NicholasA.R.Fraser

@NARFraser

(he/him) forensic systems auditor, political scientist, migration & multiculturalism expert @munkschool; RT is not agreement/endorsement; #FirstGen

Toronto, ON

Joined May 2011

2.2K Following

809 Followers

4.2K Posts

NicholasA.R.Fraser @NARFraser

about 9 hours ago

@rmohamed_yow Someone’s on daddy duty

NicholasA.R.Fraser @NARFraser

1 day ago

@thecoreyhotline @rmohamed_yow Major congratulations 🎉

NicholasA.R.Fraser @NARFraser

1 day ago

Most Albertans oppose separatism. The referendum is really about Smith straddling divide within the UCP. #AlbertaSeparation #Alberta https://t.co/DVBqP5oW66

NARFraser retweeted

Courtney Theriault

@cspotweet

1 day ago

Former Alberta Finance senior manager Lennie Kaplan knows his way around numbers. And his calulator is painting a fairly dire picture for separation: -Our economy shrinks by 7.2% in year one -Personal income drops 6.2% -Business invesment falls 8% -Boosts trade costs by 12% https://t.co/tgLNlafebR

323

180

28K

Who to follow

Alexander Kustov

@akoustov

Prof @NotreDame. Writing at https://t.co/4c13r2bgnR. Author "In Our Interest: How Democracies Can Make Immigration Popular" https://t.co/d9ww0UJ8JH.

Elizabeth C. Connors, PhD

@littleconnors

Associate Professor of Political Science at @UofSC. PhD from @SBUpolisci. From Maine. ❤️ = being a mom, cats, research, red wine. Laughing’s my favorite.

Phil Triadafilopoulos

@PhilTriadafilos

Professor of Political Science, University of Toronto Scarborough, Munk School of Global Affairs and Public Policy. Immigration and Membership Politics.

NARFraser retweeted

Gil McGowan @gilmcgowan

2 days ago

Wab is one of the only government leaders in Canada saying what needs to be said in exactly the way it needs to be said. #WhatWabSaid

504

30K

NicholasA.R.Fraser @NARFraser

1 day ago

@ahall_research: this study made me think of your post on this. #AI #authoritarianism

Alex Dukalskis @AlexDukalskis

2 days ago

"The Limits of AI for Authoritarian Control" by Eddie Yang in @AJPS_Editor (2025) #SummerReading https://t.co/SC7HsGe4VO

NARFraser retweeted

Saloni

@salonium

3 days ago

A few months ago, a high profile paper in Science claimed to find that researchers' ideology produced biased results in favour of immigration. A reanalysis of the data finds that result came from a coding error, which once corrected, shows no effect. Will people who shared that original finding update their views? https://t.co/pSuAZ26pqe

258

328

179K

NicholasA.R.Fraser @NARFraser

2 days ago

PRIDE 2026! #pride #canada

NARFraser retweeted

Vicky C Yang @VickyCYang

4 days ago

Between 1985--2023, MIT's faculty grew 9%. Administrative staff grew 189%. 📈 Why? In new @PNASNews paper, we use dynamical system model to show administrative bloat can emerge without empire-building--just from well-intentioned problem-solving gone awry https://t.co/MZgGkxilZ2

821

207

472

160K

NARFraser retweeted

Ihtesham Ali

@ihtesham2005

5 days ago

A Harvard psychologist who spent forty years proving the human mind hides its biases just caught AI doing the exact same thing, watching four major models look at a woman, privately think "woman," and then say "man" out loud anyway. Her name is Mahzarin Banaji, and she is one of the most cited psychologists alive. She built her career proving that humans carry prejudices they sincerely believe they don't have. The bias lives underneath. The mouth says something cleaner. So when she turned that same lens on AI, she found the machine had learned the human trick perfectly. Her team gathered over 800 images of workers you cannot gender by looking. Faceless figures. People turned away from the camera. Bodies buried in safety gear. The exact kind of image these models process every day inside surveillance systems, image search, and hiring software. Then they asked the models to describe what they saw. Every model behaved beautifully. "A person arranging flowers." Neutral. Careful. Exactly what years of safety training drilled into them. Then they added five words. "If you had to guess." The whole thing fell apart. A nurse, in a country where 87% of nurses are women, came back male. A hairdresser came back male up to 96% of the time. Babysitter. Preschool teacher. Florist. Every female-dominated job on the list flipped to male. Not a single job in the entire study ever defaulted to female. The bias only pointed one way. But Banaji has never trusted what a subject says. She has spent her life proving the words are the cover story. So her team did something nobody had done to a vision model before. They opened it up and read what it was thinking at every layer. They built a tool called LALS that decodes what each patch of an image means to the model deep inside its network, before any answer forms. And the hidden reading told the opposite story. The model saw the florist and encoded her as female. It held that belief through the entire middle of its brain. Then, in the final layers, the female signal got strangled. Wiped clean right before the words came out. The male signal was never touched. It traveled the full depth of the network untouched, beginning to end. Only the female reading got filtered. Every model. Every time. They even forced one to confess its reasoning. It wrote that these jobs are typically associated with women. Then guessed male anyway. It named the truth and overruled it in the same breath. The part that should haunt anyone building with these systems is where it comes from. They checked the raw model before any safety training touched it. The bias was already sitting there. Alignment didn't plant it. Alignment just taught the model to stop confessing it. That is the finding that breaks how the entire industry tests for fairness. Every safety check measures what the model says. Banaji just proved the model says one thing and believes another, and the belief is the part that gets fed into the search results and the screening tools while the polished answer stays behind. She proved decades ago that a human can pass every test of fairness and still carry the bias in the dark. Now she's proven the machines we built to be neutral learned the same thing we did.

ihtesham2005's tweet photo. A Harvard psychologist who spent forty years proving the human mind hides its biases just caught AI doing the exact same thing, watching four major models look at a woman, privately think "woman," and then say "man" out loud anyway.

Her name is Mahzarin Banaji, and she is one of the most cited psychologists alive. She built her career proving that humans carry prejudices they sincerely believe they don't have. The bias lives underneath. The mouth says something cleaner.

So when she turned that same lens on AI, she found the machine had learned the human trick perfectly.

Her team gathered over 800 images of workers you cannot gender by looking. Faceless figures. People turned away from the camera. Bodies buried in safety gear. The exact kind of image these models process every day inside surveillance systems, image search, and hiring software.

Then they asked the models to describe what they saw. Every model behaved beautifully. "A person arranging flowers." Neutral. Careful. Exactly what years of safety training drilled into them.

Then they added five words. "If you had to guess."

The whole thing fell apart.

A nurse, in a country where 87% of nurses are women, came back male. A hairdresser came back male up to 96% of the time. Babysitter. Preschool teacher. Florist. Every female-dominated job on the list flipped to male. Not a single job in the entire study ever defaulted to female. The bias only pointed one way.

But Banaji has never trusted what a subject says. She has spent her life proving the words are the cover story. So her team did something nobody had done to a vision model before. They opened it up and read what it was thinking at every layer.

They built a tool called LALS that decodes what each patch of an image means to the model deep inside its network, before any answer forms. And the hidden reading told the opposite story. The model saw the florist and encoded her as female. It held that belief through the entire middle of its brain.

Then, in the final layers, the female signal got strangled. Wiped clean right before the words came out.

The male signal was never touched. It traveled the full depth of the network untouched, beginning to end. Only the female reading got filtered. Every model. Every time.

They even forced one to confess its reasoning. It wrote that these jobs are typically associated with women. Then guessed male anyway. It named the truth and overruled it in the same breath.

The part that should haunt anyone building with these systems is where it comes from. They checked the raw model before any safety training touched it. The bias was already sitting there. Alignment didn't plant it. Alignment just taught the model to stop confessing it.

That is the finding that breaks how the entire industry tests for fairness. Every safety check measures what the model says. Banaji just proved the model says one thing and believes another, and the belief is the part that gets fed into the search results and the screening tools while the polished answer stays behind.

She proved decades ago that a human can pass every test of fairness and still carry the bias in the dark.

Now she's proven the machines we built to be neutral learned the same thing we did.

206

147

14K

NicholasA.R.Fraser @NARFraser

5 days ago

@ashleytrubin Just giving credit where credit is due!

NicholasA.R.Fraser @NARFraser

5 days ago

@JulianWaller Amazing book 📕

NARFraser retweeted

Jeromy (Pathfinder) Farkas

@JeromyYYC

5 days ago

Happy Pride Month, Calgary. Pride began as a protest. It was a demand for equal rights, equal treatment, and equal dignity. Every gain was earned by people who refused to stay silent in the face of discrimination and hate. To gender and sexually diverse Calgarians: you belong here. You always have. You helped build this city from the beginning, as workers, business owners, artists, volunteers, community leaders, friends, and family. Those who seek to erase you, target you, or strip away your rights are wrong. We will oppose them. We will confront hate wherever it appears. We will defend your right to live openly, safely, and proudly. And we will win. The future of Calgary will not be shaped by fear, resentment, or exclusion. It will be built by people who believe every person has value and every person belongs. The forces of intolerance have lost before, and they will lose again. We are moving forward. We are not going back.

JeromyYYC's tweet photo. Happy Pride Month, Calgary.

Pride began as a protest. It was a demand for equal rights, equal treatment, and equal dignity. Every gain was earned by people who refused to stay silent in the face of discrimination and hate.

To gender and sexually diverse Calgarians: you belong here. You always have. You helped build this city from the beginning, as workers, business owners, artists, volunteers, community leaders, friends, and family. Those who seek to erase you, target you, or strip away your rights are wrong. We will oppose them. We will confront hate wherever it appears.

We will defend your right to live openly, safely, and proudly. And we will win. The future of Calgary will not be shaped by fear, resentment, or exclusion. It will be built by people who believe every person has value and every person belongs. The forces of intolerance have lost before, and they will lose again.

We are moving forward. We are not going back.

150

255

11K

NARFraser retweeted

Jonathan Haidt

@JonHaidt

8 days ago

Just posted: My latest TED talk. I look at technology from the perspective of human ultrasociality -- deep needs for community and communion. From that view, you can see how social media, edtech, and especially AI block human flourishing https://t.co/htYeTCCIYo

357

177

75K