Roya Pakzad

Evaluating Multilingual Guardrails in Humanitarian AI A technical evaluation of multilingual, context-aware AI guardrails, analyzing how English and Farsi responses are scored under identical policies. The findings surface scoring gaps, reasoning…

What if “safe” AI isn’t safe in every language?

@royapak.bsky.social and Daniel Nissani explore how AI safety guardrails behave outside English, and why language can affect fact-checking and risk detection in humanitarian settings.

See how the evaluation was done: link.mozilla.ai/multilingual...

A poster for the Digital Democracy from Below taking place on February 20, 2026, 12:00pm-1:30pm EST. Free and open to the public. Speakers are: Julie Bolthouse, Luã Cruz, Bianca Wylie, Tamara Kneese, Dhanaraj Thakur and Roya Pakzad. Co-organized by Jess Reia, Jonathan van Geuns and Nicole West Bassoff. Sponsored by the Digital Technology for Democracy Lab at UVA.

As we navigate the undermining of democratic processes & witness how tech facilitates/resists authoritarianism, join us on Feb 20@12pm to hear from @technolua.bsky.social @tamigraph.bsky.social @thakurdhanaraj.bsky.social @royapak.bsky.social B.Wylie, J.Bolthouse
dtdlab.virginia.edu/event/digita...

ICYMI, highly recommend this human rights impact assessment of @wikimediafoundation.org AI/ML work by @royapak.bsky.social (w/ @farzdusa.bsky.social and david liu)

Steve Shirley, Ann Moffatt, and co-worker Dee Shermer renting time on a mainframe, 1960s. Photo courtesy of Ann Moffatt. Black and white photo of three women behind a machine console.

Ann Moffatt writes software for the Concorde at her kitchen table as her baby daughter looks on. Moffatt's memoir, The IT Girl, details the story behind the photograph. Photo courtesy of Ann Moffatt.

I wrote about her and some of her company’s work in this chapter called “The Baby and the Black Box”: marhicks.com/writing/Hick...

AI Procurement — Taraaz New Report by Taraaz and CRCR: Key Considerations for Responsible AI Procurement in the Public Sector

New report: Key Considerations for AI Procurement in the Public Sector

Covers: Third-party dependencies • Large vs small vendors • AI demos/pilots • Master Service agreements • Performance-based contracts • Union engagement • Function creep • Model contracts

taraazresearch.org/ai-procurement

So great/helpful that is open sourced, thanks for sharing.

Yes, I’m interested: rpakzad@taraazresearch.org
Also interested in discussing multilingual inconsistencies and their trust & safety implications.

I missed your session at TrustCon, but just tested Zentropi for labeling human rights violations, works great in English! Have you considered multilingual support? Here’s a screenshot of the same text in EN and FA.
Also, formatting markdown for better legibility would be helpful.

🤦‍♀️ but congratulations, Wafa 🐣🎉

Less frustrated-sounding and more productive version 🙂
royapakzad.substack.com/p/guide-for-...

Many thanks, Juliet. I was very much looking forward to meeting you -- hopefully virtually or at Stanford's Trust and Safety Research conference if you are planning to go? :)

Many thanks for your kind words, Jen :) Looking forward to our further conversation. I hope the red teaming workshop went well.

Anyway, I guess I'm the "unprofessional" one! but I’m leaving #TrustCon with barely any connections made, regret and disappointment. But if you’re interested in my work on multilingual AI evaluation and human rights impact assessment, I’d be happy to connect: rpakzad@taraazresearch.org

I also don't recall seeing or hearing anything about a room for breastfeeding/pumping (I might be mistaken) or an area for those who, for any number of reasons in the world of babies and toddlers, need a space during a 3-day, 7:30am to 8:30pm event to satisfy their 24/7 parenting responsibilities.

Happy hours (for whom?!): All 5 to 8pm networking events are built around child-free evenings, sipping drinks assuming you have no dependents. How about sponsors who care deeply about child protection and wellbeing in T&S world creating child-inclusive events for parents to be able to network too!

Badging policy: My partner, here to help with our kids, was questioned couple of times when he *had to* bring them to me in one of the conference-designated areas. With close to $2000 ticket, couldn’t there be caregiver, color-coded badges to avoid this kind of policing and unwelcome encounter?

I know I'm not the only one. You can hear the parent crowd at 5 or 6am (Call of Duty time!) in lobbies doing childcare and conference prep. Yet, the conference logistics was designed as if they don't know anything about this world. Specific examples:

As a mother of a 3.5 and 1.5 year old (still breastfeeding), just deciding to go to a conference means juggling between caregiving, professional goals, emotional labor, etc. My options was to leave them at home for 4 days with military level planning or bring them with me. I chose the latter!

I'm at #TrustCon. Half the sessions are on tech companies' responsibilities on child safety (not Palestinian children though) and the well-being of trust and safety professionals, yet the conference design completely ignores the realities of working parents, especially for mothers. My experience 👇

Graphic for a CDT Research report, entitled “Content Moderation in the Global South: A Comparative Study of Four Low-Resource Languages.” An illustration of four different hands, all being touched by / wrapped around a bright blue connective internet cord.

Over 18 months, CDT studied content moderation in four low-resource languages—Maghrebi Arabic, Kiswahili, Tamil, and Quechua—across South Asia, North/East Africa, and South America. cdt.org/insights/con...

Happy to be a part of this cohort!

My fellowship project, "Equitable AI Benchmarking for Linguistic Diversity," helps civil society to engage in AI evaluation, focusing on multilingual inconsistencies with human rights impacts.

Automation on the Move - AlgorithmWatch Systems based on Artificial Intelligence (AI) and automated decision-making (ADM) are increasingly being experimented with and used on migrants, refugees, and travelers. Too often, this is done withou...

🚨 Great new project by Algorithm Watch lead by @fabiochiusi.bsky.social pulling together research, policy briefs, and coalition building on #bordertech in the EU - including a great list tracking private companies and research projects!

algorithmwatch.org/en/automatio...

My latest newsletter is about parenthood and GenAI. Read about: multilingual inconsistency in LLM generated responses, breastfeeding as a taboo in image generation tools, voice-cloned lullabies, embedded ads in LLM-based parenting advice, and more. open.substack.com/pub/royapakz...

Old Advocacy, New Algorithms: How 16th century "Devil's Advocates” Shaped AI Red Teaming The Humane AI newsletter has moved to Substack

I wrote about red teaming, from its origin in 16th-century Catholic Church to red teaming generative AI systems. Read here:

https://open.substack.com/pub/royapakzad/p/old-advocacy-new-algorithms?r=9cbk0&utm_medium=ios&utm_campaign=post