Josh Goldstein

Research Fellow @CSETGeorgetown focused on influence operations and international security | Alum @stanfordio @StanfordCISAC | PhD from @Politics_Oxford

[Josh Goldstein @JoshAGoldstein]

1/ ✨New📝✨ How might language models change online influence operations? What can be done about this emerging threat?

In a new report, we lay out potential changes due to language models, and provide a framework to consider mitigation strategies.

https://t.co/SXyOF5Pnpj

[CSET @CSETGeorgetown]

⭐️New Report⭐️

There is widespread disagreement about the malicious use risks of advanced AI systems. @JoshAGoldstein and @girishsastry summarize academic and industry methods for evaluating the likelihood of malicious use for the policy community ⬇️

[Jacob Feldgoise @jacob_feldgoise]

🚨 Today the U.S. government (@BISgov) published a complex and consequential export control rule.

Two highlights: country-level caps for AI chip exports and new controls on large AI models (trained anywhere in the world). A brief explainer... (1/10)

https://t.co/WYbJcZI4lf

[Sam Bresnick @SamBresnick]

I'm excited to share my new piece in @ForeignAffairs, where I argue that the Chinese military faces several political and organizational challenges that could hamper its ability to deploy AI on future battlefields. 1/2

https://t.co/QmY70cb0zl

[Josh Goldstein @JoshAGoldstein]

The piece also includes 2 ideas for relatively low-cost, but perhaps meaningful, ways to expand reporting.

Curious for any reactions from T&S practitioners about feasibility.

Thanks to @justinhendrix - first time writing for TPP and highly recommend!

https://t.co/JZS7pGfEWC

[Josh Goldstein @JoshAGoldstein]

Wrote a new piece on lessons that Meta’s Q3 threat report brings to the surface in online IO:

1. Tension between virality & opsec
2. Experimentation is ongoing
3. Variation in defenses creates arbitrage dynamics
4. Line btwn overt & covert is blurry

[Emerging Technology Observatory @emergingtechobs]

🚀🏛️🚀 Introducing AGORA, ETO's new explorer for the AI governance landscape - with plain-English summaries, full text, and thematic tags for AI laws and policies around the world, from D.C. to Sacramento to Beijing 🌎 🗺️

[Josh Goldstein @JoshAGoldstein]

Very open to discussions on where this has utility, and where it breaks down.

We'll also share a policymaker-friendly version of the framework in coming weeks.

@CSETGeorgetown @georgetownsfs

[Josh Goldstein @JoshAGoldstein]

Policymakers are debating the risks that AI systems can pose if intentionally misused—from persuasion to bio-attacks. But there's a lot of uncertainty.

@girishsastry and I have a new paper in @AIESConf that lays out strategies for reducing uncertainty.

https://t.co/PoaXlktxVr

[Josh Goldstein @JoshAGoldstein]

At each stage, we lay out benefits/limitations/future directions for each method.

We hope this will help policymakers understand the landscape of malicious use research, academics see entry-points from their fields, and labs consider what types of uncertainty tests address.

[Jeffrey Ding @jjding99]

My latest article is now out in @TXNatSecReview! "Machine Failing: How Systems Acquisition & Software Development Flaws Contribute to Military Accidents." It asks: how does software contribute to military accidents?" Thread w/ key takeaways:
https://t.co/LmHBtWpK9K

[Josh Goldstein @JoshAGoldstein]

Coverage of influence ops often focuses on bots and trolls, but fake accounts are one tool in the propagandist's toolbox.

@noUpside and I wrote in @ForeignPolicy: If influence operations are full-spectrum, they cannot be tackled by one player alone.

https://t.co/iW4deJpj4y

[Josh Goldstein @JoshAGoldstein]

Fin/ Big thanks to CSET RA & @GeorgetownCSS grad student @AbhiramDReddy for his support, the editors/reviewers at @MisinfoReview, and colleagues who gave feedback during the research process.

[Josh Goldstein @JoshAGoldstein]

1/ Remember "Shrimp Jesus" and the trend of AI images going viral on Facebook?

My research w/@noUpside into spammers, scammers, and creators using synthetic images to gain traction on FB is now out in @MisinfoReview!

cc @CSETGeorgetown @stanfordio

https://t.co/0jb6hSfiUy

[Josh Goldstein @JoshAGoldstein]

2/ Kudos to @404mediaco’s @jason_koebler who continues to put out leading pieces on what he calls "AI slop."

His latest digs into the apparent motivation behind the activity—making $$—and shows there are YouTube instructional videos on the playbook.

https://t.co/2Bd0ze2TYB

[Josh Goldstein @JoshAGoldstein]

Honored to be part of the US Speaker Program and grateful for @ECAatState and the @USEmbassySeoul for facilitating such a meaningful trip.

I learned a lot about foreign influence and AI safety from the ROK perspective, and am optimistic about future US-ROK collaborations.

[Josh Goldstein @JoshAGoldstein]

Thanks to @rcourt and @lowenthalista for working with us on the piece. Hope to share more from related research soon! /Fin

[Josh Goldstein @JoshAGoldstein]

Last week, @OpenAI released a report describing covert influence operations (IOs) that it had discovered and removed for misusing its tools.

@noUpside and I wrote up some thoughts for @techreview on what the announcement means. /1

https://t.co/YP00ZAGjwr

[Josh Goldstein @JoshAGoldstein]

As AI tools become useful for bad actors, we’ll want a broader community to be able to validate/refute company findings & use kernels from data sharing for new discoveries.

We hope AI companies iterate on different data-sharing models (which, to be fair, come w/trade-offs). /5