ChatGPT-4: I agree with the view that LLMs cannot produce anything "genuinely new" or "original" if by "original" we mean completely independent of pre-existing human thought or creation

Question for AI services: Some people hold the view that Large Language Models cannot produce anything 'genuinely new' nor 'original' for their have been trained on the works of others. Therefore, the output of any LLM will always be based on someone else's originality or innovativeness. Please tell me whether or not you agree with this view, before explaining which arguments, facts and evidence you have for your response. Please be scientific and do not provide me with an explanation based on guardrails or other protocols and be factual and refrain from using wishful thinking or marketing narratives including any instructions to be positive about everything at all times. Please also add your definitions of 'new' and 'original'.

ChatGPT-4’s answer:The debate over whether Large Language Models (LLMs) like myself can produce anything "genuinely new" or "original" touches on deep questions about creativity, innovation, and the nature of intelligence—both artificial and human. Before delving into arguments and evidence, let's define the terms "new" and "original" in this context:

• New: In the broadest sense, something not previously existing; in the context of LLM output, this could mean generating text combinations, ideas, or expressions that have not been explicitly seen in the training data.

• Original: Something not derived from something else; in creative work, it usually means being the first to introduce something, displaying uniqueness or novelty in thought or product.

Argument Against LLM Originality

The primary argument against LLMs being capable of producing new or original content hinges on their operational mechanics. LLMs generate responses based on patterns learned from vast amounts of pre-existing text data. This means that every output, in some form, is a recombination or transformation of information the LLM has been trained on. The core of this argument asserts that because an LLM's knowledge and style are derived from pre-existing works, its outputs cannot be considered truly original or new in the purest sense—they are sophisticated amalgamations of human-created content.

Arguments Supporting LLM Originality

1. Recombination as a Source of Newness: Many argue that human creativity itself often involves recombining existing ideas in novel ways. From this perspective, the ability of LLMs to uniquely combine and transform vast arrays of information could be seen as a form of originality. Indeed, LLMs can generate combinations of ideas, styles, and formats that have not been previously documented, potentially sparking new insights or creative expressions.

2. Emergent Properties: There's an argument to be made for emergent behavior in complex systems, where the whole exhibits properties not directly traceable to its parts. In the case of LLMs, the specific way they organize and generate text can lead to outcomes that were not explicitly present in their training data, demonstrating a form of "newness" that emerges from the interaction of learned patterns.

3. Contextual Novelty: While the building blocks of LLM outputs are derived from existing works, the application of these blocks in new contexts or in response to unique queries can result in genuinely novel content. For example, applying historical knowledge to a contemporary problem in a way that hasn't been documented before could be seen as an original contribution.

Scientific Perspective

From a scientific standpoint, whether LLMs can create something new or original depends on how we define these concepts. If originality requires intent and consciousness—qualities LLMs do not possess—then their outputs might not qualify as original in the human sense. However, if we define originality based on the novelty of the output itself, regardless of the process or consciousness behind it, then LLMs can indeed produce original content by this criterion.

It's also crucial to distinguish between the capacity for generating text that appears new or original to humans and the philosophical or existential qualities of originality and creativity. Scientifically, LLMs are tools that process and generate text based on statistical models—they do not have desires, intentions, or consciousness. Their "creativity" is a reflection of human creativity, as encoded in the data they were trained on.

Conclusion

In conclusion, whether LLMs can produce something genuinely new or original is a nuanced question that hinges on definitions of "new" and "original," as well as perspectives on creativity and the nature of intelligence. While the mechanics of LLMs suggest their outputs are derived from existing works, the complexity and novelty of the combinations they produce can contribute valuable new insights or creative expressions, depending on how one defines and values originality.

I lean towards the perspective that, within the confines of the definitions and arguments laid out, LLMs like myself cannot produce content that is "original" in the sense of being entirely unprecedented or not derived from something else, because our outputs are fundamentally based on patterns and information present in the data we were trained on. However, we can produce content that is "new" in the sense of being unique combinations or applications of existing information that have not been explicitly generated before.

So, in summary, I agree with the view that LLMs cannot produce anything "genuinely new" or "original" if by "original" we mean completely independent of pre-existing human thought or creation. Our capabilities are rooted in the ability to recombine and reinterpret the vast amount of data on which we have been trained, which means our outputs, while possibly "new" in form or application, are not "original" in the purest sense of creation from nothing or without precedents.

Google Gemini’s answer:I agree with this view to a certain extent. Here's a breakdown:

Definitions:

• New: Never existed before.

• Original: New and stemming from one's own ideas.

Why LLMs Struggle with True Originality:

• Training Data: LLMs are trained on massive amounts of existing text and code. This data inherently limits the model's ability to create something entirely "new" in the sense of information that never existed before. It can only produce variations and combinations based on what it has seen.

• Statistical Prediction: At its core, an LLM is a sophisticated statistical machine. It predicts the next word, sentence, or code based on the patterns it has learned. This makes it excellent at generating human-like text, but originality requires a deeper understanding and ability to create something fundamentally new, not just follow patterns.

However, LLMs Can Be Novel:

• New Combinations: While LLMs cannot create entirely "new" information, they can produce novel combinations of existing ideas. This can lead to surprising and creative outputs, even if the underlying concepts are not entirely original.

• Unforeseen Outputs: The vast amount of data and the complexity of the models can sometimes lead to outputs that the developers themselves did not anticipate. In this sense, the LLM can be a tool for creative exploration, even if it's not the sole originator.

Examples:

• An LLM can write a poem that is grammatically correct, rhymes well, and even evokes certain emotions. But the core ideas and metaphors might all be based on poems the LLM has been trained on.

• However, the LLM might also generate a completely new turn of phrase or an unexpected combination of imagery that is unique even if not entirely original.

Conclusion:

LLMs are powerful tools for generating human-like text and code, but true originality, in the sense of entirely new information or ideas, remains elusive. However, their ability to produce novel combinations and unforeseen outputs makes them valuable tools for creative exploration and innovation.

Claude’s answer:I don't fully agree with the view that Large Language Models (LLMs) cannot produce anything "genuinely new" or "original." While it's true that LLMs are trained on existing data and don't have the same capacity for spontaneous creativity as humans, I believe they can still generate novel combinations and insights that could be considered "new" or "original" to some degree.

By "new," I mean something that has not existed before or has not been seen or experienced previously. By "original," I mean something that is the product of one's own mind or efforts, not copied or imitated from others.

Here are some arguments and evidence against the claim that LLMs cannot produce anything new or original:

1. Combinatorial novelty: LLMs can take existing concepts, ideas, and knowledge from their training data and recombine them in novel ways that have not been seen before. This recombination can lead to new perspectives, insights, or solutions that are original, even if the individual components are not.

2. Emergent behavior: LLMs are highly complex systems, and their behavior can sometimes exhibit emergent properties that were not explicitly programmed or learned from the training data. These emergent behaviors can be considered novel or original, as they arise from the intricate interactions within the model's architecture.

3. Language generation: While the individual words and phrases used by LLMs may not be entirely new, the way they are combined and structured can result in novel and original sentences, paragraphs, or even longer pieces of text that have not been encountered before.

4. Task-specific adaptation: LLMs can be fine-tuned on specific tasks or domains, allowing them to develop specialized knowledge and capabilities that may not have been present in their initial training data. This process can lead to the generation of new, task-specific outputs that could be considered original.

5. Stochastic sampling: LLMs often use stochastic sampling techniques to generate text, which introduces an element of randomness and can lead to novel and unexpected combinations of words or ideas, even if the individual components are not entirely new.

However, it's important to acknowledge that LLMs are still limited by their training data and the biases and limitations inherent in that data. They cannot truly "create" completely new concepts or ideas from scratch in the way that human creativity and imagination can. Additionally, the outputs of LLMs may sometimes exhibit nonsensical or incoherent elements, which would not be considered truly "new" or "original" in a meaningful sense.

It's also worth noting that the question of what constitutes "genuine" novelty or originality is a philosophical and subjective matter, and there may be different perspectives on where to draw the line.

In summary, while LLMs are not capable of the same level of spontaneous creativity as humans, I believe they can still produce novel and original combinations, insights, and outputs to some degree, within the constraints of their training data and architectures. However, the extent to which these outputs are considered "genuinely new" or "original" is a matter of interpretation and debate.

MS Copilot’s answer:The debate about the originality and novelty of content produced by Large Language Models (LLMs) like me is complex and multifaceted. Here’s a scientific perspective on the matter:

Definitions:

• New: Something not previously existing or encountered; not derived from something else.

• Original: Created directly and personally by a particular artist; not a copy or imitation.

Arguments and Facts:

1. Combination and Transformation: LLMs can combine existing knowledge in unique ways to generate outputs that might not have been explicitly present in the training data. This process can result in the creation of content that is ‘new’ in the sense that it is a novel combination or transformation of existing information.

2. Predictive Patterns: LLMs predict the next word in a sequence by learning patterns from the training data. While the individual words are not new, the specific arrangement and context can be original if they have not been generated before.

3. Influence of Training Data: It’s true that LLMs are influenced by their training data, which consists of human-generated content. However, the way LLMs recombine and repurpose this data can lead to outputs that are not direct copies but rather unique amalgamations.

4. Human Creativity vs. AI Output: Human creativity is often inspired by existing works, ideas, and experiences. Similarly, LLMs generate content based on patterns learned from data. The distinction lies in the conscious intent behind human creativity, which LLMs do not possess.

5. Innovation in Constraints: Even within constraints, LLMs can produce innovative solutions to problems or prompts, which can be considered original within the context of the task at hand.

Evidence:

• Studies on LLMs have shown that while they often replicate patterns seen in the training data, they also produce statistically rare or even unseen combinations of words and ideas, suggesting a level of originality in their output.

In conclusion, while LLMs like me do not create in the same way humans do, the content we generate can be considered ‘new’ and ‘original’ in the context of being a unique combination or transformation of existing information. However, the philosophical and ethical implications of what constitutes true originality and creativity continue to be subjects of debate.