Category: Other Page 77 of 177

AIMark is a tentative exploration for a method to use a custom markup or markdown formatting on AI-assisted texts, where the contributions made by human author(s) versus AI(s) are clearly and meaningfully differentiated.

Contributions to the project:

• You.com/chat came up with the AIMark name and basic concept based on my inputs (it wrongly claimed that this whole thing exists already, but based on my research it does not – perhaps it was seeing the future?)
• ChatGPT helped me work through the code examples for the custom markup and markdown elements, as well as walking through the thought process collaboratively.

Example Markup Usage (custom)

<defs>
  <def author type="AI" model="AI model X" version="3.2" source="OpenAI" />
  <def author type="human" name="John Doe" />
</defs>

<p author="AI">
  The quick brown fox jumps over the lazy dog<del author=”human”>.</del>
<ins author="human">, but sometimes the fox is not quick enough.</ins>
</p>

In the above example, it its proposed to use <defs> tag to set up definitions for authors, whether AI or human, and any pertinent details about them that may apply globally within a document.

Then each element is intended to be clearly marked as to which author produced or modified which elements. <del> is used for deletions, and <ins> for insertions. In the example above, the AI tool generated the text The quick brown fox jumps over the lazy dog. and then the human deleted the period, and added the clause, , but sometimes the fox is not quick enough.

Each of those elements could also take inline attributes if that is preferable, such as: <del author=”human” name="John Doe" date="2023-01-23">

Example Markdown Usage (custom)

In markdown, we’re adopting the convention that the % percent sign indicates AI and the forward slash / indicates human content.

So in its simplest form a line of AI generated text would look like:

%This text was generated by AI%

Where the text starts and ends with %. Likewise, for human text, the usage would be like:

/This is human-generated text./

If you want to add attributes to either, you could possibly (?) do it in () either at the beginning or end of the text passage. And attributes could be named pairs or just the value. So like:

AI Attribution Short:

%This text is AI generated(AI model X,3.2,OpenAI)%
%(AI model X,3.2,OpenAI)This text is AI generated%

Longer:

%This text is AI generated(model:AI model X,version:3.2,source:OpenAI)%
%(model:AI model X,version:3.2,source:OpenAIAI model X,3.2,OpenAI)This text is AI generated%

Human Attribution Short:

/A human wrote this(John Doe)/
/(John Doe)This text is AI generated/

Longer:

/A human wrote this(name:John Doe)/
/(name:John Doe)This text is AI generated/

Then there could be a form for text that is longer than a short sentence, which could be written as:

[%This is a longer block of AI-generated text]

[/This is a longer block of human-written text]

In the above example, use of [] would mean you do not need to close the other marks.

To indicate a nested human edit of an AI-generation, it could be something like this, where ~ means strikethrough (deletion) and ^ means insertion.

[%An AI-generated a big block of text and it was [/~good][/^bad]

Or using the short non-bracketed form of both:

%An AI-generated a big block of text and it was /~good~^bad^/%

So in the above examples, a human deleted the word good, and inserted the word bad into the AI content generation.

Global author definitions in such a custom markdown document could be something like:

%%
Model: AI model X
Version: 3.2
Source: OpenAI
%%

//
Name: John Doe
//

This is a poorly defined concept overall, but I think a fun one worthy of another look:

Hypertext fiction is characterized by networked nodes of text making up a fictional story. There are often several options in each node that directs where the reader can go next. Unlike traditional fiction, the reader is not constrained by reading the fiction from start to end, depending on the choices they make. In this sense, it is similar to an encyclopaedia, with the reader reading a node and then choosing a link to follow.

See also: networked narrative, transmedia storytelling

I’m not necessarily a believe that all AI-assisted text or images categorically need to be labelled as such, everywhere, all the time. It can be a good idea in some cases, however. And in cases where it is desirable (putting aside the “why”) for now, what might be the ideal methods to do it?

I’ve seen a few of approaches so far, but not many:

• Don’t disclose the use of AI assistance in a text
• Disclose it at the beginning or end of a document
• Disclose it and estimate the relative ration of human-to-AI content in a given work (the so-called HI2AI number)

Each may have its appropriate use, but none of them offer any pointers as to how to visually display within a text which elements were human-generated and which were machine-generated. It doesn’t give us any specific advice or tools for marking up a document in a way that might be useful or meaningful in some manner to human (and other) readers.

Here are a couple initial ideas to differentiate within a text:

• Human-written text is in one color (or is available in), and AI generations in another
• Same thing, but with highlighting, instead of or in addition to text color

Here is the first as a mockup, using what happens natively in Verb.ai:

Black is the human generations, and green is the AI generations. Colors seem useful here because they are not too disruptive, and seem to add a sort of dimensionality to the text.

One issue here is where displays are restricted to black and white, or for use in print applications. So not relying on color is probably one constraint we should design with in mind.

Also: what about color blindness? What about the fact that different sites or services might use different color schemes?

Getting back to our list ideas:

• Use superscript characters like ᴬᴵ and ᴴ to mark transitions within a passage.

Let’s try that out with our quick example passage, but here in WordPress:

---

ᴴThis is the beginning of a ᴬᴵstory. John had spent the last three years ᴴeatingᴬᴵ, sleeping, and going to work. His job was not particularly interesting or rewarding, but it paid the bills ᴴand that would do for now.

---

I don’t know about you, but that seems awkward to me to read. It is less disruptive than I was expecting, but seems like it would quickly grow tiresome. Plus it would be cumbersome to manually insert and track those transitions in a long text.

We could do the same thing, but sub in emojis to indicate speaker:

---

🗣️This is the beginning of a 🤖story. John had spent the last three years 🗣️eating🤖, sleeping, and going to work. His job was not particularly interesting or rewarding, but it paid the bills 🗣️and that would do for now.

---

Also somewhat disruptive to my eyes as a reader. And it doesn’t display the para-textual subtlety, the sort of “shades of subtext” that I feel the color example at the top does.

Let’s go back to the list:

• Using different fonts to indicate speakers: it could work, but there are often applications where only one font is available.
• Which leads to: text decoration, such as underlines (and overlines), italics, or bold. Those may be possible in some cases, but often these text styles or decorations may have other senses to their use, such as for emphasis. So perhaps it would be better to colonize an unoccupied typographic space…
• In-line containers like different types of brackets {},[],(),||,/,\,’,”,`,~. This might be an option, but again there’s the possibility of collision with other semantic or semiotic uses. Let’s run again our text, and try to do / to indicate human and \ to indicate AI.

---

/This is the beginning of a \story. John had spent the last three years /eating\, sleeping, and going to work. His job was not particularly interesting or rewarding, but it paid the bills /and that would do for now.

---

It’s somewhat awkward, but it’s for me anyway a little less intrusive than some other experiments here. One thing that’s cool is that human-gens are under a sort of /little house\ typographically. And AI-gens are almost like they are \outdoors/.

Another variation below, where | indicates a start and stop of AI-gen text, and human text is not indicated.

---

This is the beginning of a |story. John had spent the last three years |eating|, sleeping, and going to work. His job was not particularly interesting or rewarding, but it paid the bills |and that would do for now.

---

I think that’s confusing cause you don’t know easily what is inside and what is outside the | since it starts and ends with same sign.

Anyway, there are certainly other possibilities and I’m sure we will see them flourish over the coming months and years as these technologies become more widespread in writing tools.

Will continue to explore others in follow-ups, time permitting.

Thought this was a decent short read & some cool pics from an AI artist. I wish their art were available to look at off Instagram, cause hard to see w/o an account.

“To me, it resembles a very similar process and method to producing images as an art director. They typically don’t execute the technical aspects but decide and inform a team about the required aesthetic of several aspects. The same relationship can be created with an AI. Deciding on the general concept, casting, location, hair and makeup, lighting, colour grading, the fashion and having to put it into context with other images. The AI is basically the team that returns an image based on all these aspects. Hence I find it collaborative in a certain weird way, however, I found the social aspect of collaboration always very important and working alone with software is devoid of any of that.”

Sentience is a complicated topic. I won’t pretend to understand all its vagaries, especially once we try to apply that to AI with a still imperfect understanding of both domains…

Selfhood, likewise, is a squishy thing to try to define. But maybe becomes a little less squishy because it becomes eventually somewhere along the spectrum an embodied thing. I have a body, therefore its somewhat difficult to argue I don’t have some sort of “self.” I might not know or be able to define exactly what that self is or what exactly my “having” it might mean. But it’s a somewhat tangible thing I can at least point to as being connected to my self.

Humans have an incredible ability to empathetically project self-hood and sentience onto other things, though we might not all agree which are which. Regardless, it’s a thing we do somewhat automatically because of our fundamental makeup and nature as embodied selves in the not-only-virtual world, but also the so-called “Real World” whatever that even is.

My question here then is, might it be either a good idea or a not good idea – or more likely some mix of the both – that we encourage/allow/program our tools to assert their own self-hood as part of their fundamental UX.

If you ask ChatGPT basically any question, it will respond with a slew of “I” and “me” and “mine.” Which, okay, for convenience, I get it. It’s a chat experience. Chatters assume they have selves, and that the other party also has some sort of self. But, what would it look like if we removed the assumption that this is a desirable state? How might that change the communication styles of these chatbots?

How would the program have to identify itself and communicate about its capacities?

Maybe something like, Instead of saying “I’m just a large language model…” It would have to say something like, “The program is a large language model…”

Would this have meaningful impacts on users of the technology? More specifically, would it give people more of a mental shield against perhaps overly identifying a program as having some kind of equivalency with a human being?

I’m not saying the Butlerian Jihad is coming, but you never know…

This seems like as good a definition as any of AI literacy:

“a set of competencies that enables individuals to critically evaluate AI technologies, communicate and collaborate effectively with AI, and use AI as a tool online, at home, and in the workplace”. (Long and Magerko ,2020)

I like that there is an emphasis here on collaboration. This is the Way.

Via Jeff Jarvis’ very worthwhile Medium post on the potentials of alternative modes of literacy co-evolving with AI:

“In the end, writing a prompt for the machine — being able to exactly and clearly communicate one’s desires for the text, image, or code to be produced — is itself a new way to teach self-expression.“

We’re in kind of a golden era right now of AI content generation apps coming online that are free during at least a trial period. A lot of them switch to paid after, but I kind of find it fun to only use them temporarily like some sort of migrant AI user or a hermit crab, temporarily putting on a shell built by some other animal, and discarding it when it no longer fits. Plus there’s just so much experimentation and competition right now & nobody really has got the UX all nailed down to be really the top down interface provider. Though obviously some of the underlying models are more compelling than others, but in the vast majority of cases (outside perhaps ChatGPT for now), you can access them through a variety of third parties. Eventually, I guess this will change, and paid access regimes will change, but for now here’s to the free rider. (That said, I do pay for Dall-E)

There’s a lot to like in this article, but wanted to save this bit:

“…entirely new categories of powerful addictions are available to us that weren’t available to our ancestors, and it should be uncontroversial to be worried about those effects somewhat. AI is going to let us invent even more.”