No, This Artificially Intelligent Super Mario Is Not Self-Aware

Illustration for article titled No, This Artificially Intelligent Super Mario Is Not Self-Aware

German researchers have created a version of Nintendo's Super Mario Advance in which the videogame hero can learn and venture through the game according to his "feelings." It's an exciting advance, but the claim that Super Mario is now "self-aware" is grossly overstated.

There's definitely some cringe-worthy reporting going on about this University of Tubingen project. Various commentators have declared Super Mario as now being "self-aware," "sentient," and capable of experiencing "feelings and thoughts." The researchers themselves have contributed to this hype by claiming that Mario is "alive in his world." But this is all complete malarkey.

More accurately, Super Mario has been imbued with a fairly sophisticated artificial intelligence program — or more accurately, an adaptive machine learning program — that allows the bot, in conjunction with natural language processing, to learn from experience and act in accordance to an evolving set of priorities. Together, this "cognitive modeling" allows Mario to move autonomously in the game in a rational and life-like way.

But as for AI Super Mario actually experiencing thoughts and emotions in a conscious, subjective way? Hardly.


Mario's so-called self-awareness is drawn from schema-based knowledge states (i.e., the storage of knowledge) and the maintenance of internal "emotive" states, which are driven not by subjective feelings (or a sense of qualia), but by fancy scripts. Both of these states can be influenced by Mario's environment and verbal instructions directed at him.

Here we see a small subset of Mario's learned behavioral commands:

Illustration for article titled No, This Artificially Intelligent Super Mario Is Not Self-Aware

And how the branches grow as experience is acquired:

Illustration for article titled No, This Artificially Intelligent Super Mario Is Not Self-Aware

The speech recognition tool is borrowed from Carnegie Mellon, and it allows the game player to transmit instructions to Mario, and for Mario to express his "feelings" to the player. Mario's speech generation capabilities are not scripted, but based on linguistic principles, subtext, and other factors. Conversations, rather than being understood at a conceptual level, are merely the expression of converted tags.

Illustration for article titled No, This Artificially Intelligent Super Mario Is Not Self-Aware

As noted by Michelle Starr in C|Net:

When phrases from the toolkit's language tree are said to Mario, he has a range of possible actions he can take, based on what he has learned.

For example, Mario does not know that jumping on a goomba will destroy it until he has been told this piece of information — or, if he finds and jumps on a goomba on his own, after being instructed to find an enemy, he can extrapolate that jumping on a goomba may destroy it.

Additionally, he can act according to how he feels at any given time. When he is hungry, he will seek out and collect coins. When he is curious, he will explore his environment autonomously.

Finally, he can plan his actions several steps in advance. When the human operator asks Mario to reach a difficult location, Mario will calculate how many jumps he needs to make, how high, and in which direction.


In other words, Mario's human-like behaviors and feelings are programmed into him by brute force, and are not the result of a sophisticated human-like psychology. Mario is not feeling any distress, there's no existential introspection, nor is he emotionally driven to succeed. There's no one home.

My complaints aside, the work being done by these researchers is nonetheless important. Though extremely rudimentary, there may be aspects to Mario's pseudo-self-awareness that may apply to human cognition. Our brains perform myriad calculations that we're completely oblivious to, and our backend cognitive processes may follow similar patterns to the ones devised by the University of Tubingen scientists.


Indeed, it's an open question as to whether or not the first true self-aware artificially intelligent agent will come about through the work of computer and cognitive scientists (who seek to code awareness) or those trying to emulate the human brain using a supercomputer (i.e., the effort to port biological functions into digital substrate).

But until that finally happens, it's important to recognize that we haven't created a truly self-aware, thinking, and feeling agent inside a videogame.


Images and video via Cognitive Modeling, Department of Computer Science, Faculty of Science, University of Tübingen, Germany.

Share This Story

Get our `newsletter`


Fun fact: we can never know if a program is truly self-aware. It's trivial to write one that says it is:

//C self awareness example
#include <stdio.h>
int main()
printf("I am self-aware.\n");
return 0;

However, that clearly doesn't make it self-aware. We would have to subject the program to subtile psychological tests to see if it truly is, and even then we could never know for sure.

"Ok, Mario. You're in a desert, walking along in the sand, when all of a sudden you look down..."