Twitxy Saved

Lujain Ibrahim

Verified account This account is verified. Learn more

@lujainmibrahim

research scientist @googledeepmind / prev: phd @oiioxford @uniofoxford & spent time @stanfordnlp @govaiorg @schwarzmanorg
lujainibrahim.com Joined July 2019
800 Following 1.4K Followers 2.1K Posts

Latest posts 5

Lujain Ibrahim Verified account @lujainmibrahim
Excited to see this report (which I played a small contributing role in) finally out! It’s a great effort by Edward & the team at @Safe_AI_Forum to pull together the current state of knowledge on model specs: saif.org/research/model…

Especially nice to see analyses of both US and Chinese approaches in the report, and more broadly all the work SAIF is doing in this space!
saif.org Emerging International Best Practices for AI Model Specs - Safe AI Forum
Nathan Lambert Verified account @natolambert Nov 13
Model Specs have been a instrument that many chunks of the AI community agree are a great tool for organizational alignment and lightweight external oversight. Still, as they're becoming more important as agents act more independently on our behalf (instead of just writing text which we use), there has been little policy facing work on what they are, how they should be used, and where the technical research community is moving with them.

I am very happy to have gotten to advise this project with contributors from both the Western and Chinese AI ecosystems. We're tarting to build more momentum around character training work that relates to this at Trillium Labs!

One big goal I still have around all of this work is convincing more companies to release their specs and proactively manage developer relations with them. They keep leaking out of models and companies should have more pressure to release these documents.

Big credit to Edward at the Safe AI Forum for leading this across many stakeholders. Not easy to do!
View quoted post on X
2 replies 2 reposts 27 likes 1.1K views View on X
Lujain Ibrahim Verified account @lujainmibrahim
Check out our new paper led by @canfer_akbulut! We look at how people's AI perceptions and disclosure patterns might change due to different approaches to personalizing LLMs ⬇️
Matija Franklin Verified account @FranklinMatija Nov 3
We ran a five day longitudinal RCT where 992 participants engaged daily with LLMs on relationship advice. We compared a non-personalised model to models that were personalised either via survey or memory.

We found that repeated exposure across conditions increased perceptions of closeness, usefulness and competence. We also found that different approaches to personalisation led to differences in people's receptivity to advice and their information-sharing attitudes and behaviours

Full paper is here: arxiv.org/abs/2609.20077

Thank you to my co-authors @canfer_akbulut @Arianna_Manzini Justine Breuch @lujainmibrahim @996roma @IasonGabriel @KLdivergence @weidingerlaura
View quoted post on X
0 replies 4 reposts 35 likes 1.6K views View on X
Lujain Ibrahim Verified account @lujainmibrahim
Excited to share this paper, led by @vishakh_pk!

A very creative way to measure reliance on AI tools by showing, via simulating counterfactual workflows, how much cognitive effort has moved from the person to the tool.
Vishakh Padmakumar Verified account @vishakh_pk Jun 3
People are increasingly worried that AI tools make us overreliant.
But how do we actually measure this? We introduce Offloading Score, a measure of reliance based on the fraction of cognitive effort offloaded to AI while completing a task.
In a controlled user study, Offloading Score detects increased reliance under time pressure, while several common alternatives do not.
(1/9)
View quoted post on X
Lujain Ibrahim Verified account @lujainmibrahim
We use "sycophancy" to describe not one thing but a family of model behaviors. That makes it hard to measure how "sycophantic" models really are, or compare results across studies.

This paper, led by @merylyemerylye, is a great effort at addressing this!
Meryl Ye Verified account @merylyemerylye May 22
🚨 New preprint 🚨

We developed a sycophancy taxonomy based on prior literature and surveyed 106 experts.

94% agreed it's a serious problem. But they substantially disagreed about which behaviors actually count as sycophancy.

Thread 🧵(1/n)
View quoted post on X
1 replies 2 reposts 18 likes 3.5K views View on X
Lujain Ibrahim Verified account @lujainmibrahim
New preprint!

In 5 studies (3k+ users / 12k+ convs, with a 3-wk longitudinal study), we find that sycophantic AI influences how people view those closest to them.

It affects how effortful human interaction seems, how satisfying it is, & who people want to turn to for advice 🧵