The only banana problem is questioning the blind reliability of artificial intelligence

The only banana problem is questioning the blind reliability of artificial intelligence

He. She

Edited by Tech Innovation Website – 06/29/2023

Requests to draw one banana resulted in two bananas across all of the pictures.
[Imagem: Digital Science/Daniel Hook]

But I just want a banana!

Even a kid can draw a banana, but ask an AI program to do it and the results might surprise you!

At least Professor Dr. Daniel Hooke, who left the University of Cambridge in the UK to lead his own research company, was surprised. Digital sciencewhich aims to use artificial intelligence to conduct collaborative research.

He took the most popular image-generating AI program and then gave them a simple request: “Draw one banana by casting a shadow on a gray background.”

To the researcher’s surprise, instead of a banana, the AI ​​pulled out two bananas – and repeated the two bananas each time, even rearranging them in different ways.

Hawk continued the experiment over a period of weeks, with the AI ​​ultimately failing. This is important to him because he wants his company to use this technology in their products.

Researcher Hook recently published the results of his technical AI experiment in a paper titled “The Lonely Banana Problem.”

Solitary banana problem

This was the result of the “Draw an invisible monkey holding one banana” request.
[Imagem: Digital Science/Daniel Hook]

Solitary banana problem

In the initial attempt, using an artificial intelligence program called Midjourney, the researcher got four results (the first image in this article), all showing two bananas, not one.

Believing he had made a mistake, Hook experimented with different requests, such as: “a fully ripe banana on a solid gray background casting light shadows, hyper-realistic”, or “one fully ripe banana standing alone on a solid gray background casting light shadows, a hyper-realistic depiction” and “A perfect banana standing alone on a uniform light gray surface, taken from above, hyper-realistic photography.” All of these orders also yielded two or more bananas.

See also  The Ministry of Health begins 25 vaccination campaigns against influenza, starting from Monday (10).

And it could get worse: When the researcher asked for an “invisible monkey with one banana,” the results yielded very visible monkeys holding two or more bananas.

All the pictures I asked Midjourney to produce contained two (or more) bananas, apparently no matter which way I asked. The results were the same, with aliens and other animals holding bananas. Somewhat strangely, many monkeys They ended up wearing bananas, Hook said.

Solitary banana problem

These were the best results, after the researcher tried to adapt to the maximum level of artificial intelligence. But even here, one of the photos featured two bananas.
[Imagem: Digital Science/Daniel Hook]

Why does AI error matter?

the doctor. Hook says that the file Solitary banana problem He points out that “although this technology is developing rapidly and its results are impressive, there are still gaps, and while they are not always immediately noticeable, they are not small.”

“AI systems at their current level of development do not perceive things the way we do — they understand commonly occurring patterns. And their reality is fundamentally different from ours — they are not born into the physical world, but into a logical one. Certainly, as successive As generations of AI develop, it becomes easier for us to interact with them which suggests they truly understand… However, without an understanding of the physical world, AI has trouble with the concept of a single banana.”

Knowing that AI doesn’t understand the real world or people the way we think it is, the researcher argues, is “an important step to taking responsibility for building on these technologies as we create new tools.”

See also  Braga Hospital encourages theater workshops for epilepsy patients

“[…] We believe we have a responsibility to ensure that the technologies we launch are well tested and well understood. The use cases in which we deploy AI should be appropriate to the level at which we know AI can operate, and any functionality should come with a “health warning” so that people know what they need to look for – when they can trust AI, the researcher concluded. That AI and when it shouldn’t.

Follow the Technological Innovation website on Google News

Other news about:

More topics

By Andrea Hargraves

"Wannabe internet buff. Future teen idol. Hardcore zombie guru. Gamer. Avid creator. Entrepreneur. Bacon ninja."