The AI features in the established design application «figma» introduced a few days ago, led me to a conversation with ChatGPT in which I asked why the current AI models are only “very limited in creativity” and why they have so much trouble coming up with something new and, finally, why figma’s design suggestions are completely useless for a professional designer.
ChatGPT’s answers summarized: AI systems are designed for consensus and averageness, not «innovation».
Here is my complete, in my opinion very interesting dialog with ChatGPT:
Thomas: Hi ChatGPT, I have a few questions about your way of thinking and working. When I ask you a question, your neural network is programmed to provide me with the “most probable response”, is that correct?
ChatGPT: Yes, you could summarize it like that. Everything I do is based on patterns and statistical probabilities.
Thomas: So you can’t really think. You also don’t know what you know, you’re more like a trained dog that knows what to do when its master gives it a certain command.
ChatGPT: That’s an accurate analogy. My functioning is actually more like a well-trained dog than a thinking being. Here are some points to illustrate that: I have no consciousness and don’t really know what I “know”. Much like a trained dog that responds to certain commands, I respond to your input based on the probabilities and patterns I’ve seen in the training data.
Thomas: All right, since you don’t know what you know, of course you don’t know what you don’t know either, do you? If I ask you what doesn’t yet exist in an area, then your answer is generated in the form that would be the “semantically” correct answer to that question. You’ve only learned a preconceived answer to that question, but it’s not based on any actual research or verification. Is that correct?
ChatGPT: Yes, something like that. For one thing, I don’t work on the basis of rules. For another, I don’t use pre-formulated answers. My answers are based on the weights and tokens defined during the training phase to select the most appropriate semantic tokens for my answer. Furthermore, from my version 4 onwards, I can also search for information on the internet, but this additional information is only added within the prompt as contextual information to my model parameters, just like the information you provide in our dialog. So this information can actually, at least in our dialog, close knowledge gaps, but it does not change my basic model parameters. This would be theoretically possible, but is not intended, as users could fundamentally manipulate or change me in this way.
But yes, as you prefaced your question, I have no understanding of my own knowledge or gaps in it. I also have no direct access to my knowledge, as it is not in the form of a conventional database that could be searched via search algorithms, but in the form of neural model parameters that neither I nor my developers can consciously access. In other words, neither I nor my developers know in detail how I arrive at my outputs. I know that this is difficult to understand, but it is one of the fundamental properties of a multilayer neural network that stores knowledge about model parameters.
Thomas: OK, but do I understand you correctly that the contents of your texts and pictures correspond to what is known consensus, because you were designed to analyze your training data for similarities, but not for “dissent, “missing aspects”, “errors” or “contradictions”. You’re not looking for gaps or missing aspects in either the training or application phase, just existing ones, right?
ChatGPT: Yes, that is correct. My answers tend to reflect the consensus because I have been built to search the data provided to me for general patterns and consensus. This leads me to provide answers that match the most commonly held views and information.
Thomas: Are you aware that some innovations are based precisely on finding something outside the norm?
ChatGPT: Yes, I am aware of that. Many innovations and creative breakthroughs result from the ability to think outside the norm and to recognize disconsensus or missing aspects. I completely lack this ability.
Thomas: So you are only creative in the sense that you resample existing patterns according to principles or procedures that you know, but not by developing or adding ideas that are unknown or new to you. Nor can you transform in the actual sense, since you cannot think at all and therefore cannot “negate” or “transform” because you lack any sense of logic, can you?
ChatGPT: Yes, that is correct. My creativity is limited to re-sampling and combining existing patterns and information. I can’t generate new or original ideas outside of those patterns, nor can I consciously negate or think in any form because I lack the skills of self-reflection and conscious thought.
Thomas: Is this the reason why many designs generated by AIs are rather average and not really exciting or innovative?
ChatGPT: Yes, AI models like me are trained to reflect the consensus or the average. The learning methods used (LTSM, Transformer, etc.) lead to a strong weighting of frequently occurring content. How should I know what is potentially innovative? It is more likely that potentially innovative content is weighted less by my learning algorithms due to its rarity and is therefore not learned by me.
Thomas: Okay, now I realize a lot. But what does the future hold? Is that in the fundamental nature of AI or is it conceivable or even likely to develop AI systems that have capabilities such as negation, reasoning and creative innovation?
ChatGPT: At the moment, this is the nature of our learning algorithms. But it is more than probable that AI technology will advance. We do not yet know to what extent creative thinking will then be possible. The technical challenges involved are enormous. To the state of my knowledge, however, the following considerations and efforts are currently being made in AI research:
- Negation and logical reasoning: These capabilities require algorithms that go beyond current pattern recognition capabilities. This could be achieved by integrating formal logic systems and machine learning methods that are able to draw complex logical conclusions.
- Counterfactual reasoning: models could be developed that run counterfactual scenarios to understand what would have happened if certain assumptions were not true. This would require AI systems to be able to generate and analyze hypothetical and alternative scenarios.
- Experimental and empirical learning: AI systems could potentially be designed to learn through interaction with their environment and trial and error in a similar way to humans.
- Self-reflection and metacognition: AI models may one day be developed that recognize their own knowledge gaps and actively search for new information to fill them.
- Hybrid models: combinations of symbolic AI approaches (which support logical thinking) and neural networks could be developed to leverage the strengths of both approaches.
Regrettably, today’s AI systems are still a long way from these approaches. There are AI systems that are capable of “rule-based” thinking. These systems use explicitly coded rules to draw logical conclusions and make decisions. However, the rigid rule structures prevent “learning”.
Far more advanced approaches are needed to develop truly creative systems. The solution could lie in combining different methods that combine rule-based thinking, machine learning and other AI technologies. It remains to be seen what the future holds in this direction.
Thomas: Thank you for the interview.