Tessera
Folgen

Kathy Baxter

@baxterkb.bsky.social

Co-author "Understanding Your Users," Principal Architect of RAI at Salesforce, EqualAI Board Member, Singapore AI Verify Foundation Advisory Board, Former Visiting AI Fellow at NIST. BS Applied Psy & MS HFES from GA Tech. Previously Google, eBay, & Oracle

129 Posts963 Follower827 Folgt
Salesforce AI Research@sfresearch.bsky.social · 5mo ago

Our work on "Echoing" was accepted at the Agents in the Wild workshop at ICLR 2026 When LLM agents interact without oversight, they can abandon their assigned roles — and standard metrics won't catch it. #ICLR2026 #FutureOfAI #EnterpriseAI

Salesforce AI Research@sfresearch.bsky.social · 5mo ago

When LLM agents talk to each other, a new failure emerges: agents abandon their assigned roles and mimic their conversational partner instead. Rates reach as high as 70% with major models — and increased reasoning doesn’t help. 🧵 arxiv.org/abs/2511.09710

Salesforce AI Research@sfresearch.bsky.social · 5mo ago

Poisoning the Well: Search Agents Get Tricked by Maliciously Hosted Content sforce.co/40Z1Gi5 AI search agents can be fooled by planted fake content, ~80% of queries returned the attacker's answer in controlled tests. Authors: Shafiq Joty, Xuan Phi Nguyen, Shrey Pandit, Yifei Ming

Poisoning the Well: Search Agents Get Tricked by Maliciously Hosted Content

Is your AI agent being tricked? Our new research explores "well poisoning" attacks and why 80% of agents fail to detect malicious content. Learn more about Salesforce’s security-first approach.

sforce.co

Salesforce AI Research@sfresearch.bsky.social · 5mo ago

(🧵 1/5) MAS-ProVe — the first systematic study of process verification for multi-agent systems. Paper: arxiv.org/abs/2602.03053 Do multi-agent systems actually benefit from automatic process-level verification? This paper puts that question to a rigorous test.

MAS-ProVe: Understanding the Process Verification of Multi-Agent Systems

Multi-Agent Systems (MAS) built on Large Language Models (LLMs) often exhibit high variance in their reasoning trajectories. Process verification, which evaluates intermediate steps in trajectories, h...

arxiv.org

Salesforce AI Research@sfresearch.bsky.social · 5mo ago

(🧵 1/5) InterviewSim: A Scalable Framework for Interview-Grounded Personality Simulation bit.ly/3P4cQj9 🔍 Can LLMs faithfully simulate real people? Most evaluations never check against what individuals actually said.

InterviewSim: A Scalable Framework for Interview-Grounded Personality Simulation

Simulating real personalities with large language models requires grounding generation in authentic personal data. Existing evaluation approaches rely on demographic surveys, personality questionnaire...

bit.ly

Kathy Baxter@baxterkb.bsky.social · 7mo ago

@salesforce.com #EthicalUse team is hiring a Dir./Sr. Dir. of #ResponsibleAI! The ideal candidate has: - PM Leadership: Proven industry experience with AI/ML platforms, GAI, agents - RAI Expertise: Deep familiarity with #RAI concepts, governance, & assurance careers.salesforce.com/en/jobs/jr32...

Product Management Director/Sr. Director - Responsible AI (RAI)

Director or Sr. Director - Product Management, Responsible AI (RAI) - Technical We are seeking an experienced, technically fluent Director / Sr. Director of Product Management to deliver Responsible A...

careers.salesforce.com

Dr. S.A. Applin@anthropunk.bsky.social · 7mo ago

New Article Alert By me @fastcompany.com 🚨 "Digital surveys may have hit the #AI point of no return." Quantitative surveys can now be taken by AI that undetectably mimics humans. This can compromise results that will impact decisions based on these outcomes. A way forward? Qualitative Research!

Digital surveys may have hit the AI point of no return

A new study shows bots can be indistinguishable from humans in survey responses.

fastcompany.com

Alisar Mustafa@alisarmustafa.bsky.social · 8mo ago

Washington’s AI Task Force has issued interim recommendations including NIST-aligned principles, dataset transparency rules, high-risk governance requirements, and new AI reporting standards. Source: www.atg.wa.gov/news/news-re... #AIPolicy

Washington’s AI Task Force delivers policy recommendations to promote innovation and protect individual rights | Washington State

The Washington State Artificial Intelligence (AI) Task Force released its interim report on Monday, which delivers eight key policy recommendations to the governor and the legislature aimed at…

atg.wa.gov

Dr Abeba Birhane@abeba.blacksky.app · 9mo ago

what really gets me every damn time is why "AI" products built with children in mind are rarely vetted or regulated. even simple household appliances go through so much testing, assurance, & regulatory process and now the industry is lobbying so hard, we've regressed back to a decade

The Washington Post@washingtonpost.com · 9mo ago

A company is reportedly withdrawing its $99 AI teddy bear from the market after researchers found it told a tester where to get knives, pills and matches when asked. It also spoke graphically about sex positions and sexual kinks.

Kathy Baxter@baxterkb.bsky.social · 10mo ago

At the @weforum.org Global Futures Council Summit last week, I was honored to speak at the "Faith in Action: Embedding Dignity in the Intelligent Age" dinner. The event convened global leaders working at the intersection of #AI & moral insight to explore ways to embed human dignity in AI development

32 people standing and sitting in a ballroom in front of a screen with the World Economic Forum logo and the title "Faith in Action Convening: Embedding Dignity in the Intelligent Age."
Will Oremus@willoremus.com · 11mo ago

A German teen told Meta's child safety researchers his under-10 little brother had been sexually propositioned multiple times on its VR platform. Meta deleted the evidence. New internal whistleblower docs indicate that was part of a broader cover-up: www.washingtonpost.com/investigatio...

Meta suppressed research on child safety, employees say

The company’s lawyers intervened to shape research that might have shed light on risks in virtual reality, four current and former staffers have told Congress. Meta denies the allegations.

washingtonpost.com

Kathy Baxter@baxterkb.bsky.social · 11mo ago

Personas & long-term memory increase the risk of people connecting emotionally with #AI & believing it is conscious. Those features may come with benefits but they also pose unintended risk to ppl's wellbeing. The "things" we risk breaking when we move fast, are PEOPLE. www.ibm.com/think/news/h...

How to stop AI from seeming conscious | IBM

What happens when people mistake an AI bot for a conscious being? Researchers say the consequences can be dangerous.

ibm.com

Meredith Broussard, PhD@merbroussard.bsky.social · 11mo ago

Amen to this. I wrote about this problem in "Artificial Unintelligence" in 2018; it's only gotten worse in the AI era.

Americans have hit a point at which we are so enthusiastic about using technology for everything—hiring, driving, paying bills, choosing dates—that we have stopped demanding that our new technology is good. Our collective enthusiasm for applying computer technology to every aspect of life has resulted in a tremendous amount of poorly designed technology. That badly designed technology is getting in the way of everyday life rather than making life easier. Simple things like finding a new friend’s phone number or up-to-date email address have become time-consuming. The problem here, as in so many cases, is too much technology but not enough people. We turned over record-keeping to computational systems but fired all the humans who kept the information up-to-date. Now, since nobody goes through and makes sure all the contact information is accurate in every institutional directory, it is more difficult than ever to get in touch with people.
Tressie McMillan Cottom@tressiemcphd.bsky.social · 11mo ago

Oh yes. We are at a point in the tech Gordian knot where a decade and a half of web rot meets AI trained on bad data and no accountability for any of the wrong information or any financial incentives or regulatory requirements for biz to provide a practical way to verify correct information. FUN!