9 ms·

Human-Centered AI: Building Trust, Democracy and Human Rights by Design

71 points by benbreen 8y ago

d--b 8y ago
This is not specific to AI, this should apply to capitalism as a whole. AI is just a tool. It's the value system that needs to be addressed. Data-driven capitalism usually has one objective: maximize profit. Take youtube recommendation engine for instance. I am an engineer and I am really interested in a lot of stuff, physics, arts, economy, you name it. But on youtube I don't get recommendation for Feynman's lectures on Physics, or guided visits of the Louvre collection, I get shit like: "what happens if you throw a ton of dry ice in a pool" or "the sharpest knife made of cardboard" or "10 unforgettable goals". This is pushing addictive crap on me, because the objective function of the recommendation engine literally is: maximize the time people spend watching stuff. Why? Because of ad revenues! What culture needs to curb is capitalism. AI will follow.
- YouAreGreat 8y ago
  > the objective function ... is: maximize the time people spend watching stuff Sure, but the real problem is how well it works. Humans eagerly eat it up because their behavioral heuristics aren't adapted and get gamed. Humans want it and enjoy it and can't get enough, even though it sucks up their time without tangible benefit. Reducing the problem to "capitalism" is like saying the only problem humans have with opioids is that the Sacklers get filthy rich and therefore humanity would be fine if only everybody could get the stuff for free at a friendly neighborhood government drug outlet.
  - rustyboy 8y ago
    That's not the most outlandish idea though is it? Plenty of people argue that if there were more living wage paying jobs, better universal healthcare, and legalized drugs then the opioid crisis wouldn't be as huge, or even exist as it currently does. That being said I agree that > is like saying the only problem capitalism cannot be the only thing to blame, but separating out the correlations versus the side effects of capitalism is a huge academic discussion in it's own right.
- tzahola 8y ago
  >the sharpest knife made of cardboard OMG, so I'm not the only one who gets these idiotic "sharpest knife made from cardboard/pasta/tinfoil/rice/wood" in their recommendations!
- cirgue 8y ago
  > What culture needs to curb is capitalism. I think this is even more specific: culture needs to curb advertising monetization. Google, FB, youtube, et al generally suck because they are rapidly evolving toward convergence with tabloid publishers.
- white-flame 8y ago
  No, what we need to curb is agreeably unethical behaviors, and to delineate basic rights that need to be expressed in light of new technologies (like data privacy rights). This is regardless of any other socioeconomic structures, and ends up being the fundamental basis of complaints like what you raise.
vowelless 8y ago
RE: the idea of "human centered AI" I am quite saddened by this view of "human-centered AI". I read the linked piece by Fei-Fei Li [0], which this talk is essentially based around. The three goals mentioned seem extremely limited and immature. Additionally, I think human-centered AI should not be based around goals, but rather, axioms / laws -- agreed on and debated by humans. The goals are immature because they are extremely broad, ripe for misuse. They are attempting to describe some end result, which, I am not really sure is necessarily a good approach for the concept of "human centered AI". For example: * How does enhance human capability, not replace it protect against discrimination of humans who don't have access to AI from the humans who have "enhanced capabilities" due to "human-centered AI"? If the answer is "get AI capability to all humans", then isn't it extremely important for the first goal to be "get equal opportunity of access to AI capability to all humans"? * Fei-Fei Li says No amount of ingenuity, however, will fully eliminate the threat of job displacement. Addressing this concern is the third goal of human-centered A.I.: ensuring that the development of this technology is guided, at each step, by concern for its effect on humans. Why should human-centered AI not be a precursor to a post-job world for humans? * Donahoe paraphrases Fei-Fei Li's first goal: Goal 1 — making AI more human-like in its intelligence — is essentially a technological task. Does human-centered AI need to have human like intelligence? For now, I am unconvinced. YouAreGreat mentioned: Sure, but the real problem is how well it works. Humans eagerly eat it up because their behavioral heuristics aren't adapted and get gamed. Humans want it and enjoy it and can't get enough, even though it sucks up their time without tangible benefit. And they are absolutely right. Things like slavery also worked really well for a lot of people for thousands of years. Many generated lucrative profits off of that practice and it struck at some core human failing in how we looked at each other. It took deep philosophical works, wars and strict enforcements to alleviate those problems (and we still haven't complete 'solved' it I guess). Maximizing an objective function based on some human desire is a horrible paradigm for human-centered AI, and the three goals listed by Fei-Fei Li don't seem to address this fundamental issue. As an alternative take (maybe others can chime in), I think a human-centered AI should be built around the great philosophical developments of our world. Things like equality of opportunity should be baked in as an axiom for developing human-centered AI (perhaps this should be part of the objective function). No discrimination based on protected classes [1] should be a at the foundation of whatever system the 'human-centered AI' constitutes. (These are examples, maybe there are better core principles). RE: Ethics Donahoe says: > A key theme we will emphasize today is that human-centered AI will require new thinking about democratic accountability for data-driven machine-based governance decisions, as well as richer development of the concepts of algorithmic scrutability and interpretability for governance actors. This is incomprehensible to me. Can someone actually explain what she is talking about here? I am glad she says that the third ethics point is of their primary focus. But I don't see what solutions are provided to manage / enforce that point. Is it left to the benevolence of the "AI engineers"? She paraphrases: > They use slightly different terminology but all revolve around some variation of the concept that AI should incorporate “human values,” reinforce “human dignity, or benefit human beings and humanity. To date, most of these initiatives remain at a relatively high level of abstraction, so it’s hard to know what they might actually require in practice. I'm glad there is some discussion going on about this. I'll have to look into the linked websites at a later date. > In a parallel way, the roots of today’s human-centered AI movement reach back to the Universal Declaration of Human Rights drafted in the aftermath of World War II, and to the body of international human rights law developed in the 70 years since. It is well attested that these human rights declarations have a strong western bias. I am personally fine with that. But what about other societies that don't agree with the declarations? There will need to be international treaties and policies, similar to nuclear policies, that keep everyone on board. But this can get messy, fast. -------------------------------------------------------------------------- [0] https://www.nytimes.com/2018/03/07/opinion/artificial-intelligence-human.html https://www.nytimes.com/2018/03/07/opinion/artificial-intell... [1] Fortunately, this is an issue that is being heavily looked at. A lot of papers at ML conferences tend to focus on this issue.

71 points by benbreen 8y ago

d--b 8y ago

This is not specific to AI, this should apply to capitalism as a whole. AI is just a tool. It's the value system that needs to be addressed. Data-driven capitalism usually has one objective: maximize profit. Take youtube recommendation engine for instance. I am an engineer and I am really interested in a lot of stuff, physics, arts, economy, you name it. But on youtube I don't get recommendation for Feynman's lectures on Physics, or guided visits of the Louvre collection, I get shit like: "what happens if you throw a ton of dry ice in a pool" or "the sharpest knife made of cardboard" or "10 unforgettable goals". This is pushing addictive crap on me, because the objective function of the recommendation engine literally is: maximize the time people spend watching stuff. Why? Because of ad revenues! What culture needs to curb is capitalism. AI will follow.

YouAreGreat 8y ago
> the objective function ... is: maximize the time people spend watching stuff Sure, but the real problem is how well it works. Humans eagerly eat it up because their behavioral heuristics aren't adapted and get gamed. Humans want it and enjoy it and can't get enough, even though it sucks up their time without tangible benefit. Reducing the problem to "capitalism" is like saying the only problem humans have with opioids is that the Sacklers get filthy rich and therefore humanity would be fine if only everybody could get the stuff for free at a friendly neighborhood government drug outlet.
- rustyboy 8y ago
  That's not the most outlandish idea though is it? Plenty of people argue that if there were more living wage paying jobs, better universal healthcare, and legalized drugs then the opioid crisis wouldn't be as huge, or even exist as it currently does. That being said I agree that > is like saying the only problem capitalism cannot be the only thing to blame, but separating out the correlations versus the side effects of capitalism is a huge academic discussion in it's own right.
tzahola 8y ago
>the sharpest knife made of cardboard OMG, so I'm not the only one who gets these idiotic "sharpest knife made from cardboard/pasta/tinfoil/rice/wood" in their recommendations!
cirgue 8y ago
> What culture needs to curb is capitalism. I think this is even more specific: culture needs to curb advertising monetization. Google, FB, youtube, et al generally suck because they are rapidly evolving toward convergence with tabloid publishers.
white-flame 8y ago
No, what we need to curb is agreeably unethical behaviors, and to delineate basic rights that need to be expressed in light of new technologies (like data privacy rights). This is regardless of any other socioeconomic structures, and ends up being the fundamental basis of complaints like what you raise.

vowelless 8y ago

RE: the idea of "human centered AI" I am quite saddened by this view of "human-centered AI". I read the linked piece by Fei-Fei Li [0], which this talk is essentially based around. The three goals mentioned seem extremely limited and immature. Additionally, I think human-centered AI should not be based around goals, but rather, axioms / laws -- agreed on and debated by humans. The goals are immature because they are extremely broad, ripe for misuse. They are attempting to describe some end result, which, I am not really sure is necessarily a good approach for the concept of "human centered AI". For example: * How does enhance human capability, not replace it protect against discrimination of humans who don't have access to AI from the humans who have "enhanced capabilities" due to "human-centered AI"? If the answer is "get AI capability to all humans", then isn't it extremely important for the first goal to be "get equal opportunity of access to AI capability to all humans"? * Fei-Fei Li says No amount of ingenuity, however, will fully eliminate the threat of job displacement. Addressing this concern is the third goal of human-centered A.I.: ensuring that the development of this technology is guided, at each step, by concern for its effect on humans. Why should human-centered AI not be a precursor to a post-job world for humans? * Donahoe paraphrases Fei-Fei Li's first goal: Goal 1 — making AI more human-like in its intelligence — is essentially a technological task. Does human-centered AI need to have human like intelligence? For now, I am unconvinced. YouAreGreat mentioned: Sure, but the real problem is how well it works. Humans eagerly eat it up because their behavioral heuristics aren't adapted and get gamed. Humans want it and enjoy it and can't get enough, even though it sucks up their time without tangible benefit. And they are absolutely right. Things like slavery also worked really well for a lot of people for thousands of years. Many generated lucrative profits off of that practice and it struck at some core human failing in how we looked at each other. It took deep philosophical works, wars and strict enforcements to alleviate those problems (and we still haven't complete 'solved' it I guess). Maximizing an objective function based on some human desire is a horrible paradigm for human-centered AI, and the three goals listed by Fei-Fei Li don't seem to address this fundamental issue. As an alternative take (maybe others can chime in), I think a human-centered AI should be built around the great philosophical developments of our world. Things like equality of opportunity should be baked in as an axiom for developing human-centered AI (perhaps this should be part of the objective function). No discrimination based on protected classes [1] should be a at the foundation of whatever system the 'human-centered AI' constitutes. (These are examples, maybe there are better core principles). RE: Ethics Donahoe says: > A key theme we will emphasize today is that human-centered AI will require new thinking about democratic accountability for data-driven machine-based governance decisions, as well as richer development of the concepts of algorithmic scrutability and interpretability for governance actors. This is incomprehensible to me. Can someone actually explain what she is talking about here? I am glad she says that the third ethics point is of their primary focus. But I don't see what solutions are provided to manage / enforce that point. Is it left to the benevolence of the "AI engineers"? She paraphrases: > They use slightly different terminology but all revolve around some variation of the concept that AI should incorporate “human values,” reinforce “human dignity, or benefit human beings and humanity. To date, most of these initiatives remain at a relatively high level of abstraction, so it’s hard to know what they might actually require in practice. I'm glad there is some discussion going on about this. I'll have to look into the linked websites at a later date. > In a parallel way, the roots of today’s human-centered AI movement reach back to the Universal Declaration of Human Rights drafted in the aftermath of World War II, and to the body of international human rights law developed in the 70 years since. It is well attested that these human rights declarations have a strong western bias. I am personally fine with that. But what about other societies that don't agree with the declarations? There will need to be international treaties and policies, similar to nuclear policies, that keep everyone on board. But this can get messy, fast. -------------------------------------------------------------------------- [0] https://www.nytimes.com/2018/03/07/opinion/artificial-intelligence-human.html https://www.nytimes.com/2018/03/07/opinion/artificial-intell... [1] Fortunately, this is an issue that is being heavily looked at. A lot of papers at ML conferences tend to focus on this issue.