iiinova — where your digital self keeps evolving

Ziv Epstein

@ziv-e.bsky.social

284 followers 141 following 15 posts

Postdoc at MIT College of Computing • PhD from MIT Media Lab • Creativity, values and (more-than-)human-centered design

Posts Replies Media

Ziv Epstein @ziv-e.bsky.social · 19/08/2026

Important caveat: We ran our study in October 2024 & platforms change over time. See the below tweet re X's tweak in July 2026. Beyond any particular design, our paper raises the question, if we can measure the values the algorithm amplifies, whose values should they amplify? x.com/nikitabier/s...

020

Ziv Epstein @ziv-e.bsky.social · 19/08/2026

But why is this happening? It seems that the algorithm places more weight on replies — and we are often more likely to reply to content we disagree with, which the algorithm (mis)takes as a positive signal. It treats those rare replies as a stronger signal than the flood of value-aligned likes.

120

Ziv Epstein @ziv-e.bsky.social · 19/08/2026

When contrasting these values to the values that the users actually hold, we find a disconnect: there's a negative correlation (misalignment) between users’ explicit values and the values in content the algorithm is more likely to amplify

120

Ziv Epstein @ziv-e.bsky.social · 19/08/2026

We find that the FYP algorithm is amplifying conservation values such as rule conformity, societal security and tradition, while demoting self-transcendent values such as caring , dependability, as well as universal concern and preservation of nature.

120

Ziv Epstein @ziv-e.bsky.social · 19/08/2026

Do our social media algorithms correctly reflect our values? Our new article published today in @pnas.org shows that the answer is often not, and that the content that gets promoted into their ranked feeds is often actively counter to our values.

23012

Ziv Epstein @ziv-e.bsky.social · 25/03/2026

In evaluation, we find that people agree with labels from off-the-shelf LLMs less than a random other person! But fine-tuning and then applying our personalization method yields a 66% relative improvement in agreement compared to human-human agreement rates, leading to SOTA performance.

110

Ziv Epstein @ziv-e.bsky.social · 25/03/2026

We collect 32K ground truth value expression annotations from 1K people on 5K representative social media posts using Schwarz Values. We then construct a personalization architecture for predicting value expressions by learning from a small number of informative calibration annotations per user.

110

Ziv Epstein @ziv-e.bsky.social · 25/03/2026

Measuring values is key for aligning systems with people’s values, but value expression is fundamentally subjective, leading to divergent labels from different people. We find meaningful disagreement in value annotations across raters. See for yourself, do you think the example contains humility?

100

Ziv Epstein @ziv-e.bsky.social · 25/03/2026

Pleased to share our new paper forthcoming in @icwsm.bsky.social! We introduce a novel framework to measure value expressions in social media posts at scale, leveraging personalization to handle the inherent subjectivity of human values. arxiv.org/abs/2511.08453

2329