Abstract

This paper recounts the author's personal and professional journey from skepticism to appreciation of artificial intelligence (AI), particularly in the context of structured content and XML. Drawing parallels to the satirical film Dr. Strangelove, the author reflects on initial fears of AI and the moment of realization that, like the "Doomsday Machine," AI's impact is profound, misunderstood, and even humorous. The paper explores practical applications of AI—including prompt engineering, personas, custom instructions, and multimodal interactions—demonstrating how these tools enhance content creation, summarization, tagging, and validation. Legal, cultural, and technical concerns are also addressed, emphasizing responsible and effective use of AI in professional publishing environments.[1]

Introduction

For years we have heard about the potential of AI. I was skeptical. Then, a couple years ago it seemed like AI was about to go big. I saw some talks at Balisage that discussed AI in an XML world, and I was intrigued. I dove in. I had some “aha!” moments. This talk will describe what I learned since then and how it can be applied to challenges in a structured content environment.

The Doomsday Machine

I had another significant “aha” moment when I was young. When I grew up in the 1960’s we always worried about the Cold War and nuclear weapons. Nukes were powerful, dangerous, and worrisome, but kind of abstract. But when I saw Dr. Strangelove for the first time, I understood nukes differently. I discovered a new coping mechanism, dark humor. Nukes are funny, if you do it right. If you are not familiar with the film, I asked ChatGPT to provide a brief description.

Dr. Strangelove or: How I Learned to Stop Worrying and Love the Bomb is a 1964 satirical black comedy directed by Stanley Kubrick.[2] The film follows a series of absurd events that lead the U.S. and Soviet Union toward nuclear war after a rogue American general launches a nuclear attack. At the center of the story is the Doomsday Machine, a Soviet device designed to automatically destroy all life on Earth if the USSR is attacked, meant as a deterrent, but ironically kept secret until it's too late. The film lampoons Cold War paranoia and nuclear strategy, blending dark humor with chilling realism.

The threat of nuclear war and the Doomsday Machine were both absurd and terrifying. That is when I became more aware. And more worried. About the bomb anyway.

The New Big Scary Thing

A similar catharsis of sorts happened to me, but with big scary AI. I was a skeptic, that is until I got laid off and started doing job interviews for XML roles but had to explain my experience with AI. I wasn’t aware of the connection between the 2 technologies. Fortunately for me, that was just before Balisage last year. I listened to Steve DeRose[3] and others describe their early forays into AI. It was time to dig deeper. It was time to stop worrying and learn to love AI. And AI can also be funny.

About AI

The Art of the Prompt

An AI prompt is the input or instruction you give to an AI to generate a response. It can be a question, a command, or a detailed description—written in natural language or structured formats. To create an effective prompt:

  • Be clear and specific about what you want.

  • Provide context when needed (background info, examples, goals).

  • Use a tone or style appropriate to the task (e.g., casual, professional, concise).

  • Break complex requests into steps or bullet points if helpful.

For instance, the prompt:

Explain how solar panels work.

May produce results like this:

How Solar Panels Work[4]

Solar panels convert sunlight into electricity using a technology called the photovoltaic effect. Solar panels are made of many solar cells, usually made from silicon, a material that can conduct electricity. When sunlight (made of tiny particles called photons) hits the cells, it knocks electrons loose inside silicon. These freed electrons start to move, and that flow of electrons is what we call electricity. The electricity created is in direct current (DC) form. An inverter changes it into alternating current (AC), which is what homes and appliances use. This electricity can power your home, be stored in batteries, or sent back to the power grid.

The more information included in the prompt, the more detailed and specific the results can be.

Personas

A persona is a role or identity you assign to the AI to make its responses more relevant, styled, or useful for a specific audience. This can affect tone, vocabulary, level of detail, and even point of view. Using a persona helps you:

  • Get answers tailored to a specific audience (e.g., kids, executives, beginners)

  • Match a tone or voice (funny, serious, formal, sarcastic)

  • Simulate a professional role (doctor, teacher, marketer, coder)

  • Guide style for creative writing (e.g., noir detective, medieval bard)

Now consider this slightly longer version of the previous prompt that has additional information about the age of the audience:

Explain how solar panels work in simple terms for a 10-year-old.

Might return:

How Do Solar Panels Work?[5]

Imagine the sun is a big flashlight shining down on Earth. When that sunlight hits a solar panel (which is like a special magic sheet), the panel turns the sunlight into electricity—kind of like turning sunshine into power to charge your tablet or play video games. Inside the solar panel are tiny parts called solar cells. When sunlight hits these cells, they get excited and start moving—and that movement creates electric energy. That energy then travels through wires and goes to your home to power lights, TVs, or even a fridge. And guess what? It does all this without making smoke or pollution—so it's good for the Earth!

The voice, reading level, and casual style are more suitable for a ten-year-old than the response from the previous prompt.

Custom Instructions

Custom instructions let you personalize how the AI responds to you and how it understands your preferences. You can set these once and they’ll apply to all future prompts, unless you turn them off. They can be used to set the AI’s tone or personality, tell the AI what to prioritize in responses, and provide background about who you are or what you're working on. These features are especially helpful if you're using AI regularly for business, learning, or creative projects.

You start by including some custom instructions in the prompt. For example:

I am a technology blogger writing for a trendy tech audience.

This persona can be saved and reused avoiding the need to reenter it many times. Then it will be included in all of your prompts until you turn it off. For example, a second prompt might be:

Write a short paragraph about Markdown. Include humor and target the Gen Z audience.

The response may be something like this:

Markdown is basically the chill version of HTML—it lets you make your text look all snazzy without needing to know coding, like using emojis instead of learning 1000 random symbols. Want to make something bold? Just slap some asterisks on it. Want to make a list? Just type a hyphen and boom, bullet points. It's like the lazy genius of formatting, letting you look like a pro while keeping things low-key.[6]

Multi-Modality

Multi-modality refers to an AI’s ability to understand and work with multiple types (or “modalities”) of data—not just text, but also images, audio, video, and even structured inputs like charts or code. Common Modalities in AI include:

Table I

Modality Examples

Modality Examples
Text Messages, documents, code, web pages
Images Photos, drawings, screenshots, diagrams
Audio Voice commands, music, podcasts
Video Surveillance clips, social media, tutorials
Structured data Tables, spreadsheets, sensor logs

When designing a multi-modal prompt, you’re giving the AI more than just text. You’re combining modalities—for example, uploading an image and asking a question about it in text.

Multi-modal prompts require that you start with a primary instruction in text such as “Can you identify the insect in this picture and describe its habitat?” You must also upload a photo or another non-text object, then ask questions about that object. You can also optionally add conditions or constraints, such as "Focus on the insect on the left.”

Usually, multi-modal requests are only supported in a paid version of the AI engine. Multi-modality is currently available in GPT-4 Turbo, part of the ChatGPT Plus plan at $20/month (as of April 2025).

The Socratic Method

AI engines can use the Socratic Method to clarify your prompt and collect additional information by asking questions. For example, you could enter the prompt:

Please tell me how the Socratic method can be used in an AI prompt to ask questions to clarify a request. Let’s say I asked you to recommend a destination for a vacation that is within my budget and is to a place that matches my interests.

The results would look something like this:

Great question! The Socratic method in an AI prompt context means asking thoughtful, open-ended questions to clarify, explore, and narrow down a user's request rather than jumping straight to an answer.

When you say:

"Recommend a destination for a vacation that is within my budget and is to a place that matches my interests"

An AI using the Socratic method wouldn't immediately throw out a list of destinations. Instead, it would ask a series of targeted, guiding questions to better understand your context and preferences.

Here's how the Socratic method could be applied step by step:

Happy to help! To narrow it down, can I ask a few quick things first? 1. What's your budget (roughly)? 2. Do you prefer cities, nature, beaches, or something else? 3. Domestic or international? 4. How long is the trip, and when are you planning to go?[7]

Using the Socratic Method allows you to start with an incomplete or vague prompt and have the AI ask enough questions to clarify and refine the results.

Stupid AI & XML Tricks

Below are several examples of “exercises” related to the kind of tasks that often occur in an XML content environment. You can try these in your favorite AI tool.

Content Generation

First, let me say that content generation is rife with copyright issues, especially when you use AI in a “generative” way to create text, or even images and video. Read up on this and consider generating content as not producing material you can claim copyright to, at least until the legal issues have been clarified. The Large Language Model learns from using all the information out there on the internet whether it is public domain or not. AI may produce content that is potentially copyright infringement. All the examples discussed earlier show various ways to generate text content. Thanks to fair use rules, I can include them in this article. Generating content can be fun, and useful to explain things. It like a Google search on steroids. Try this example:

Please write a short paper about Fibonacci, the Fibonacci sequence and the golden ratio. Provide a brief bio of the man. Describe the Sequence and its relation to the ratio. Provide several examples found in nature.

Abstract Summarizing

Giving a document to the AI engine as a PDF or text file, or even in an XML format allows you to add the prompt to “create a short, abstracted summary of this document.” Experiment by adding details or personas to make the results targeted for a specific audience.

The Abstract for this paper was completely generated by ChatGPT and has not been edited. After copying and pasting the content of the paper I used this prompt:

Create an abstract of this paper written in MS Word.

Then I cut and pasted a MS Word version of this paper into the prompt. The results were impressive. Including this gem:

If you include or refer to a still from the movie, you must note its copyright: "Still image from Dr. Strangelove, © Columbia Pictures Industries, Inc., 1964. Used here under fair use for commentary."

The paper just happened to be in Word document format, but XML documents work too.

Citations, Dates, References, and Style Validation

Content often has specific styles for how some information is represented. These rules provide consistency in elements such as citations, cross references, Dates, URLs and email addresses, and many others. Since AI has learned from everything, it knows about things like Chicago Manual of Style, Blue Book citations, etc. You can ask AI to check that all of your controlled vocabulary and style rules are observed in the content. Try this example:

Are all the dates in this document expressed in the same date format? Are all the dates valid date values?

Then give it a document to evaluate.

Auto Tagging

AI knows a surprising amount about XML markup, standards, and terminology. You can ask it to take a plain text file and tag it using a standard schema like DocBook or Dita and get a pretty good result. Of course, you need to validate it and check that no content was lost or tagged incorrectly. One area people find this approach useful for is tagging complex content like tables or large repetitive files. Try this example. You copy some plain text table content into the prompt and ask it to:

Convert this tabular data to the XHTML Table element and attribute markup. Make sure there is a colspec for each column. Make sure there are the correct number of <td> elements and they correspond with the definitions in the colspecs. 

Modality          Examples

Text              Messages, documents, code, web pages
Images            Photos, drawings, screenshots, diagrams
Audio             Voice commands, music, podcasts
Video             Surveillance clips, social media, tutorials
Structured data   Tables, spreadsheets, sensor logs

Final Thoughts

It is important to understand the limits of AI before you apply it to important things like XML documents at work. It is important to be aware of the risks and, especially when working with document content and XML markup. Following is a breakdown of the key issues across legal, cultural, and technical dimensions. I am not a lawyer. I don’t feel qualified to advise on an emerging legal issue. I am going to let ChaptGPT explain some important issues. The following is mostly output from ChaptGPT when I asked: “Explain the legal, cultural, and technical challenges to using AI.”

Key Issues in Using AI for Document Content and XML Markup

Intellectual Property & Copyright: AI-generated content may raise concerns about ownership and copyright infringement.[8]

Data Privacy & Security: Compliance with regulations like GDPR and HIPAA is crucial when handling sensitive information.[9]

AI Transparency & Accountability: Some industries require documentation on AI-generated content to ensure accuracy and prevent misinformation.[10]

Cultural Issues

Bias & Fairness: AI models can reflect biases present in their training data, affecting neutrality and inclusivity.[11]

Human-AI Collaboration: AI should complement human oversight rather than replace it, ensuring alignment with organizational values.[12]

Adoption & Trust: Some employees may resist AI integration, requiring training and clear guidelines.[13]

Technical Issues

Quality of AI Output: AI-generated XML markup may contain errors, requiring validation and testing.[14]

Integration with Existing Systems: AI tools must work seamlessly with document management systems and XML processing pipelines.[15]

Version Control & Auditability: AI-generated content should be trackable and versioned to maintain accountability.[16]

Appendix A. Writing a Paper with AI about Writing a Paper with AI

This appendix describes the process and decision involved in writing this paper. It details the practical methods and challenges involved, from initial prompt-learning and iterative drafting in Word, to XHTML table conversion and attempts at DocBook XML tagging. Emphasizing AI as a contextual assistant rather than an autonomous writer, it explores ethical concerns like copyright, highlights the tool’s strengths in formatting and summarization, and calls for clearer publishing guidelines around AI-generated content. Through this experience, the author demonstrates that while AI can extend understanding and enhance markup, it must be used with caution and human oversight.

Writing Process

I began by taking several on-line courses on writing prompts for AI.[17] I focused on prompt writing and tasks that might be useful in a publishing environment (URLs not available). I did not take courses that could show me how to make funny pictures or deep fake videos. That training helped me form an outline of ideas to write about.

To write the paper, I first drafted it in Microsoft Word (aka comfort zone) by fleshing out the outline. Occasionally I would use ChatGPT to help describe a concept. For instance, I had it summarize the description of the movie Dr. Strangelove. All other content generated by AI that appears in the paper is considered a quote and is also indented.

Often, while editing in Word, Microsoft CoPilot would constantly offer to rewrite sentences that may not be perfect grammar. That quickly became annoying. I prefer prompting ChatGPT every now and then, rather than having AI disruptively work in real time.

It would have been very easy simply to let ChatGPT write large sections of the paper, but you can get very odd results, even erroneous information. Instead, I chose to treat it like a help system and asked it questions periodically about specific topics, mostly about topics I had written already. Also, the copyright issues are still very unclear, so I assumed the strict interpretation of the concept of Fair Use.

Once I had the main text done. I fed the article into ChatGPT and asked to summarize it to create the abstract Strangelove (as described in the corresponding footnote). I liked the results so much I didn’t rewrite it much. That and the summary of issues at the end of the paper are the only blocks of AI generated text that I used without much editing. There may be a sentence or phrase here and there that were only modestly edited, and since they were inline and very short, I made them into my own voice and felt like they did not need to be quoted.

I used ChatGPT to convert simple tabular text into XHTML table markup. It worked well and produced valid XHTML tables. See Summary of XHTML Table Conversion below for more details.

Conversion Process

When I felt the paper was mostly done, I tried to use ChatGPT to convert it to Balisage DocBook tagging with the following prompt.

I would like you to convert a Word document into DocBook, specifically the Balisage-1-5 version. Note that I used quote styles to indicate content that is a response from ChatGPT. Also, look for references to intellectual property, dates, cross references. If you have any questions ask me.

I may have asked for too much. It ran all night and didn’t produce output. I was using the free version so it may just have exceeded a processing limit. It also made me wonder about how much energy was used during my sessions with ChatGPT. Perhaps the carbon equivalent of several cross-country flights I fear. I hope it just stopped. I intentionally told it to stop in the morning, just in case. I hope it wasn’t too late.

So, instead of doing the conversion by hand, I used the XML Mind utility[18] on the Web. The results were valid DocBook XML.

Finishing Touches

The XML document I produced was valid. Then I asked ChatGPT to look for anything in the paper that might need copyright references:

Can you review the same document and identify significant references or information that might need a footnote added to capture rights information.

The result was useful:

If you include or refer to a still from the movie, you must note its copyright: "Still image from Dr. Strangelove, © Columbia Pictures Industries, Inc., 1964. Used here under fair use for commentary."

XHTML Table Conversion

This section summarizes a discussion with ChatGPT about converting plain text into a styled XHTML table.

Original Request: I requested ChatGPT to convert a plain text block into an XHTML table. The prompt and input data were submitted like this: 

Convert the following text into an HTML table. Use the first row as the header. Set the first column width to 35% and the second column to 65%.

Modality          Examples 

Text              Messages, documents, code, web pages
Images            Photos, drawings, screenshots, diagrams 
Audio             Voice commands, music, podcasts 
Video             Surveillance clips, social media, tutorials 
Structured data   Tables, spreadsheets, sensor logs

Generated XHTML Table Output:

<table border="1" style="width:100%;"> 
<thead> 
<tr> 
<th style="width:35%;">Modality</th> 
<th style="width:65%;">Examples</th> 
</tr> 
</thead>

<tbody> 
<tr> 
<td>Text</td>
<td>Messages, documents, code, web pages</td>
</tr> 
<tr> 
<td>Images</td> 
<td>Photos, drawings, screenshots, diagrams</td>
</tr>
<tr> 
<td>Audio</td> 
<td>Voice commands, music, podcasts</td> 
</tr> 
<tr> 
<td>Video</td> 
<td>Surveillance clips, social media, tutorials</td>
</tr> 
<tr> 
<td>Structured data</td> 
<td>Tables, spreadsheets, sensor logs</td> 
</tr> 
</tbody> 
</table>

Observations

AI can provide some serious benefits to a variety of publishing activities. And it also can introduce serious problems to the publishing process. Best practices and guidelines need to be clarified for it to be used appropriately.

AI is surprisingly easy to learn and use. Treat it like a smart helper and ask it questions.

Prompts should be worded like the kind of search request you always dreamed of in the Google world. In plain language with conditions and other info. Like sentences you would ask a mentor. AI understands you in an eerie way.

Everything AI produces should be carefully verified.

Applications include fact checking, date and format validation, style consistency checking, and many other time-consuming activities.

Paying attention to AI as its use expands is important to the publishing world.

For the record, I used the same process to create this auxiliary paper.


[1] This abstract was generated using ChatGPT by copying the text into the prompt and asking it to create an abstract. It was not edited.

[2] Film synopsis of Dr. Strangelove (1964) directed by Stanley Kubrick. Source: ChatGPT summary.

[3] DeRose, Steven J. “Can LLMs help with XML?” Presented at Balisage: The Markup Conference 2024, Washington, DC, July 29 - August 2, 2024

[4] Generated by ChatGPT

[5] Generated by ChatGPT

[6] Generated by ChatGPT

[7] Generated by ChatGPT

[8] See: "Legal Implications of AI-Generated Content" - [https://www.wipo.int](https://www.wipo.int)

[9] Reference: GDPR Compliance Guide - [https://gdpr.eu](https://gdpr.eu)

[10] Article: "AI Ethics and Regulatory Compliance" - [https://ai.gov](https://ai.gov)

[11] Study: "Bias in AI Models" - [https://www.aialliance.org](https://www.aialliance.org)

[12] Paper: "Human-Centric AI Design" - [https://arxiv.org](https://arxiv.org)

[13] Report: "AI Adoption in the Workplace" - [https://www.forrester.com](https://www.forrester.com)

[14] Best Practices: "Validating AI-Generated Code" - [https://developer.mozilla.org](https://developer.mozilla.org)

[15] Guide: "AI Integration Strategies" - [https://www.ibm.com/cloud](https://www.ibm.com/cloud)

[16] Whitepaper: "AI & Content Versioning" - [https://www.microsoft.com](https://www.microsoft.com)

[17] Prompt Engineering with ChatGPT - Ronnie Sheer, LinkedIn Learning

What is Generative AI - Pinar Seyhan Demirdag, LinkedIn Learning

Ahead of the Curve: Technical Writing, Semantics, and AI - Leif Erikson amd Rob Morrison, A Content Wrangler Talk

[18] https://www.xmlmind.com/w2x/docx_to_docbook.html