity:
“It’s true human art draws on memories, bodies, dreams, and arbitrary but somehow meaningful linkages between abstract, surreal thoughts. None of those things are things ChatGPT has. At least not in a literal way. But then again, there is something strange about the way fiction is about pretended bodies, pretended dreams, pretended lived experiences. If you read a modern story about, say, a 17th-century Dutch merchant, that’s not based on anybody’s real-life experience as a 17th-century merchant. Similarly, because these large language models don’t have what we have, that doesn’t mean that their work has none of the qualities that ours does. It’s not that easy.”
The future of writing
According to Michaels, increased usage of large language models is expected to lead “to a certain convergence of styles, with writing becoming less diverse and more monolithic as we use the same algorithms.”
Florent Vinchon, a PhD student at Université Paris Cité studying the creative skills of generative AI, suggests that — although it is still too soon to say with certainty — machine learning “might have the same impact on writing as photography had on painting,” a view shared by other experts in the field. Just as the camera pushed painters from realism to abstraction, spawning artistic movements that concerned themselves with how we perceive rather than what, AI could inspire new generations of authors to experiment with their writing.
Painting used to look like this. (Credit: Art UK / Wikipedia )
Will technology make writing more abstract, too? (Credit: Hennie Niemann / Wikimedia Commons)
Since programs like ChatGPT allow anyone to produce competent prose, the novels and novellas of tomorrow may be deliberately less polished, riddled with grammatical errors and linguistic idiosyncrasies that point to the limits of the human brain and the uniqueness of the individual perspective. Far from a threat to the survival of creative writing or human creativity in general, AI, like the innovative technologies that came before it, may become a force of change and a source of inspiration.
In this article
artificial intelligence
books
technology
