4 ms·

What a strange comment section. A customer service chatbot telling your customers that an issue is solved, when the issue isn't solved, is a core product failu

What a strange comment section. A customer service chatbot telling your customers that an issue is solved, when the issue isn't solved, is a core product failure. It's also a Responsible AI issue. A code generation model suggesting GPL code without attribution is a core product failure and serious legal risk [1]. It's also a Responsible AI issue. A diffusion model generating child porn or nazi propaganda is a core product failure and a serious legal risk [2]. It's also a Responsible AI issue. In other words, the problem isn't that Responsible AI is BS. The problem is that Responsible AI is a conglomeration of serious legal and product risks that companies NEED to solve in the context of specific products. More importantly, how will those risks change over time? The western world public is overwhelmingly skeptical of AI and in favor of regulation. Even if new government regulation can be gummed up, you can expect juries to be less than sympathetic. But you’re not going to solve those problems by hiring a bunch of philosophers and interdisciplinary types that obsess over “nutrition labels for models” and the like. Everyone knows that Tay being a nazi is bad. And I don’t need a PhD in Sociology or whatever to tell you that you probably shouldn’t be openly condescending toward your customers. But constraining and aligning the output of models is a really hard technical problem, and needs a different type of expertise. So, I think we’ll see Responsible AI teams start to look a lot like e.g. safety at automotive companies, verification & validation at hardware companies, or security at tech companies. Mostly staffed with technical people who are strong engineers and happen to specialize in data governance/constrained generation/rotics safety/etc. There will be dedicated teams that kinda-help-kinda-police the rest of the company (think security). But there will also be a lot of diffused responsibility within each product group, with ICs or small groups within each group owning this space. [1] https://githubcopilotlitigation.com/ https://githubcopilotlitigation.com/ [2] https://en.wikipedia.org/wiki/Strafgesetzbuch_section_86a https://en.wikipedia.org/wiki/Strafgesetzbuch_section_86a

giraffe_lady 4y ago
But it has a whiff of what used to be called political correctness so we must RAGE and reject it entirely.
cypress66 4y ago
You don't need an ethics team to tell you about those risks, they are pretty obvious. What you need is your legal team to tell you how risky it is, and management to decide how risk averse they want to be.
- thwayunion 4y ago
  > What you need is your legal team to tell you how risky it is, and management to decide how risk averse they want to be. No. 1. Even if you're not going to do any mitigation, you do need engineers and researchers to help do the risk analysis. It's a collaborative process; rarely can a lawyer provide a full risk analysis without asking questions of eng/research, and sometimes answering those questions requires doing novel R&D and/or non-trivial engineering work 2. What happens when the lawyers say "very risky" and management isn't willing to take that risk but also doesn't want to "sit out"? In that case, technical folks are needed who can help mitigate risks. Constraining and aligning the output of these models is non-trivial and in many cases remains a completely open research problem. And that's only the legal risk. As my comment pointed out, there are also functional requirements. Consider a chip maker fabricating a new chip: eliminating bugs in the architecture demands hardware V&V, regardless of what legal determines about liability, because the chip doing what you want it to do is a core product functionality. Similar for most applications of text generation models (image less so, for now, but any integration with the real economy will introduce the same sort of core product functionality requirements).

giraffe_lady 4y ago

But it has a whiff of what used to be called political correctness so we must RAGE and reject it entirely.

cypress66 4y ago

You don't need an ethics team to tell you about those risks, they are pretty obvious. What you need is your legal team to tell you how risky it is, and management to decide how risk averse they want to be.

thwayunion 4y ago
> What you need is your legal team to tell you how risky it is, and management to decide how risk averse they want to be. No. 1. Even if you're not going to do any mitigation, you do need engineers and researchers to help do the risk analysis. It's a collaborative process; rarely can a lawyer provide a full risk analysis without asking questions of eng/research, and sometimes answering those questions requires doing novel R&D and/or non-trivial engineering work 2. What happens when the lawyers say "very risky" and management isn't willing to take that risk but also doesn't want to "sit out"? In that case, technical folks are needed who can help mitigate risks. Constraining and aligning the output of these models is non-trivial and in many cases remains a completely open research problem. And that's only the legal risk. As my comment pointed out, there are also functional requirements. Consider a chip maker fabricating a new chip: eliminating bugs in the architecture demands hardware V&V, regardless of what legal determines about liability, because the chip doing what you want it to do is a core product functionality. Similar for most applications of text generation models (image less so, for now, but any integration with the real economy will introduce the same sort of core product functionality requirements).