Hacker News Reader

Comment by strogonoff | Hacker News Reader

I used to think that unbiased news is an intractable problem purely based on information theory: to have an unbiased picture of the world, you need to have all the facts, which is physically impossible; and any layer that simplifies those facts (be it a journalist or an LLM) will add bias (either by journalist, journalist’s employer, LLM’s training data or operator preferences, and so on). In short, when your selection of facts to report is already biased, there’s not much you can do. In even shorter, map is not the territory (and all you have is a map).

However, I realized that’s not really true.

I mean, of course you cannot have unbiased news, and reevaluating the definition of “news” is not going to help you.

It’s just that there is actually a very good mechanism for combatting this—and it’s called “be aware of your bias and upfront about it”. There is any other way around it.

View on HN

To add: where I wrote “be aware of your bias and upfront about it” I of course meant journalists and news organizations.

Corollary: since this awareness of the bias requires self-honesty and introspection, it’s drastically more difficult with an LLM because you also need to know all about your LLM developer and operator’s accidental and intentional biases.

Technically speaking, if every day you take an uniformly random selection of news articles, on average what you read will not be biased. Or, to be precise, it will follow the same biases as the population of news articles you sample from.

> what you read will not be biased. Or to be precise [...]

Your "to be precise" sentence states directly that what you read WILL be biased.

Replace “news articles” with “facts” and you’re closer to the point, but then how useful would that input be to you?

The difference is that facts are harder to consume than news articles, as the latter is essentially packaged and distributed collections of the former.

Furthermore, the intent of parent's bias averaging, as I understand it, is...

1) There are a finite number of major biases axes in mass media (e.g. political parties, wealthy-poor, conservative-liberal, etc.), ergo averaging these out is a helpful reduction of major biases

2) It assumes the underlying material has a bias, and seeks to counteract it. If you build a fact-based method, you leave yourself open to treating biased opinions as facts, if they're sufficiently masked

Indeed, the first challenge to unbiased approach would be the impossible in practical terms task of separating facts from fiction.

I don’t know if it is clear but I really meant it that unbiased news is not a possibility due to fundamental limitations of information theory and finiteness of individual information processing capabilities; self-awareness and bias transparency is the only way.

> helpful reduction of major biases

Only if those major biases are opposite between media.

Imagine biases as vectors pointing in arbitrary directions. Best case is when bias vector in one cluster is compensated by exactly opposite bias vector in another cluster, but how often do you expect that to happen? A more typical scenario is a vector pointing in merely a different, rather than opposite, direction. A worst-case scenario, and also common (because bias manifests itself in selection if facts to report, and most sources report on the same events), is when all bias vectors point in a similar direction.

In your original (upthread) argument that "unbiased news is not a possibility", are you stating that as a fundamental consequence of the news' lossy compression of {all facts} into a smaller product... or something else? It's unclear.

As for averaging out, yes, the vector model was what I was using internally.

Disclosure: I'm US-based, so our politics are traditionally bipolar as a result of our election system (in contrast to Europe et al.).

I would be surprised if vectors don't cancel often though, because think what we're really talking about in a capitalist information space: popular perspectives that are profitably marketable.

What is the easiest way to find an audience? Do what your competitors are not.

So on the whole, high revenue and production cost media (tv, web) in a capitalist information space will converge into bipolar pairs on any given issue, in order to maximize market share.

In lower cost media or operations, it will more likely be a random scattering of vectors.

> Disclosure: I'm US-based, so our politics are traditionally bipolar as a result of our election system (in contrast to Europe et al.).

There's a whole lot of bias that's "baked in" with US politics / media. There's broad consensus on neoliberal capitalism, personal freedoms vs collectivism, common ground on opposing states like Russia and China.

On a few wedge issues you will average out to a common ground, but ultimately you're not eliminating bias, you're just eliminating a relatively small bias between left-wing and right-wing US and replacing it with an "overall US" bias.

From mass market sources, true. Probably because they're aiming towards the common American, who generally has similar feelings. Ergo, they're that + whichever politically-polarized biases.

As for whether averaging that out is useful...?

I'd be happy to have news from a "overall US" bias, without the right/left political spin.

It's almost impossible to arrive at a pure random sampling like that. As an easy example, if you limit yourself to English news articles, you're going to have a pretty large bias towards "the West", which right now probably includes a strong bias against China or Russian's viewpoints on things (not to say that including those viewpoints would be more truthful necessarily, but certainly less biased).

Unbiased news algorithm:

1. Identify all clusters of power and opinion, eg. Republicans, Democrats, Chinese Communist Party, UK Labour Party, Green Party, etc.

2. Associate each cluster with news organizations that tightly align with their policies.

3. Evaluate stories - headlines in only one cluster (under reported elsewhere) represent a view. Headlines that have different takes, likewise, give you a spectrum of thought on an issue or topic.

4. Collect fine points and build a report. Bonus for providing contrasting opinions and associating them with the party/power responsible for that thinking.

I want this tool. It'd be okay if it wasn't perfect. Just getting 80% of the way there would be incredibly valuable. I'd pay for it.

When I started working on Zeitgaist [1] I immediately recognized that biased information is the main problem. I'm currently thinking about attaching additional information to the sources I present like ground [2] is doing with political spectrums. I really like your idea for the news algorithm. If you or someone elsw wants to build something like that on top of Zeitgaist with me, don't hesitate to contact me.

[1] https://zeitgaist.ai [2] https://ground.news/

Your "etc" includes some very powerful ethnic groups. Are you planning on "identifying" them?

I didn't think so.

I don't see why you can't. Throw religion in while you're at it. Even interest groups around other classes: disability, socioeconomic, etc.

You can absolutely say a newspaper (or individual columnist or editor) is pro-LGBT, pro-Israel, pro-Palestine, pro-Mormon, pro-Amish, pro-Atheist, etc. That shouldn't be controversial. Wikipedia can and does do this in a relatively neutral way.

If your classification is meant to disparage a group, then I'd see the trouble. Don't do that.

I want an up to date, well rounded view of how everyone sees the world. A system capable of getting me to see geopolitical events in other perspectives is useful, even if I ultimately choose to uphold my existing beliefs.

FWIW, I subscribe to all sorts of diverse political view subreddits and Twitter handles. Unfortunately, it's still hardly comprehensive is a lot of work to analyze and distill.

What we're really talking about here is boting any deviation from a tabula rasa appraisal of reality, then clustering that into an affiliation/quality, right?

I.e. Christians (as an averaged whole) can look at the same situation as a blank slate and have a statistically significant and consistently different perspective

As can scientists. As can maybe microbiologists and industrial chemists.

The point of the clustering isn't that the tags are pejorative, but that they're predictive metadata on any source's bias.

And often times that's currently unexposed. I see a post by echelon, but only by the content of that single post (and maybe some prior HN contact) can I reconstruct his typical perspectives.

It would probably include some institutions masquerading as representatives of ethnic groups but that isnt quite the same thing.

1. Will your cluster identification be unbiased? How do you avoid providing more fine-grained taxonomy for clusters that hit close to home for you (or this tool’s operator) and shoving the rest into as few buckets as possible (because you don’t care)? 2. What about biases shared by all news organizations, regardless of cluster?

Reply on news.ycombinator.com