Timnit Gebru Is Calling Attention to the Pitfalls of AI | Hacker News Reader

Timnit Gebru Is Calling Attention to the Pitfalls of AI

wsj.com

15 pointsbookofjoe14 comments

I respect her as a researcher and all but the constant reminder that there are biases in AI is utterly pointless. This is going to go the way of privacy - companies simply do not give a shit. They will pay lip service in some way or another for a little while, but growth and desperation if revenues fall will result in completely dropping any feigned morality and they will attempt anything to survive.

Fortunately we have governments to regulate all this?

My point is it is also non trivial, using privacy again (a much simpler problem IMO) the US utterly failed here. The EU tried but I don’t think we can call their legislation a success either. Biases are present in every statistical model, the idea the government will magically come up with easily verifiable metrics that need to be met is pure fantasy. (If such metrics are even possible) enforcement of something like that is probably impossible anyway due to IP concerns.

>“I didn’t anticipate using chatbots in search engines, which is a bonkers idea that everyone is now racing to do to.”

>‘If the input data is biased, then the output can amplify such biases.’

Didn't bias exist before AI chatbots?

I don't like the idea of censoring a technology because it could show bias without an acknowledgement of the bias present in all information.

Where is a source today of unbiased information?

>I don't like the idea of censoring a technology because it could show bias without an acknowledgement of the bias present in all information.

Ok. Bias is present in all information. I'm not aware that anyone was denying that, but there you go.

Is that acknowledgement sufficient or will you require an explicit disclaimer to that effect in every conversation had about the subject, for the forseeable future?

>Where is a source today of unbiased information?

I don't believe anyone claimed there was one. Do you believe the existence of an unbiased source of information is required before any attempt can be made to correct the bias in any specific source of information?

>Ok. Bias is present in all information. I'm not aware that anyone was denying that, but there you go. >Do you believe the existence of an unbiased source of information is required before any attempt can be made to correct the bias in any specific source of information?

If bias in information must be fixed and all information has bias then all information must be fixed. That may be the case but you cannot escape the bias of the fact checkers so we will end up where we started? How do you fix bias without creating bias?

>If bias in information must be fixed and all information has bias then all information must be fixed.

No. It is possible to focus on one thing at a time, or to address similar problems in tandem. It isn't the case that we must remove bias from all forms of information, or else remove no bias from any form of information. And even if "fixing all information" in general was a worthwhile goal, it doesn't place limitations on the effort to identify or address the specific problem of bias in AI.

> That may be the case but you cannot escape the bias of the fact checkers so we will end up where we started?

No. We won't end up where we started. Not all biases are equal to all others. Bias towards racism doesn't have the same effect as bias against racism, for instance nor does a bias towards science lead to the same place as a bias towards religion.

>How do you fix bias without creating bias?

The context you're ignoring in this conversation is relevant to all of your questions.

It says right there in your quote that amplification is the problem.

I would add, inability for users to critically examine facts to determine bias is also an issue.

There isn't one.

I think the point is that people take it as unbiased.

"GPT hallucinated" "GPT said this was true and it wasn't" "GPT has a political leaning"

People want it to be "unbiased" and "factual" but that doesn't exist. The world is not a math problem.

Fair enough. But isn't this how search engines already work?

It's interesting to note that the premise of DIARI is "[to] foster the development of new AI systems created by and for Black people in the US and Africa. DAIR will prioritize leadership by members of these communities and create spaces where contributors are able to do this work with psychological safety." which explicitly declares a substantive bias itself. Will this also be amplified?

There are groups worried about the possibility that AIs will be racist or offensive. And there are groups worried that AI will recursively self improve and kill everyone. These are not the same, and they don't overlap much at all.

It's much, much easier to tune an AI of less-than-human intelligence until it doesn't say racist things, than it is to tune a superintelligent AI into not killing everyone. And, compared to the risk from possible future superintelligences, there just isn't much scope for a present-day system like ChatGPT to cause real harm. So, as someone coming from the AI Notkilleveryoneism perspective, I see the "what if AI is prejudiced" perspective as a dangerous distraction, I see Gebru and similar researchers as a dangerous distraction: the issues she raises will be solved without too much difficulty, whereas the risk of artificial superintelligence will not.

You have no evidence for the relative difficulty of tuning some hypothetical "superintelligent" AI. You are just just guessing or expressing an article of faith. I see your science fiction nonsense claim as a dangerous distraction from solving real problems that impact real people.

https://archive.ph/W90fE

Reply on news.ycombinator.com