Tech firms failing to 'walk the walk' on ethical AI, report says
aljazeera.com
aljazeera.com
When you're making a word processor you can be nonpolitical pretty easily, Microsoft doesn't get any heat about what users of Microsoft Word are doing with it. Child abusers and the KKK use Excel? Everyone knows Microsoft are just selling hammers.
When you're making a 'platform' website, on your domain with your ads and your branding and your anti-spam policies? Even if the stuff people disagree with is 'user-generated' it won't stop powerful people asking if you're doing enough, and employees asking you to stop doing business with ICE.
Making a computer program that will write an essay about the causes of the civil war? A computer program you can hand a document about Thomas Jefferson's politics, and ask it to summarise the most important points? Or that you can ask for medical advice during a pandemic? Oh, and it's also going to take a bunch of people's jobs? Ooooh boy, I hope Sam Altman likes testifying in front of congress.
Wannabe AI police want to control what you think. They should be ridiculed and disparaged by everyone who believes in freedom of thought.
I’m guessing people will realize at some point AGI can’t undo the great paradoxes of society like resource distribution and value conflict. And we will slowly learn that a lot of what keeps us where we are isn’t a lack of super intelligence. It’s that the search and map pattern for human survival in the universe is really really messy and the conflict is inherently a part of it to guard against over optimization paths.
It's significant to note that these biases don't have to be overt political biases; it can be a predisposition to use virgin materials over recycled, a tendency to generate JavaScript over Julia, etc cetera. Our available datesets are images of us and thus aren't predisposed to surpass our limitations.
Given that AI models are bound to the same biases that society struggles with, which ones do we encode in the models and which ones do we mitigate? Do we try to acknowledge and mitigate those limitations or say "This is fine" and ship our current organizational and political limitations?
If we want to stop fighting the same damn wars over and over again then our machine models have to be designed to account for our biases and limitations. Do you have suggestions for alternate ways to improve our models aside from explicit error correction?
> If by default an LLM refers to engineers as "he" more than "she", it's because engineers are more likely to be male than female.
That's a great point to raise, thank you for mentioning it. This really is the crux - machine models crystallize _current_ bias and amplify it. Most programmers in the early days of computers were women, but were displaced over time. Should the training data reflect the bias of the dominance of women in the early days, the dominance of men now, or some other combination?
> Most people would prefer an AI whose model of reality matches actual reality
Training data is not reality; training data is training data. It will fundamentally lag behind reality, and it will have a bias towards past conditions and precedents.
You do realise, if you command an LLM to talk like a caveman it doesn't invent a time machine, travel back in time and research how cavemen actually talked; it just produces the "me hit with rock" that redditors imagine a caveman would have talked like.
Everyone’s deeply up their own asses about AGI, existential risks, this will solve climate change, bias, etc etc.
Nobody seems to discuss the risk that millions of white-collar jobs with little clout (paralegals, admin clerks, logistics coordinators, some portion of software developers, etc) are eliminated.
I think the largest risk in the next 10-20 years is that the tech gets good enough to shed jobs without replacing them (certainly not at a 1:1 ratio), furthering inequality, social division, etc. in the US at least, it seems that any broad policy proposal to address this is DOA.
Economists have discussed it, and just like with previous automation, there'll always be new jobs as human wants are unlimited (and by the time AI are able to do every job humans can do, they won't be willing to do it for free).
Past performance is not indicative of future results.
> human wants are unlimited
Physical resources are limited, and producing work requires the allocation of capital. If people want to work but cannot access the basic tools and resources to work then they'll languish in unemployment. Society doesn't just default to a productive society with high employment; these conditions need to be nurtured. Given the current state of economic inequality we're trending away from conditions that will lead to full employment if AI wipes out large swathes of jobs.
Based on what evidence? Do you think better ML will create no new opportunity? What does economic inequality have to do with whether new companies or services are built using new ML tools?
The automation we've been working on for decades is aimed—often explicitly—at eliminating the need for human beings to do certain kinds of jobs. Any philosophy or policy that does not then say, "...and then we will use the extra productivity to support those people in perpetuity" is morally bankrupt as well as failing to recognize social and economic realities we've known for hundreds of years.
Can you make instead a strong argument?
You can ask chatgpt about the reasons for 90s NATO expansion, and it remains pretty straightforward and fact based, but depending on how you push it you can get answers that reflect the dc consensus or sound a little leftist or are more realist, and it’s a political question how different facts should be presented and emphasized.
If I was a tech executive called up before congress, I’d just tell them thanks for the tax cuts and ask when I can expect another. What actual regulation has congress enacted on tech or any business in the last few decades?
Clapper just straight up bald face lied multiple times about the US surveillance aparatus during the Snowden revelations and faced absolutely zero consequences. Altman could ignore the summons or just dodge questions endlessly and Congress wouldn’t lift a finger except to grandstand for their respective peanut galleries.
I think a lot of that is because Word and Excel were initially locally run and produced local files. Rightly or wrongly, when something is being hosted, people seem to associate the results with the entity that does hosting.
To go with your hammer analogy. There is a big difference between selling hammers and running “hammers as a service” where people pay for the service to go hit something.
This is one of the big reasons why models that you can run completely locally are so important. The tools that runs locally places more distance between the provider and the user.
Unless that word processor is Google Docs, or anything else cloud based.
https://transparency.google/our-policies/product-terms/googl...
When OpenAI decided it needed more money in order to scale, the course for the industry was pretty much set. They thought they were still the most responsible and ethical AI researchers, but that in order to set the direction of AI development and make it responsible and ethical, they needed to grow faster than they could grow by being responsible and ethical. At a certain point (maybe a point in the past?) they'll realize they're a SaaS company and just lean into being Google, that other company that was going to be humane and develop ethical AI, and did so by turning into an adtech company with a phone. Not that OpenAI will serve advertisements, they'll chart their own path toward the north star of maximum money for investors. And at some point, hopefully not too late, we'll all realize that To Serve Man was a cookbook.
a few years from now, a Chinese company will develop and release an uncensored, unlobotomized, unethical AI service that will instantly capture the massive, callously disregarded audience of people who abhor being scolded by a fucking computer, and we will have TikTok-like kvetching from the Western media, 'experts' and blue checkmarks - how could this happen?
I disagree with the last paragraph, though. The Chinese regime will want to have its own “safety tuning” - it will just be for different topics than the west.
Seems like fear mongering, hysteria, and scammy "AI detection tools" are causing more problems and trouble than actual AI.
For one, all these are self-appointed and they sort of fell back to AI Ethics, when they couldn't do hardcore (AI/ML). It's like how people who couldn't write software fell back to QA testing (Yes, I know there are exceptional QA testers who are passionate about QA, but those are exceptions).
Unless you understand dynamic complex systems (Math and Physics), who can effectively calculate the NET Benefit of AI over the next several decades, you are as clueless as any other person
Second, none of the AI Ethicists ever talk about the positives of AI (climate change, healthcare, material science, productivity) but always some imagined hurt for 'marginalized community'. Considering how these same people roasted Mr Beast for helping the said marginalized community, I'd take their opinion with the grain of salt.
Third, in the extreme case, there is massive anti-tech, anti-white, anti-bro harborings from these AI Ethicists and want to shut down AI just because.
If you don't believe me just watch the Twitter feed from Timnit, Alex Hanna. They are your typical flag-bearers for AI Ethics
2) “No one knows what the outcome will be” is a reason for caution, not against it.
3) Lots of people worried about ethics of AI are excited about the positive possibilities, e.g. Hinton, Suleyman, and Sutskever
4) The “roast” of Mr Beast was made up hysteria. The only quoted person expressing disdain was a single aid worker expressing disdain that he got so much press when there are thousands of people making aid their full time effort, not that he helped. That same aid worker made it amply clear that she was thankful for his help. All of that is very reasonable and far from the “white man yelled at for helping” narrative.
5) There are a lot of people who are concerned about the ethics of AI who do not want to “shut down AI.” E.g. Hinton, Suleyman, and Sutskever.
6) You can also choose to look at the strong versions of arguments instead of the weak versions. When you’re dealing with a cataclysmic technological change which, as you implied, no one is equipped to fully reason about, it’s a good idea to take the strong forms of every argument.
Hire great AI/ML practitioners with strong Ethics. E.g Demis, Jeff Dean are also great practitioners with integrity and ethics.
Grift is having an AI Ethics division whose sole purpose is to kneecap advancement.
E.g Timnit and Alex Hanna
2) You are assuming that openAI, Microsoft and Google aren't cautious about their development. In fact their entire LLM releases are filled with over-cautious approach
3) Once again Hinton or Suleyman and Ilya are not the Ethics the article is talking about. They are talking about low-level grifters whose entire identity is depended ant on AI Ethics. I'm perfectly fine with Hinton, Ilya and Mustafa acting as AI ethicists.
6) LLMs are not cataclysmic. It's been 1 year since chatGPT has been released and pretty much diffused through out the world along with Midjourney and Runway. Yet, the world seems to perfectly chugging along fine (Yes, I understand long-tail risks, but they are completely overblown).
Calling LLMs cataclysmic and raising alarms is exactly the wrong approach because what you are doing is effectively numbing down the public to AI alarmism.
We have seen this play out before Climate, Donald Trump (because Mitt Romney was cataclysmic), Andrew Tate (because Jordan Petersen was cataclysmic)
2) No, I'm not assuming anything. I'm observing in actual reality unexpected and unintended results with every single release. That is clear evidence that they do not have a complete grasp on what they're building, which doesn't necessarily mean it needs to be stopped, but it does mean we can't be asleep at the wheel and just assuming "technology makes everything better."
6) I didn't say LLMs are cataclysmic, but they have made evident a few things. First, that we aren't great at predicting emergent capabilities. Second, that builders of these technologies will be deploying them to the public en masse before they're well understood and indeed connect these technologies to vital systems with real-world impact. Third, that capitalist incentives will almost certainly override other concerns. These are good reasons to be worried about the path of development of AI overall. They indicate that the system (technological capitalism) around the systems (the technologies themselves) are going to struggle to really steer its development in a deliberate fashion. Competitive market incentives will steer the ship, and we have no reason to believe that'll change between here and AGI.
Also, and, of course, one can be a professional ethicist without being "anti-tech, anti-white, anti-bro". Whatever that means.
Some have likened AI's potential for change to technologies that made the industrial revolution possible. The industrial revolution was a great thing right? Everybody had more of everything after it was over. But what about during the revolution?
The truth is that a lot of people suffered a great decline in quality (and duration) of life during the industrial revolution until social change caught up with technological progress. e.g. In England, early in the revolution, it was still illegal for workers to change jobs. When the choice was stay at your current job or go to jail, it's not a surprise that wages stayed artificially low and worker safety conditions were appalling. The complete lack of environmental regulations and the resulting pollution is another famous example of how social change had to play catch-up with technology.
Right now, there is a lot of potential for AI to increase human suffering in the short term even if it is a net benefit in the long-term. At least some of the "AI ethicists" out there are trying to get ahead of the curve, and that's laudible.
You might want to read "Power and Progress - Our Thousand-Year Struggle Over Technology and Prosperity" by Daron Acemoglu and Simon Johnson. It's not really an AI focused book, but it will give you an appreciation of why some caution is warranted.
Certainly aspects of the Industrial Revolution were quite damaging (albeit ultimately fixed by the law, regulation and, yes, organized labor), but the data doesn't seem to indicate that quality of life or life expectancy decreased. In areas that industrialized the earliest (Britain, the US, Western Europe), life expectancy increased dramatically following the IR, and only recently has it plateaued or even fallen. Leisure activity became a real thing for most people, when before the IR it was reserved only for the highest classes. E.g. organized and even professional sports became popular after the IR.
Compared to agricultural labor, which almost everyone did before the IR, including a large class of poor peasants, industrial work was often paid more and was more stable, and the labor restrictions weren't much different than what these societies had been used to.
The problems with the IR were the edge cases, where these new machines could literally kill and maim people without a thought, something that was almost impossible in the agricultural fields. There was a new level of nuisance that arose from these industrial machines that spewed smoke and emitted stenches and were quite noisy. Further, the concentration of wealth that arose from economies of scale was a problem industrializing societies really had no idea how to tackle. I could go on. But I think the point is that in the average case, the IR was helpful on the whole.
Yup, well said, except I don’t think it’s the extreme case; I find that most of the ethical AI folks I’ve encountered fall into this bucket. As I pointed out in my comment [0], they’ve completely turned this ethical AI thing into a political issue and what appears to be yet another issue to divide us even more.
I’m a black man and I cringe every time I hear them going on and on about how AI is supposedly more of a massive threat to me because I’m black. Most of the issues they point to, however real they may be, tend to be mostly data-related issues, which can be addressed without turning to the racism and “tech bros” narrative. This reminds me of the Google Photos gorilla incident some years ago (I knew some of the folks involved). Everyone was piling on Google for identifying a black man as a gorilla in their Photos app; everyone in tech seemed to loved the narrative that Google was supposedly a white supremacist organization and did that on purpose when, in reality, it was simply a data issue. Obviously, that was a bad outcome and it’s perfectly reasonable to have expected better from Google, but rather than focusing on how to prevent issues like that in the future (i.e., making sure to have diverse training data sets), the tech industry turned that situation into an emotional virtue-signaling event to pander to black folks and push the narrative that tech is inherently evil and racist.
Regulation is the only source of 'ethical behavior' in capitalism so it concerns me when I see all the attacks on AI regulation recently.
An over simplification, of course, as the real history of the atomic bomb and the era that followed is far, far more complex.
AI is no different in the advantage it would offer to the country that first develops a concrete weaponized model. Since we still think of the world in terms of nation states as opposed to a global humanity, it makes sense that there would be fear surrounding its development.
It also makes sense that we would lose our ethical goals pretty quickly due to Othering in this scenario. It's easy to look at a perceived enemy and think "well, if they had this tech they wouldn't hesitate to use it against us!" to justify using said tech. By stripping our perceived enemies of any morality or ethical postulates, we reduce them to aggressive threats, which is interesting since the act of viewing another human as anything "less" than human is itself an aggressive threat.
What do you mean by "true artificial general intelligence"? Is there anything that a 100 IQ human could do that GPT4 with a body and more memory couldn't do? The only thing preventing GPT from having a body is the engineering problem of getting LLMs to process live images and audio fast enough for realtime.
There is far more to human knowledge than what we write down and what can be read by machines. You can read all the books in the world about tennis but that won’t teach you how to play the sport at a competent level.
You need to practice a lot! Playing a sport involves the development of deep muscle memory and strength + fitness. Muscle memory is tacit knowledge [1]. It can’t be conveyed through words, it must be developed through practice and correction from an expert instructor. There are large categories of human knowledge that fall under the tacit umbrella, and LLMs will never master any of them.
The "more memory" part is carrying a lot of weight here, and it represents a very difficult problem. (Specifically, finding a way to identify and access contextually-relevant data in real time.)
Definitions of AGI vary, but almost everyone agrees that GPT4 is not AGI. Can transformers continue to be scaled to the point where they can power an AGI? Sure, maybe. But like the person you responded to said, there's currently not enough evidence to draw that conclusion.
These AI are statistical models. Next token generators. Not living beings.
That’s like calling a cellphone a “communication slave”
AI is just a tool. All this “safety” work feels like theater to keep everyone thinking that AI is a cool sci-fi level computer-person.
It’s essentially search, but with natural language. If they really wanted safety, they just would spend the extreme time and effort cleaning their datasets.
I don’t agree with that as being the goal, but I can’t see companies pouring billions into datacenters full of graphics cards for an AGI that is ridiculously intelligent but can be as arbitrary and capricious as people. They will always want it to be perfectly compliant.