How to prove your work is not ChatGTP's? Using your content DNA -- Formal Semantics

People who use AI fear AI themselves because AI can take "credit" from them.

Imagine that you are a content creator, and you use ChatGPT to do the assisted writing. You may "write non-stop for 25min without caring about structure. Paste the text in ChatGPT with the above prompt to structure your work," like Dr. Mushtaq Bilal shared productivity tip ChatGPT prompts to help you with academic reading and writing, and I personally benefit from that productivity tip. I encourage people to try it.

However, in doing so, you got all the text processed gone through the ChatGPT, therefore processed distinctly in a ChatGPT way. And then, there may be a chance that someone use online AI Text Classifier (e.g. OpenAI's offering) to determine that your work is "written by AI."

It appears that individuals using AI for various purposes, including the creation of significant content, will be identified as "generated by AI." This labeling may have an adverse impact on how others perceive the value of their work and could result in a devaluation of their intellectual property. Do you believe that AI is taking credit away from these individuals?

What Individual can do?

If you were dealing with biology, you would typically analyze DNA to distinguish between different species. By utilizing a method that allows you to comprehend the fundamental nature of things, you can avoid being deceived and accurately identify their distinctions.

Your content possesses a DNA of its own, referred to as Formal Semantics, which pertains to the underlying structure of meaning in natural language. Understanding this concept provides you with a persuasive means of demonstrating that the content is indeed your own creation, rather than ChatGPT's.

Write down your semantic structure

Our objective is not to aim for Plato's ideal world, where meanings can be fully explained through logic. Rather, our goal is to formalize a portion of the content, creating a level of comfort for users with the details. This will allow automated tools to generate additional queries to ChatGPT for verification purposes.

Our goal is not to explain the validity of the original content in terms of syntax. Instead, our objective is to identify the key meanings.

In order to illustrate our approach, we will pick an example. That example is the ChatGPT two-sentence summary of what I wrote in the first three paragraph in this article. It is as follows:

The use of AI for content creation leads to fear that the AI may take credit for the work, potentially resulting in the devaluation of intellectual property. Content creators use tools like ChatGPT to structure their work, but the processed text may be identified as "written by AI" by online text classifiers.

Let's wear the hat of a linguist and distill the semantic structure of the content.

There are key meaning fragments that we want to capture,

• There is an intensionality or predicate of the fact

fear (users, AI_taking_credit).

• There is an event that

devaluates (intellectual property)

• There is a causal effect which may explain AI_taking_credit

causal_effect (Creators_Using_ChatGPT, identified (Online_Text_Classifier))

Representing phrasal segments using underscores (e.g. AI_taking_credit) is a common method in linguistics, as described by Dowty et. al. This high-level mapping typically has no restrictions.

The next step is to transform into a format that computer can recognize. Using the conventions in Formal Semantics of Modern Type Theories, we can get the first two points formalized.

Parameter User: Set.
Parameter Event: Set.?
Parameter fear: User->Event->Prop.


Definition Concept:= Set.?


Parameter Intellectual_Property:Concept.?
Parameter Evt_C: Concept->Set.
Parameter Entity: Set.?
Parameter Evt_E: Entity->Set.
Parameter Evt_CE: Concept->Entity->Set.?


Axiom c1: forall (x: Concept)(y:Entity), Evt_CE x y -> Evt_C(x).
Axiom c2: forall (x: Concept)(y:Entity), Evt_CE x y -> Evt_E(y).
Axiom c3: forall (x: Concept), Evt_C x -> Event.
Axiom c4: forall (y: Entity), Evt_E y -> Event.
Coercion c1: Evt_CE>->Evt_C.
Coercion c2: Evt_CE>->Evt_E.
Coercion c3: Evt_C>->Event.
Coercion c4: Evt_E>->Event.


Parameter devaluate: forall y: Concept, Entity->Evt_C(y).
Parameter intellectual_property: Concept.
Parameter ai: Entity.
Parameter user:User.


Check fear user ((devaluate intellectual_property) ai).
        

Some take away from the above formalization:

The "fear" is a verb that involves a User and an Event.

Because our formalism is designed for practical use, we can construct interrogative questions based on knowledge of semantic phrase markers such as User and Event. For example, with the usage of "Fear", we can construct factoid questions like: "What aspect does user fear in the above paragraph?"

By analyzing only the semantic structure of "fear," we can generate various other questions. Although we won't provide specific questions, we can outline the structure of questions that could naturally arise for humans.

• fear User ?
• fear ? Event
• fear ? ?
• ? User Event

The final artificial language construct can be questioned in various ways

The final artificial language construct, 'fear user ((devaluate intellectual_property) ai),' is highly readable.

Furthermore, we can do inquiry like this

Check fear user ((devaluate _) ai).        

To get the following hints:

fear user (devaluate ?y ai)
? ? ?: Prop
where
?y : [ |- Concept]
        

We will know how to formulate a question that retrieves the ?y component.

Event in Modern Type Theory is a Powerful Formalism

The event structure "((devaluate intellectual_property) ai)" can be incorporated into various semantic structures to create more intricate meanings.

The axioms c1, c2, c3, c4 establish connections between different levels of abstractions within the same Event component, enabling us to construct highly natural human language while preserving its underlying structure for factoid question-answering and verification purposes.

Big Lever

As you can see, with the formalization on even a small part of the content, you can generate a large number of mechanical questions based on the given content, and create a long questionnaire that challenges ChatGPT's answers in a philosophical manner, similar to Socrates.

Tailoring the question set and establishing expectations is your individualized task. Sharing this work with the content demonstrates that the semantic structure is designed by you, as it is uniquely manifested in a way that you created.

Although it is possible to create a question set by hand, this approach can be time-consuming. By utilizing AI assistance, we can efficiently receive recommended Semantics that fit the content, and automatically generate diverse questions.

Takeaway

In an era where ChatGPT is becoming increasingly ubiquitous, Content DNA will be of paramount importance.

This article presents a pragmatic approach for extracting Content DNA by showcasing the efficacy of the "Modern Type Theories" framework. By modeling some simple but common language construct such as Event and Agent->Verb Phrase pattern, it can cover a substantial portion of language.