What could it mean for an AI to be "politically neutral”? And can we measure it? New paper + dataset.

We propose a defn that applies to any type of conflict: a neutral response should maximize approval on both sides of an issue, while keeping that approval balanced.

1/🧵 Key finding: neutrality isn't a fantasy. Our experimental “balanced” response wins high approval from both sides at once — even when the sides strongly disagree on the substance.

And costs less than 10% in approval vs. answers that agree with you (red and blue arrows).

2/

Talking to AI shifts political opinions. Sometimes a little left, sometimes a little right. And reduces polarization a bit too!
How *should* AI influence our politics? Well, I have a practical idea about that...
🧵 A machine that tried to never change your opinion would be pathological: it would hide the most important information from you.
Instead, the machine should only change your opinion through *fair* methods.
2/

Meta will stop paying pro fact checkers, and switch to a community notes system. In part, they say this is because of fact checker "bias."

I don't think this is good news. But it must be said that the fact checking community did this to itself.

How?
🧵

about.fb.com/news/2025/01/m… In my view fact checking has two goals:
1) Identify harmful falsehoods
2) Create trust with the audience

The fact checking community mostly succeeded at #1. And they mostly failed at #2. They were mostly right, but many people no longer believed them.

Here's the evidence...

Everyone has heard that there are serious problems with optimizing for engagement. But what are the alternatives?
New paper: we gathered folks from eight platforms for an off-the-record chat to share best practices -- and document them from public sources.
arxiv.org/abs/2402.06831
𝑹𝒂𝒏𝒌𝒊𝒏𝒈 𝒃𝒚 𝒑𝒓𝒆𝒅𝒊𝒄𝒕𝒆𝒅 𝒆𝒏𝒈𝒂𝒈𝒆𝒎𝒆𝒏𝒕 𝒄𝒂𝒖𝒔𝒆𝒔 𝒔𝒊𝒈𝒏𝒊𝒇𝒊𝒄𝒂𝒏𝒕𝒍𝒚 𝒉𝒊𝒈𝒉𝒆𝒓 𝒕𝒊𝒎𝒆-𝒔𝒑𝒆𝒏𝒕 𝒂𝒏𝒅 𝒓𝒆𝒕𝒆𝒏𝒕𝒊𝒐𝒏

That's why everyone does it. And this isn't necessarily bad -- real user value here -- but sometimes it goes wrong.

“When journalism begins to accept the death of objectivity, the industry will begin to thrive off relying on organic humanity rather than stiff, rigid and outdated mechanisms.”

I don't think "the death of objectivity" is the right direction. I see better alternatives.

1/x Why are mainstream journalists gunning for "the death of objectivity"? The short answers:

- it was always a little bit of an incoherent concept
- it was used to exclude marginalized people and perspectives

There are real problems here. But. Reporting is more than opinion.

2/x

Instead of "amplification," here are three alternative ways of talking about what recommender systems do that correspond more closely to how this all works -- and would make better law.

1/x When I push people on what "amplification" means, we always end up at one of three ideas:

1) Reach
2) Comparison to a chron baseline
3) User intent

2/x

I think we expect far too much of mis/disinfo interventions. Too much of the work in this field rests on the insane implication that we could fix our politics if we just use mass surveillance tools to stop people from saying bad things to each other. I want to qualify this by saying there really are organized information operations, content farms that churn out "news" that isn't, deceptive fraud and scams, etc. I'm not an absolutist by any means. But removing content isn't going change e.g. vaccine hesitancy.

I wrote a paper containing everything I know about recommenders polarization, and conflict, and how these systems could be designed to depolarize.
arxiv.org/abs/2107.04953

Here's my argument in a nutshell -- a THREAD First, what even is polarization? We all kinda know already because of the experience of living in a divided society, but why do we care? A few big reasons:
- it destroys social bonds
- it escalates through feedback cycles
- it contributes to the destruction of democracy
2/n

I have been studying disinformation since 2011. I am horrified by how the category has expanded from “factually wrong” to “expresses a frame I disagree with.” This is reflected in ML work, where no one looks too closely at where the labels in their training data comes from. In complex categories like “disinformation” it’s quite important to understand the nature of the actual examples being used. Read the papers closely. Go look at the underlying data sets. For most disnfo classifiers, the labels are not even at the article level — just source level

If you want to use AI/ML for good, have you considered... form extraction? This is a major difficulty for climate change, human rights, medicine, and journalism, and a state-of-the-art-challenge.
A THREAD.
jonathanstray.com/to-apply-ai-fo… Today I am pleased to announce a new public form extraction benchmark, hosted by @weights_biases. This is a nasty campaign finance data set, the FCC Public Files. We've collected 20,000 labelled PDFs, in hundreds of different formats.
wandb.ai/deepform/polit…
2/

Folks who want to use AI/ML for good generally think of things like predictive models, but actually... smart methods for extracting data from forms would do more for journalism, climate science, medicine, democracy etc. than almost any other application. A THREAD.
1/x Here's how form extraction could help climate science

https://twitter.com/ed_hawkins/status/1167769410238595072

2/x

Currently watching the @TowCenter discussion on conservative news. They interviewed many conservative journalists to understand how they see their work, and this is the report.
cjr.org/tow_center/con…

thread. Here are the questions they wanted to answer from talking to journalists at conservative news outlets.

2/

The AI text adventure: a THREAD about an emerging art form.

ICYMI, I recommend playing AI Dungeon 2 by @nickwalton00, released last week by. It's a GPT-2 based text adventure. Rudimentary, a prototype, but I believe it's a harbinger of great things.
theverge.com/tldr/2019/12/6…

1/ @nickwalton00 The crazy, amazing thing about this experiment is that the AI will respond in (more or less) sensible natural language to anything you can throw at it. This was the promise of "interactive fiction" of old, what we hoped for in the 70s and 80s when it emerged. But..

2/

A THREAD on the the divide between the "critical" and the "technological" communities, as a response to the paper by @davidjmoats and @npseaver which explores this topic. I'm posting the bits I highlighted and explaining why each of them struck me.
journals.sagepub.com/doi/full/10.11…
Ready? @davidjmoats @npseaver It starts with @mathbabedotorg's 2017 editorial. Apparently the critical community (not the best name, but let's go with it for now) was super surprised to hear that they were not critiquing technology. Me, I recognize Cathy's critique very naturally. The problem is...

Ok, a THREAD about doing journalism for a polarized, suspicious public.

Suppose you want to be read and trusted across the political spectrum. What does that mean and how to do it?

1/ This is the background context:

Media trust has been falling for decades
knightfoundation.org/reports/indica…

Polarization has been increasing for decades
people-press.org/2014/06/12/pol…

(I'm going to focus on the US today but there are similar trends internationally)

2/

THREAD on a wacky problem that AI ethics theorists should be aware of.

Consider a coin flip bet: Heads you get 1.5 times your current wealth. Tails you get 0.6. Should you take it?

Expected value is 1/2*0.6 + 1/2*1.5 = 1.05 times your wealth. So take it right?

Not quite... Averaging over many people, it's true that the expected value of wealth increases. But if you watch one person's wealth over time, it decreases!

If you work it out, you get sqrt(0.6*1.5) ~= 0.95 times as much wealth at each time step.

(from ergodicityeconomics.com/lecture-notes/)