Can an AI write a credible novel with Wikipedia and YouTube?

I always wanted to create a toolbox of critical analysis tools for young people to analyze, debate, and anticipate the near term future given our human nature and our history, especially since the scientific revolution. This is what my novel is intended to do, but writing a first (crap) edition, second (okay-ish) edition, and now third (pretty good?) edition has led to corollary insights and questions dealing with strong artificial intelligence (AI).

Do we use large sets of algorithms like Watson does? Perhaps. But thousands of algorithms smack of inefficiencies. Is a dog solving a differential equation when it catches a ball? I don't think so. Yet it might be argued that a dog is solving a differential equation in the sense of isomorphism. I think a guide towards a strong AI has to included a concept of algorithmic efficiency.

If we assume that there is a sufficient representation of humanity online, like our literature, our news stories, our movies, our poems, our wars, our science and mathematics, and so on, can we build a strong artificial intelligence (AI) based on humanity's online representation capable of experiencing love, hate, fear, etc., so that it can write believable novels?

If so, how would we know that we are succeeding? Applying the Turing test is one way, but think about well written novels like War and Peace, Moby-Dick, or some other more contemporary novel like Contact. These novels capture microcosms of humanity and its complexities. They are set within historical contexts as humans understand history. They include simple minded people, extremely complex people, and even insane people. They include good, bad, sex, love, hatred, and all the other singular words with which we describe ourselves. They are used in literature classes to get students to learn about human nature. Would it not then be a test for strong AI for us to be unable to detect a novel written by a strong AI?

The first question concerns increasingly higher order strong AI meta algorithmics perquisites such as:
* reading our human writings via optical character recognition,
* understanding our spoken language,
* performing static pattern recognition -- That's Joe Bob.
(realm of statistical methods, e.g., principal component analysis, filtering, compression of data,...)
* performing spatio temporal recognition -- That's Claire on the beach having sex with Joe Bob.
* performing expanding contextualization recognition. Claire and Joe Bob are making love on the beach because (1) humans are biochemically driven to reproduce, (2) Claire and Joe Bob love each other, (3) Joe Bob has just returned from a war he didn't think he would survive because country X invaded his country to force their religion on the world starting with his country because many people cannot think for themselves and can be driven by religion to do evil, especially when national leaders start wars for the purpose of capturing resources and/or for a need to conquer to become immortals in history.
* ...
This seems (to me) to be a realm suited for the engineer (and the tinkerer). The engineer will use tools to solve some of the problem space of strong AI.

The second question concerns relational modeling.
* The Standard Model of physics is limited to particle physics.
* Organic chemists use rules of thumb.
* There are different approaches to taxation in capitalism.
* There are linkages between any two Wikipedia articles, e.g., from the Standard Model to pornography via quantum physics, transistor, computer, internet, webcam,... Mapping the relational networks would be key. Low order methods might begin with Shannon information theory, thermodynamics, statistical physics, cellular automata, agent based methods, dynamic networks,...
* ...

Perhaps the physicist is best prepared to tackle this question, e.g.,
Medicine = Linear Taylor series expansion = (0th order) caveman + (1st order) sanitation + (2nd order) antibiotics + (3rd order) rational drug design + (4th order) gene therapy + ...
Economics = Nonlinear coupled stochastic differential equations model.

The mathematician would be of use to both the engineer and the physicist to order and bound different sequences of AI approaches in terms of convergence in some set of metric spaces. Clearly there are other mathematical approaches (algebraic, topological,...).

My first serious attempt at novel writing (2005) would have failed the strong AI test. It had too many logical holes, and was too forced. (Its seed was based on a dream in 2003 or 2004.)
Joe Bob really liked to beat his wife. Joe Bob's wife wondered if she would get beaten again when he came home.

My second attempt at rewriting the same story (2012) might have some chance of passing the strong AI test. It had no serious logical holes and was better written, but still too forced by the use of too much narrative in the first 50 pages.
Job Bob's sister always wondered why Claire married him. "Will he beat her again when he comes home," she wondered nervously.

I'd like to think that my third pass of the same story (2014-2015) would pass the strong AI novel test. I'm going for contemporary "page-turner" readability.
"Why'd you marry that bastard?" Joe Bob's sister asked Claire while she loaded her shotgun. "Do ya like gettin' your face kicked in? You still can't see out that busted eye can ya?"
"You leave Job Bob alone!" Claire screamed. "Society made him do it."

I'm thinking of myself as some kind of positively trending strong AI from 39 years old to 49 years old whose brain network is always growing and being sculpted. Questions, opinions, suggestions, omissions, or requests to go to hell are equally welcome.
"Why'd you marry that bastard?" Joe Bob's sister asked Claire while she loaded her shotgun.
Who was loading the shotgun, Joe Bob's sister, or Claire?
Would it not then be a test for strong AI for us to be unable to detect a novel written by a strong AI?
Here are my thoughts:

A well-written novel (or short story) produced by a machine would surely be impressive. And I suppose it could work as a test of some kind. Yet, when compared to the Turing test, there are a couple of important differences. A written text is static and it is a one-way communication. A conversation (in a Turing test) is dynamic and it is a two-way communication. For these reasons, I consider the Turing test to be a stronger and better test, since the machine requires to respond to e.g. real-time dynamic context as well as new information during a conversation with a human.

A machine that would pass an AI test on a produced novel/story could be considered a good writer.
A machine that would pass the Turing test could be considered a good two-way communicator.
And we humans are not merely capable of writing, we are also capable of two-way communication.

(in case you haven't seen it already, there was a thread on a similar topic a while ago: "Asking bots for help")
Last edited:


This is too speculative and we no longer host blogs,.

Related Threads for: Can an AI write a credible novel with Wikipedia and YouTube?

  • Posted
  • Posted
2 3
  • Posted
  • Posted

Physics Forums Values

We Value Quality
• Topics based on mainstream science
• Proper English grammar and spelling
We Value Civility
• Positive and compassionate attitudes
• Patience while debating
We Value Productivity
• Disciplined to remain on-topic
• Recognition of own weaknesses
• Solo and co-op problem solving

Hot Threads