Is AI creative?

1. What exactly do we mean by originality?

The starting point was a short definition: creativity is something unconventional. Problems arose when applying it to AI. The assistant treated new combinations, analogies and transformations of learned patterns as evidence of creativity. The response was that these might still be skilful operations on patterns.

This qualification matters. The number of possible combinations may be enormous and the result unique even though the method remains routine. Uniqueness alone does not settle whether a creative idea has arisen. Nor does surprising an audience explain how a solution was produced; it also says something about their expectations.

Logic, letters and numbers are the foundation, not a finished idea. Pointing out that a new mathematical proof must be logically valid therefore does not answer whether it is creative. Basic conditions for meaningful reasoning must be distinguished from an established way of framing and solving a particular problem.

In this part we adopt a stronger criterion: creativity includes questioning the existing framing of a problem, its assumptions or criteria for success. We do not reduce it to finding a better variant within unchanged boundaries. This is the article's deliberately chosen perspective, not the only possible definition of creativity.

2. AlphaGo: a new strategy in the same game

AlphaGo Zero was raised as an argument for AI creativity. The system learned by playing itself without training on examples of human games. It combined a neural network evaluating positions and proposing moves with selective search of the game tree. This was not full brute force—enumerating every possible course of play. [1]

Its success therefore cannot fairly be explained solely by more memory and computing power. The method also mattered: learning, evaluating positions and guiding search. Large resources enabled many trials, while the algorithm helped select possibilities worth further investigation.

The rules of Go and the goal of winning nevertheless remained fixed. AlphaGo Zero developed strategies within them. It did not change the rules, the winning condition or the task itself. That is the example's limitation: acting more effectively within a framework does not by itself prove an ability to question that framework.

From this article's perspective, it is primarily an achievement in learning and optimisation. Calling a strategy unconventional describes its relationship to previous player practice. It is not automatic proof of creativity in the stronger sense at issue.

3. An objective function is not an idea for a solution

An objective function defines what the system should regard as a better result; in a game, that might be victory. Optimisation searches for a way to achieve it. This can lead to behaviour that is highly unusual from a human perspective without changing the criterion by which the system judges success.

Precision also matters: a human setting the goal does not mean they supplied every later solution. A system can find a strategy its designer did not anticipate. Dependence on a human objective and independence in seeking a way to achieve it are different issues. The origin of an instruction alone therefore does not settle whether the result is creative.

Similar caution is useful for mathematical discoveries. FunSearch, which combines a language model with a program that checks results, found new constructions improving on previous results and new heuristics for a packing problem. These are concrete achievements. They do not by themselves prove that the system independently reformulated the research goal or its basic assumptions. [2]

4. What did this conversation actually establish?

The strongest objection to readily attributing creativity to AI is that skilful processing of learned patterns does not itself demonstrate it. Likewise, more trials, more effective search and better optimisation against an objective function should not automatically be labelled creativity.

The opposite claim—that no AI can ever be creative—also needs caution. It does not follow merely because a system operates algorithmically or learns statistically. Rejecting one argument for creativity is not proof of its impossibility. A categorical conclusion would require a criterion distinguishing routine from creativity that can be consistently applied to specific cases.

The conversation also exposed a weakness in the AI's replies: “creativity” shifted between novelty of output, unconventional strategy and reframing a problem. Without a consistent definition, successive “yes” and “no” answers appeared contradictory even though they often addressed different questions.

The thesis of this part is therefore that a new pattern of play, a better method and a previously unknown result are not enough to attribute creativity to AI when creativity means going beyond an existing framework. The examples show capabilities in learning and optimisation. A stronger claim about creativity requires stronger evidence.

Sources

1. David Silver et al., “Mastering the game of Go without human knowledge”, Nature 550, 354–359 (2017). How AlphaGo Zero works.
2. Bernardino Romera-Paredes et al., “Mathematical discoveries from program search with large language models”, Nature 625, 468–475 (2024; online 2023). Description of FunSearch.