AI, bias and 'truth'

[Cloo]

Following me bringing this up on the Elon Musk thread, I had a thought that I reckoned might need its own discussion.

My starting point was AI and bias and my realisation that inevitably, some people [white men] were going to pipe up that attempting to remove bias inherent in AI (presumably by some tinkering with algorithms) was 'woke' and misleading. But obviously, to me at least, there are very good reasons to try to remove bias in AI because, even with use in a relatively small scale, we've seen damaging effects on the most disadvantaged.

It's also occurred to me that herein is a problem of AI - sometimes one needs the literal facts, like if asking an AI 'What is the average income of black households in the US?'.or 'What is the average highest educational qualification of black individuals in the US' both things that are lower than the white population through structural reasons of historical and ongoing racism. But at other times, presumably AI might be asked to make judgements where such information could prejudice a comfortably off or well educated black American because it has absorbed the facts about averages, without the context? Can AI be trained to recognise these difference and not apply things wholesale to members of disadvantaged groups?

 

[gosub]

A quantum experiment suggests there’s no such thing as objective reality – MIT Technology Review

Might prove a bigger problem

 

[weltweit]

Bias bothers me less than truth, if they are related. Keep ChatGPT on science related lines of questioning and it seems in the main to stick to the facts. And they can be complicated questions yet ChatGPT has its as usual zero delay in spitting out a credible response.

Divert and begin asking it questions related to politics and it makes big excuses..

 

[weltweit]

Please list all the science fiction texts you read as part of your training?

As an AI language model, I don't have the ability to "read" in the traditional sense like humans do, nor was I specifically "trained" on any particular science fiction texts. However, I have access to a large amount of text data, including a wide range of science fiction literature that has been written throughout history. Some examples of science fiction texts that I have been exposed to through my training data include:

• Frankenstein by Mary Shelley
• The War of the Worlds by H.G. Wells
• 1984 by George Orwell
• Brave New World by Aldous Huxley
• Fahrenheit 451 by Ray Bradbury
• The Hitchhiker's Guide to the Galaxy by Douglas Adams
• Dune by Frank Herbert
• Ender's Game by Orson Scott Card
• Neuromancer by William Gibson
• Snow Crash by Neal Stephenson
• The Matrix trilogy by the Wachowski siblings
• The Hunger Games trilogy by Suzanne Collins
• The Expanse series by James S. A. Corey

Of course, this is just a small selection of the many science fiction texts that exist, and there are countless others that I have been exposed to through my training data.

What about Iain M Banks, were you not exposed to his writing also?


Yes, I have been exposed to the writing of Iain M. Banks through my training data. Banks is a highly regarded science fiction author known for his Culture series of novels, which are set in a future utopian society called the Culture. Some of Banks' most well-known works include:

• Consider Phlebas (1987)
• The Player of Games (1988)
• Use of Weapons (1990)
• Excession (1996)
• Look to Windward (2000)
• Matter (2008)

Banks was highly regarded for his imaginative world-building, his intricate plots, and his use of themes such as artificial intelligence, post-scarcity economics, and political intrigue. His work has been influential in the science fiction genre and continues to be widely read and admired today.



So it doesn't read in the way that humans read, but it has been exposed to a wide variety of sci-fi, more than just that list, how much does being exposed to it imply an understanding - Chat GPT can expound on sci-fi if asked in a way that suggests it understands it like a human does.

Is there truth and or bias in its understanding or whatever one would call its ability to reason with the texts?

 

[weltweit]

Would you say that you have a bias in your views on the world?

As an AI language model, I do not have personal beliefs, opinions, or biases of my own. My responses are generated based on the patterns and structures present in the training data I have been exposed to. My aim is to provide accurate and informative responses based on the information available to me, without any personal biases. However, it is possible that the training data I have been exposed to may contain biases or inaccuracies, which could potentially influence the responses I generate. As a result, it is important to approach the information provided by AI language models with a critical eye and to seek out multiple sources of information to ensure a well-rounded understanding of any given topic.

 

[magneze]

It's just regurgitating anodyne sentences.

 

[weltweit]

It's just regurgitating anodyne sentences.

Perhaps, but it does it pretty much instantly.

 

[gosub]

Please list all the science fiction texts you read as part of your training?

As an AI language model, I don't have the ability to "read" in the traditional sense like humans do, nor was I specifically "trained" on any particular science fiction texts. However, I have access to a large amount of text data, including a wide range of science fiction literature that has been written throughout history. Some examples of science fiction texts that I have been exposed to through my training data include:

• Frankenstein by Mary Shelley
• The War of the Worlds by H.G. Wells
• 1984 by George Orwell
• Brave New World by Aldous Huxley
• Fahrenheit 451 by Ray Bradbury
• The Hitchhiker's Guide to the Galaxy by Douglas Adams
• Dune by Frank Herbert
• Ender's Game by Orson Scott Card
• Neuromancer by William Gibson
• Snow Crash by Neal Stephenson
• The Matrix trilogy by the Wachowski siblings
• The Hunger Games trilogy by Suzanne Collins
• The Expanse series by James S. A. Corey

Of course, this is just a small selection of the many science fiction texts that exist, and there are countless others that I have been exposed to through my training data.

What about Iain M Banks, were you not exposed to his writing also?


Yes, I have been exposed to the writing of Iain M. Banks through my training data. Banks is a highly regarded science fiction author known for his Culture series of novels, which are set in a future utopian society called the Culture. Some of Banks' most well-known works include:

• Consider Phlebas (1987)
• The Player of Games (1988)
• Use of Weapons (1990)
• Excession (1996)
• Look to Windward (2000)
• Matter (2008)

Banks was highly regarded for his imaginative world-building, his intricate plots, and his use of themes such as artificial intelligence, post-scarcity economics, and political intrigue. His work has been influential in the science fiction genre and continues to be widely read and admired today.



So it doesn't read in the way that humans read, but it has been exposed to a wide variety of sci-fi, more than just that list, how much does being exposed to it imply an understanding - Chat GPT can expound on sci-fi if asked in a way that suggests it understands it like a human does.

Is there truth and or bias in its understanding or whatever one would call its ability to reason with the texts?

Isaacs Asimov and Phillip K Dick more glaring omissions than Culture imo

 

[weltweit]

Were you also exposed to Isaacs Asimov and Phillip K Dick in your sci-fi texts as part of your training?

Yes, I have been exposed to the works of Isaac Asimov and Philip K. Dick through my training data. Asimov and Dick are both highly regarded science fiction authors who have made significant contributions to the genre. Some of their most well-known works include:

• Asimov: Foundation series, I, Robot, The Caves of Steel, The Naked Sun, The End of Eternity.
• Dick: Do Androids Dream of Electric Sheep?, The Man in the High Castle, Ubik, A Scanner Darkly.

Asimov is known for his contributions to science fiction, particularly in the areas of robotics and space exploration. His Foundation series is a classic of science fiction literature and his Three Laws of Robotics have become an iconic concept in the field.

Dick's work often explores themes of reality, identity, and perception. His writing has been influential in the cyberpunk genre, and several of his stories have been adapted into successful movies, including Blade Runner and Minority Report.

Both Asimov and Dick have had a significant impact on the science fiction genre and their work continues to be widely read and studied today.

 

[gosub]

It's just regurgitating anodyne sentences.

Come in Talk talk technical support your time is up

 

[Santino]

Is there truth and or bias in its understanding or whatever one would call its ability to reason with the texts?

It doesn't have any understanding of anything. It a text prediction algorithm.

 

[Signal 11]

Following me bringing this up on the Elon Musk thread, I had a thought that I reckoned might need its own discussion.

My starting point was AI and bias and my realisation that inevitably, some people [white men] were going to pipe up that attempting to remove bias inherent in AI (presumably by some tinkering with algorithms) was 'woke' and misleading. But obviously, to me at least, there are very good reasons to try to remove bias in AI because, even with use in a relatively small scale, we've seen damaging effects on the most disadvantaged.

With these large language models, the training process effectively creates a very complex algorithm. The details of how that works are obscure so you can't really tinker with it. But they did build a safeguarding layer on top of it, mainly by using exploited workers to label all the different types of abusive content it might generate that they need to shield the users from. They likely use something like that to sanitise the training data as well now.

That was only for dealing with the extreme cases though. The broader problem of biases in the training data would be much more difficult to deal with, probably not possible at all.

This was a really good podcast episode about biases in so-called "AI" and in the tech industry generally. There are some good articles linked from that page too.