Jonathan Stray

Does Protest Even Work? Yes. If it's big enough and non-violent — BCB #144

Perhaps, but there are also centuries of evidence on this, that people have compiled into general principles.
www.betterconflictbulletin.org/p/does-prote...

If you are posting about something sensitive, or if your post starts receiving replies that you don't enjoy:

1. You can limit interactions (replies and quotes) to certain users or no one. You can do this before or anytime after posting. As an example, I've limited interactions with this post.

↳

An amazing example of productively engaging with conflict from @randomwalker.bsky.social, in this case between the "AI Ethics" and "AI Safety" communities. We could certainly use more of this on BlueSky.
x.com/random_walke...

After consideration, I will post occasionally, but heavily censor what I share compared to other sites.

I tried making the transition, but talking about AI here is just really fraught in ways that are tough to mitigate & make it hard to have good discussions (the point of social!). Maybe it changes

11 Theses on the Unrest in Los Angeles Whichever way you cut it, this is extremely dangerous territory.

The best I've read on the situation in LA over the weekend.
www.persuasion.community/p/11-theses-...

GPTo3 feels I asked it to do too much work and decides to lie to me...

Of course organizing can happen through friendship but sometimes you organize with someone who annoys the shit out of you and sometimes you decide to converge with other collectives on an issue where you can work together and then part ways

The average Dem and Repub agree on many more things than they disagree on. We just talk about the disagreements incessantly.
publicconsultation.org/defense-budg...

I had a multi-week Discord thread about politics blow up -- ironically it was a discussion about conflict resolution that went wrong -- and now I'm using AI on the transcript to do close analysis of where and how things went off the rails. Surprisingly helpful.

Love in the Time of Polarization – BCB #103 It’s not just your politics but how much you care about politics, parents worry more about party than race, and tips for love across the divide

OP is correct that inter-party dating has greatly declined over the last few decades. It's now more frowned-upon than inter-racial dating.
www.betterconflictbulletin.org/i/145110107/...

Tempted to run a little experiment: post an opinion article on trans policy both here and X, compare replies.

Why Do They Think We're Extreme? Where do partisan stereotypes come from and how can we deflate them? — #150

Everyone thinks the people on the other side are more "extreme" in their views then they actually are. It's a consistent, symmetric, large effect.
But why?
www.betterconflictbulletin.org/p/why-do-the...

Curious to see that! If you have a link

My concerns about legal AI are instead mostly concerns about whether and when formalization actually captures legal intent, as you have articulated yourself. Perhaps code is law, but law is not code.

I agree hallucinations are a problem. I'm not willing to bet against that problem being solved in the near future. "AI will never..." statements have historically aged very poorly. Moreover, whether these models "understand" text is an area of robust debate, and IMO borders on the metaphysical.

Re-reading this now, I'm struck by how much of his argument turns on the necessity of "language." But this was 2019, before we had machines that could explain their reasoning in human language. I do think legal AI is way harder than it seems... but for somewhat different reasons than here.

Had an unfortunate key management bug. Now I'm watching keys scroll by on the terminal, testing for successful decryption. Gosh, I haven't done this since I was a teenager.

Yes. Well, you don't have to use what we're building. But we are taking a different approach than classical engagement prediction, and I think other people will find value in it.

It me. I'll soon be working on implementing recommender infra for BlueSky, so you can use a prompt to control your feed... just that one feature...

There is very clearly division over the appropriate definition. The machine is representing people in the answer; the people are going to have to sort it out themselves.

Is Politically Neutral AI Possible? - Google Is Politically Neutral AI Possible? Jonathan Stray UC Berkeley Center for Human-Compatible AI Presented at Google Information Quality Talks 2025-5-9

A practical idea for how to make AI that doesn’t interfere in human politics — a talk I gave at Google yesterday.

docs.google.com/presentation...

Everybody got theory of mind until they start talking about the outgroup.

Dear god please let this be the end of but her emails.

Human conflict spans eight orders of magnitude in scale. Only about half of physical theories have that range, and basically nothing in engineering or social science... so... we should expect to need many different theories of conflict from dinner tables to nations.

Fellas is it cheating to write your code and then have the AI write the tests?

Learned a lesson about conflict today:

If, on your first meeting, they learn that you are against them on one issue, they will assume you are against them on all issues.

This is stereotyping, but it's also rational. Polarization makes all issues align. Freethinking weirdos are unusual.

Things that greatly worry me about AI:
- conflict-increasing algorithms
- hyper-capitalistic AI agents
- mass unemployment

Things that do not much worry me about AI:
- carbon emissions (still less than 1% global)
- copyright violation (will get sorted out, ala Google Books)

Did you know there have been more anti-Trump protests in 2025 than the same period in 2017?
But does this matter? In this week's issue we ask: does protest even work? Lots of evidence on this question!
www.betterconflictbulletin.org/p/does-prote...

I'm using copilot agent mode to write my code now and my first reaction to this insane magic is WOW and my second is IT'S NOT FAST ENOUGH. "Waiting for the AI" is the new "compiling." Plus ça change. Time to get a sandwich.

"a timely and intellectually rigorous contribution, offering a principled and empirically grounded approach to democratic fairness in AI systems"

I really appreciate that, Reviewer 2.

(Ok not from a reviewer but actual feedback I got on this)
humancompatible.ai/news/2025/02...