From Resume Project to Ethical Dilemma: My Journey with AI

Back in 2022, I needed a resume project. I was in a tough work environment with a baby on the way and needed a change. As a software engineer, I wanted to build a project that would let me put the AI box on my resume. I built an NPC generator that I’ve since taken down due to ethical concerns, especially now that HalfOrc.Club is trying to become something real.

The tool would create some roll tables for defining features, roll on them, and generate a simple description:

$"{Name} the {Adjective.ToLower()} {Race.ToLower()} {Profession.ToLower()} has {FirstTrait.ToLower()} and {SecondTrait.ToLower()}";

In September 2022, I first created the NPC generator using DALL-E 2. I gave it a simple prompt, and it returned the images below after about a minute.

Fast forward to July 2023, and the new “Stable Diffusion” model marked another leap forward in capabilities. I provided the same prompts as before and got better images faster.

Then DALL-E 3 was released in September 2023 and blew everyone away. The images were good and only took a few seconds to generate.

Generate a picture of Zook the gnome druid riding a giant eagle

Let’s pause to reflect for a second on what just happened in only a year. Things seem to finally be slowing down, and I personally believe we’re near the peak of the AI craze. New open-source models like Grok 2 from X, based on Flux, generate higher quality images but have a narrow understanding of the prompt.

And here is the current state of Stability AI.

Generate a picture of Zook the gnome druid riding a giant eagle

How Does It Work?

From AWS - What is Stable Diffusion?

As a diffusion model, Stable Diffusion differs from many other image generation models. In principle, diffusion models use Gaussian noise to encode an image. Then, they use a noise predictor together with a reverse diffusion process to recreate the image.

Stable Diffusion models analyze images from the internet, calculating steps to reduce them to random noise. When a user requests a similar image, these steps are reversed from a new starting point of noise, resulting in a comparable image.

The Mona Lisa from Wikipedia

Random noise (static)

DALL-E 3 “Show me the Mona Lisa if she was a half orc”

Large Language Models (LLMs) operate on a similar principle of pattern recognition and generation. They are trained on vast amounts of text data, learning the statistical relationships between words and phrases. When given a prompt, an LLM uses these learned patterns to predict the most likely next words or sentences, generating coherent text that matches the style and context of the input. This process allows LLMs to produce human-like text across a wide range of topics and writing styles.

TL;DR:

These systems work because of billions of unattributed references baked into the model. The training data was sourced from the vast collection of publicly available data on the internet, including websites, books, articles, and social media posts. This broad dataset allows AI models to generate diverse and contextually relevant content. However, it also raises questions about the rights of original creators whose work may have been used without explicit permission in the training process.

The Ethical Dilemma: Where Do We Go from Here?

The ethical concerns over AI center on questions about training data, consent, copyright, and fair use. As AI-generated content becomes more prevalent in creative fields like tabletop role-playing games, we must respect intellectual property and artistic integrity as the technology evolves. We need an open dialogue about what's acceptable in our community. I'm going to open a conversation from my perspective. My goal in doing this is for people with other perspectives to poke back at things I can’t see from my perspective. I hope to learn from you all.

Draft of an Ethical Framework for Using AI in the TTRPG Community

This applies to those who profit from AI-generated content—the people selling the games and content we buy and play. For home games with friends, you can do what you want.

1. Attribution and transparency: When using AI-generated content, it's crucial to disclose its origin and not pass it off as human-created work. This maintains trust within the community and respects human creators' contributions.
2. Copyright and fair use: Ensure that the output doesn't infringe on others' work. This includes avoiding attempts to mimic or replicate another artist's style without permission. While AI models may draw inspiration from various sources, we should respect individual artists' unique creative expressions and seek to develop original content.
3. Artistic integrity and originality: As AI becomes more capable of generating high-quality content, we must balance its use with the value of human creativity and originality in TTRPGs. This includes considering how AI-generated content might impact the uniqueness and personal touch that many players and creators value in the hobby.
4. Collaborative approach: Use an approach where AI is used as a tool to enhance human creativity rather than replace it. This could involve using AI for brainstorming, generating initial ideas, or assisting with routine tasks, while leaving the core creative decisions and final touches to human creators.

How Do We Use AI at HalfOrc.Club?

First, I'm going to close the NPC generator. It was a fun project that I might bring back if I start making money from other ventures and can afford to offer it for free. I used it to showcase my tech ninja skills as a portfolio piece, landed a job, but never directly profited from it.

Second, I'm developing a new tool I plan to monetize. It's essentially a wiki with flexible block-based editing. You'll be able to manage privacy settings for your wiki pages and opt in to syncing them with AI, making them searchable by your campaign's lorekeeper bot. You can also connect your Twitch, YouTube, or podcast to summarize your TTRPG sessions and add session notes to the wiki, clearly labeled as AI-generated and linked to the source. Humans will create the core content and have to opt in. This means players joining a Twitch stream at episode 137 can ask questions like "Who is Lord Rogerfel and why was he exiled?" and receive summaries from and links to the campaign wiki. To be specific, I’m using AI for the following:

• Speech to text
• Text summary
• Search

Looking back, what started as a simple resume project has led me down a path of deep reflection on the role of AI in our creative spaces. As I move forward with new projects like HalfOrc.Club, these ethical considerations will be at the forefront of my work.