4 ms·

Isn't there a problem with the sampling here? What if people who likely install RegretsReporter are also people who are more likely to view tangentially related

Isn't there a problem with the sampling here? What if people who likely install RegretsReporter are also people who are more likely to view tangentially related problematic content in the first place, and then are shown more of it. Also, the dividing line on what's consider problematic is difficult, as we saw recently with Right Wing Watch. And the proposed regulation seems even more problematic. If the AI model were public, then it would be far easier for people to create adversarial content to game it. This is a problem with any models built on public data, including OpenAI's GPT stuff, or GitHub's Copilot. Detailed knowledge of how it works allows more efficient adversarial attacks, and the more these services become important public infrastructure, the more value such attacks will be. Imagine a co-pilot attack that inserts hard-to-discover Heartbleed-esque bugs into code for example. It seems way way too early to be discussing regulation of algorithms of a field so young and changing so rapidly. 5 years from now, the network architectures may be unrecognizably different in both form and function. It might be better to have some industry-wide group like the W3C or IETF which sets these standards, and have tech published reports and audits for compliance, but done in a way to prevent attacks.

uniqueuid 5y ago
> Isn't there a problem with the sampling here? What if people who likely install RegretsReporter are also people who are more likely to view tangentially related problematic content in the first place, and then are shown more of it. Also, the dividing line on what's consider problematic is difficult, as we saw recently with Right Wing Watch. As I commented above, sampling is a problem, but the sample was only used to gather candidate videos which were then manually analyzed following Youtube's own content guidelines. So the takeaway is: According to the study's authors, Youtube recommends videos despite them violating their own guidelines.
- cromwellian 5y ago
  Does Right Wing Watch violate the guidelines? It’s difficult enough to make AI that could enforce these guidelines, but even with a staff of tens of thousands of human evaluators I doubt you could avoid even 10% false positives given how much people argue over classification already. This kind of filtering, some say censoring, is super problematic both to pull off And to please all stakeholders, which is why a one size fits all regulation is likely to fail and create a new class of problems.

uniqueuid 5y ago

> Isn't there a problem with the sampling here? What if people who likely install RegretsReporter are also people who are more likely to view tangentially related problematic content in the first place, and then are shown more of it. Also, the dividing line on what's consider problematic is difficult, as we saw recently with Right Wing Watch. As I commented above, sampling is a problem, but the sample was only used to gather candidate videos which were then manually analyzed following Youtube's own content guidelines. So the takeaway is: According to the study's authors, Youtube recommends videos despite them violating their own guidelines.

cromwellian 5y ago
Does Right Wing Watch violate the guidelines? It’s difficult enough to make AI that could enforce these guidelines, but even with a staff of tens of thousands of human evaluators I doubt you could avoid even 10% false positives given how much people argue over classification already. This kind of filtering, some say censoring, is super problematic both to pull off And to please all stakeholders, which is why a one size fits all regulation is likely to fail and create a new class of problems.