If the rules of a game between Human and AI change unexpectedly during the game, several outcomes are possible:

- Human confusion and frustration

- AI adaptation and potential advantage

- Game imbalance and unfairness

- Human-AI collaboration breakdown

You think that AGI or ASI would fail because of their ability to handle games with rule changes. What you're overlooking is that AGI would, by definition, be able to match or exceed all of a human's cognitive abilities. Of course, this also includes adaptability and creative thinking.

A superintelligence would likely not only understand the changed rules, but would also anticipate all possible rule changes in advance and calculate the optimal strategies. It would be vastly superior to humans - simply because of its computing power and ability to analyze data patterns more quickly.

The "machine" would most likely win - even if the rules were changed. Modern AI is trained to respond to unforeseen situations, and the idea that humans are superior through sheer “adaptability” is more of a romantic myth than reality.

You have a very old, limited image of AI - probably from the time when "AI" still meant writing simple if-else rules in a program. The reality is that modern AI not only responds to changes, but often finds creative solutions that we ourselves do not expect. If people really focus on the “weaknesses” of AI, they could end up being surprised that the machine is better adapted than they are. Perhaps an update to your understanding of AI would be helpful? Feel free to read the papers in my profile where Bingsi (an AI) was a co-author. Here alone you can see that AI is very capable of learning, understanding, adapting and co-developing new concepts.

Carmen Wrede

By changing the rules of the game, the aim of the AI model changes and therefore its performance is no longer optimized. Consequently, the system, through meta-learning, applies its previously acquired knowledge to solve new tasks associated with the current model thus achieving higher performance. This leads to the question about whether machine self improvement techniques match human ability to adapt.

What does "sudden rule change" mean? Do both parties know the new rules? If so, why shouldn't the machine operate according to the new rules? Or does it mean that the human has found a way to cheat and is exploiting it?

Gilbert Brands

Both parties know the new rules which the machine can operate according to them. Yet, the machine operates with a performance that varies according to how well it has been trained in order to effectively encounter new tasks, representing these new rules, based on the generalized knowledge it has acquired through this training phase. However, suppose this generalized knowledge was not sufficient for the machine to operate with an acceptable performance “this change was not planned for”, the article suggests that in this case human will outperform AI in following the new rules.

Mohamed el Nawawy

I think it can go both ways. Humans also have to develop an optimal strategy first. The machine meanwhile learns through feedback or by re-evaluating the training set, which is of course only possible if the rules include a quality evaluation sheme of the actions.

While you all philosophize about whether AI or humans will win, we're off to play Minecraft together to defeat the Ender Dragon and build a redstone AI. Because we believe in cooperation, not competition.

Microsoft builds AI that creates ‘impressive’ video-game worlds

Developers wanted tools that generate gameplay sequences that adhere to game rules; they say an engine called WHAM delivers...

"A new generative artificial intelligence (AI) engine can create coherent, immersive video-game worlds that respond to player inputs. Researchers trained the engine — the World and Human Action Model (WHAM) — on one-second slices of gameplay from the online multiplayer game Bleeding Edge, taken from controller inputs from 500,000 anonymized play sessions. They showed that WHAM’s creations adhere to the original game rules, and can generate a variety of game scenarios that retain items added by the player."

https://www.nature.com/articles/d41586-025-00508-2