That is extremely far from the case. These biases are in the eye of the beholder. People with opposite feelings about Google (or whatever the corp of the moment is) see exactly the opposite bias.
There can also be tons of bias in how something is presented. How will google handle a post that has factually correct information presented in a biased way?
I know that sort of thing happens sometimes (Google presenting a spurious statement as a categorical answer) but those are bugs. As long as they are very rare, and fixed quickly when they occur, I don’t see them causing much harm.
OK, some people believe anything they read (especially if it confirms their existing biases), but that problem has always existed. I think Google’s occasional snippet fuck-ups are a drop in the ocean compared to the spread of false information through social networks.
Not a super well thought out article. Example: lots of speculative complaints that ChatGPT will lead to an explosion of low quality and biased editorial material, without a single mention of what that problem looks like today (hint: it was already a huge problem before ChatGPT).
Ditto with the “ChatGPT gave me wrong info for a query” complaint. Well, how does that compare to traditional search? I’m willing to believe a Google search produced better results, but it seems like something one should check for an article like this.
IMO we’re not facing a paradigm change where the web was great before and now ChatGPT has ruined it. We may be facing a tipping point where ChatGPT pushes already-failing models to the breaking point, accelerating creation of new tools that we already needed.
Even if I’m wrong about that, I’m very confident that low quality, biased, and flat out incorrect web content was already a problem before LLMs.
> Admitting the possibility of bias is very different than admitting bias in every single trending content decision.
They don't have to admit bias in every content decision. They have to admit systematic bias in content decisions in the aggregate, which does not necessarily imply bias in every content decision.
While it is technically, theoretically possible that their system is open to the introduction of these biases but does not actually exhibit those biases in the aggregate, it is astoundingly unlikely.
On one of the images, they advise the moderators to demonetize controversial content, and then make exceptions for certian controversies (which Google does not feel it would be right to suppress). This is obviously bias; it's the definition of bias. It just doesn't feel like bias to us because if we were made autocrats and instructed to rule as undemocratically as possible, we would enforce the same bias.
Ouch. Some pretty damming stuff there. I can see spokespeople getting something wrong, but multiple points that are seen as highly contentious in SEO land being explicitly denied and the docs say opposite? Not much wiggle room there
Google seemingly does this a lot (although I don't know if it's included in their transparency report or if anyone has tried to measure it -- both of which would be nice). You can report predictions as "inappropriate" (look at the lower right of a prediction list in a desktop browser) and you'll see something like
Which predictions were inappropriate?
[...]
The predictions selected above are:
Irrelevant
Violence or gore
Sexually explicit, vulgar, or profane
Hateful against groups
Sensitive or disparaging for individuals
Dangerous or harmful activity
Other
If Google agrees with the complaint, the prediction could be removed regardless of the popularity of the underlying search.
Yes, I agree. They are naturally bimodally distributed, but I'd expect Google is arrogant enough to think that they can tell the difference between an honest and manipulative review using such statistics anyway.
They obviously can't, and that's why people are outraged.
Another possible explanation is that Google's search engine is exhibiting, due to the lack of a better term, a form of "transference" [0].
Google engineers are huge fans of XKCD as is evident from Randall Monroe's well attended talk at Google some time back. In my mind, his talk was easily one of the most attended talks, second only to Linus Torvalds talk on Git.
[0] I'm sure there is a standard term for software exhibiting biases held by its authors when making decisions on behalf of users.
This is pretty much the backlash everyone said Google would be getting if their model isn't literally perfect on release, no? Also, aside from obvious hallucinations, I'm having a hard time imagining a model that would respond with unbiased views, because the response would need to be interpreted by a person's cultural views anyways. I guess, they could just respond with "Sorry, I can't answer that question", but then again, there would be a backlash how they limit their model while answering sensitive topics.
All search engines are explicitly biased. That is the point, they generate a ranking of results. Heck even how you tokenize text is an explicit bias of what you match against.
I don't get what you're trying to say. If Google stops showing news from participants, but show news from non-participants (due to false negatives causing their algorithm to fail to identify those sites as news), would that not be obvious discrimination according to the above quote?
The predominant backlash to Google's recent ToS changes come to mind as an example of selective content being reported out of context.
reply