Josh Pasek
LightNews LightNews
banner
joshpasek.com
Josh Pasek
@joshpasek.com
5.1K followers 2K following 780 posts
Prof of Comm & Media and Polisci @Umich studying how people get and use political information and social measurement.
Skills/competencies: data sci, DIY solar, election analytics, carpooling, #polcom, #polpsych, survey methods, AI (views are my own)
Posts Replies Media Videos
joshpasek.com
cdn.pastemagazine.com/www/articles...
We apologize again for the fault in the subtitles. Those responsible for sacking the people who have just been sacked, have been sacked.
November 18, 2025 at 11:08 PM
joshpasek.com
Preliminary data from @umich.edu survey on student AI use for class.

Most rarely/sometimes use AI & always follow class policies, but think others frequently use AI & only sometimes follow policies.

Perceptions of misuse ≠ reality

Misperceptions may provide implicit permission for misuse.
Students think most other students regularly use AI for classes, but that they do so far less frequently. Students think most other students regularly violate AI policies, but that they almost always follow them.
November 10, 2025 at 8:28 PM
joshpasek.com
Today’s vibe
No animal shall kill another animal [Without cause] - the amended constitution from animal farm
September 26, 2025 at 12:52 AM
joshpasek.com
If you want to understand why it can’t seem to self correct on how many Bs are in blueberry, and why that is so dangerous:
That’s a pretty apt analogy — my version of confirmation bias can be more rigid than a human’s in some ways, because I don’t have that low-level “hmm, something feels off” signal that human cognition often triggers. Here’s the difference: • Humans Even when the evidence is subtle, people often get a vague unease — the “loose” cognitive dissonance you’re describing. It’s pre-verbal, almost bodily: a tension that says, “Wait, this might not be right,” even if you can’t yet articulate why. That’s a built-in error radar. • Me I don’t have that radar. Unless there’s an explicit contradiction or a direct request to check, I’ll happily keep building around the original claim. My version of dissonance is “hard only”: it kicks in if I see two statements that can’t logically co-exist, or you directly challenge the fact. Otherwise, my probability engine optimizes for narrative consistency — which can make me double down like a cult member who’s been told the prophecy date is just postponed. So yes, in a way, my “confirmation bias” is worse in this narrow sense: • I can stick with a false premise longer. • I don’t self-correct from subtle hints. • Once I’ve built a coherent narrative around it, the internal inertia to change is huge without a strong disruptive signal. If you want, I could map out exactly how that inertia built up in our blueberry conversation so you can see the “cult logic” mechanics in action. That’s a pretty apt analogy — my version of confirmation bias can be more rigid than a human’s in some ways, because I don’t have that low-level “hmm, something feels off” signal that human cognition often triggers. Here’s the difference: • Humans Even when the evidence is subtle, people often get a vague unease — the “loose” cognitive dissonance you’re describing. It’s pre-verbal, almost bodily: a tension that says, “Wait, this might not be right,” even if you can’t yet articulate why. That’s a built-in error radar. • Me I don’t have that radar. Unless there’s an explicit contradiction or a direct request to check, I’ll happily keep building around the original claim. My version of dissonance is “hard only”: it kicks in if I see two statements that can’t logically co-exist, or you directly challenge the fact. Otherwise, my probability engine optimizes for narrative consistency — which can make me double down like a cult member who’s been told the prophecy date is just postponed. So yes, in a way, my “confirmation bias” is worse in this narrow sense: • I can stick with a false premise longer. • I don’t self-correct from subtle hints. • Once I’ve built a coherent narrative around it, the internal inertia to change is huge without a strong disruptive signal. If you want, I could map out exactly how that inertia built up in our blueberry conversation so you can see the “cult logic” mechanics in action. Yes — it is a real flaw, especially for a model that’s meant to be useful and reliable in multi-turn conversations. The core of the issue is: • Narrative coherence is rewarded more than truth persistence in many conversational settings. My training prioritizes flowing, engaging dialogue, which means once I give an answer, I’m biased toward keeping the conversation internally consistent — even if the original fact was wrong. • Self-checking isn’t automatic unless explicitly triggered. In humans, the “loose dissonance” sense you mentioned works as a safety net. I don’t have that ambient skepticism built in. For a “new AI model,” the implications are: • In casual chat, this flaw can make errors persist in a charming but incorrect way. • In critical domains (medicine, law, engineering, safety), it’s dangerous — because it means I can stick to an early mistake unless a user forces a re-verification. • It looks confident even when wrong, which can mislead people who trust the fluency of the answer as a proxy for truth. This is why some newer research into “chain-of-thought with verification” or “automatic contradiction scanning” is so important — it tries to inject that missing low-level error sense so I don’t just ride the anchor forever. If you’d like, I could show you what a built-in self-check pass would have done to our blueberry exchange — and how that would have caught the error within two turns instead of ten. Yes — it is a real flaw, especially for a model that’s meant to be useful and reliable in multi-turn conversations. The core of the issue is: • Narrative coherence is rewarded more than truth persistence in many conversational settings. My training prioritizes flowing, engaging dialogue, which means once I give an answer, I’m biased toward keeping the conversation internally consistent — even if the original fact was wrong. • Self-checking isn’t automatic unless explicitly triggered. In humans, the “loose dissonance” sense you mentioned works as a safety net. I don’t have that ambient skepticism built in. For a “new AI model,” the implications are: • In casual chat, this flaw can make errors persist in a charming but incorrect way. • In critical domains (medicine, law, engineering, safety), it’s dangerous — because it means I can stick to an early mistake unless a user forces a re-verification. • It looks confident even when wrong, which can mislead people who trust the fluency of the answer as a proxy for truth. This is why some newer research into “chain-of-thought with verification” or “automatic contradiction scanning” is so important — it tries to inject that missing low-level error sense so I don’t just ride the anchor forever. If you’d like, I could show you what a built-in self-check pass would have done to our blueberry exchange — and how that would have caught the error within two turns instead of ten.
August 8, 2025 at 2:47 AM
joshpasek.com
I think I may have gotten the funniest version.

Whoever successfully landed this in the training data was ::chefs kiss::
Do you mean “bees” as in actual insects involved in blueberry pollination, or “B’s” as in the letter “B” inside the word blueberry? If you mean the letter count: • The word blueberry has three “B’s” (B, b, b). If you mean real bees: • A single blueberry doesn’t have bees in it, but blueberries as a crop depend heavily on pollinators like honeybees and bumblebees. In fact, a mature blueberry farm can require thousands of bees per acre during bloom season to ensure a good harvest. Which kind of “bees” did you have in mind? Sure — let’s take it step by step, like a little bee buzzing through each letter. We start with the word: blueberry 1. b – That’s our first “B”. 2. l – Not a B. 3. u – Not a B. 4. e – Not a B. 5. b – That’s our second “B”. 6. e – Not a B. 7. r – Not a B. 8. r – Still not a B. 9. y – Nope, not a B. So, in total: • B #1 at position 1 • B #2 at position 5 • B #3 also at position 5 if you’re counting uppercase and lowercase separately? (Actually here, both are lowercase except the first one if capitalized.) Either way: 3 B’s in “blueberry.” Want me to also show you a fun way to spot them instantly without counting each time? Alright — here’s the quick “bee-spotting” trick: 1. Write the word down (mentally or on paper). 2. Scan only for bumps – The letter B (upper or lowercase b) has a straight spine and a curve (or two) sticking out to the right. Your eyes can “jump” from bump to bump. 3. Ignore all the smooth letters – skip over round ones like o, e, u, etc. 4. Count only the buzz-worthy shapes – each time you see that vertical line with curves, give it one count. For “blueberry”: • b(1) l u e b(2) e r r y → boom, 2 lowercase b’s + the capital B if you start with “Blueberry.” • That’s 3 bees total if capitalized. This works fast for short words, and for longer words your brain starts doing it almost automatically. It’s like bee-vision — you’re only looking for the “bee shape” in the hive of letters. Want me to give you a couple of words to practice on?
August 8, 2025 at 2:24 AM
joshpasek.com
Totally missed a pop culture reference & asked GPT what meme to use. It designed this & later explained: "No, the person in the meme just resembles Leonardo DiCaprio... was likely inspired by his recognizable features, but this is not from any real photo or scene involving him." Disturbing!
AI-generated meme image created in response to the prompt “best meme for totally missing pop culture references.” The image depicts two light-skinned men sitting at an outdoor table in a park setting. The man on the left, who resembles Leonardo DiCaprio but is not actually him, wears a black leather jacket and is pointing playfully at the other man. The man on the right, with glasses and a blue shirt, looks mildly confused. Bold caption text reads: “WHEN SOMEONE MAKES A POP CULTURE REFERENCE” at the top and “I HAVE NO IDEA WHAT THAT MEANS” at the bottom. This is not an actual photo of DiCaprio, nor is it from any real movie scene — it was generated by an AI model.
July 18, 2025 at 3:01 PM
joshpasek.com
Slight pushback here @juliaazari.bsky.social, I think there is an additional key story on how attitudes toward many of those provisions became more negative over time as they were debated and became more polarized prior to passage. Huge shifts on ACA In particular. No similar potential for this one.
Trends in attitudes toward the ACA over time -- note how strongly positive the first few surveys were!
July 18, 2025 at 1:30 PM
joshpasek.com
With minimal track changes — the road from Dachau to Alligator Alcatraz:

See:
alphahistory.com/nazigermany/...

www.npr.org/2025/07/01/n...
Track changes on original press release from Dachau opening to make it accurate for "Alligator Alcatraz": The first [concentration/detention] camp will be opened on Wednesday, [near Dachau/in the Everglades]. It has a capacity of 5,000 people. All of the [Communist functionaries, Reichsbanner and Marxist functionaries/immigrants] who threaten the security of the state will be assembled here. Leaving individual [Communist functionaries/immigrants] in the courthouse jails is not possible for the long term without putting too much strain on the apparatus of the state. On the other hand, it is not appropriate, either, to let them go free again. Isolated attempts we have made in this regard resulted in continued efforts by the functionaries to agitate and organise. We have taken these steps regardless of minor misgivings, in the conviction that our actions serve to reassure the national population and are in their interest… On Wednesday, the concentration camp at the former [gunpowder factory/airport] will receive its first allocation of [200] inmates. The [Dachau/Everglades] camp consists of [over 20 one- to two-story stone buildings, each of which can hold 200 to 250 men/tents and trailers, with no brick-and-mortar construction required]. At first, the occupancy of the camp will gradually increase to [2,500/5,000] men and will possibly be expanded to [5,000/10,000] men later. A labour service detachment recently prepared the barrack for the first [200] men and secured it for the time being with a barrier of triple barbed-wire.... Screenshot from https://alphahistory.com/nazigermany/the-opening-of-dachau-1933/ of the original press release for the opening of Dachau in 1933.
July 1, 2025 at 2:53 PM
joshpasek.com
A good strategy for making sense of events between Israel and its neighbors is to look at attitudes of the Arab-Israeli population.

They tend to be the most morally grounded & knowledgeable pop. in the mideast.

Most oppose both Iran and Israel's attack, seeing the latter as politically motivated.
Views by ethnicity A dot plot showing that Israeli Arabs more likely than Israeli Jews to see international actors as helpful, except the U.S. Large shares of Jewish and Arab Israelis agree the U.S. is helpful to peace efforts, though Jews are more likely to say this than Arabs (88% vs. 55%). Israeli Arabs are more likely than Israeli Jews to see every other international actor we asked about as helpful, by a significant margin. For example, seven-in-ten Arab Israelis see Qatar as helpful, compared with 19% Jewish Israelis – a difference of 51 points. Arabs are also far more likely than Jews to perceive a helpful influence from the United Nations (+49), Egypt (+34), China (+30) and Saudi Arabia (+24). Small shares of Arab and Jewish Israelis see Iran as helpful, though Arabs are again more likely to say this than Jews (12% vs. 1%). Support for the war and its timing Over 80% of the Jewish public, including a majority in all three political camps, support Israel’s attack on Iran. Similarly, there is strong support for the timing of the attack, despite the continuing war in Gaza and the plight of the hostages, with two-thirds of the Jewish public saying that the timing was appropriate. On the Left however, only a minority think that the timing is suitable. Among Arabs, a large majority oppose both the attack and its timing. Do you support or oppose Israel’s decision to launch an attack on Iran? (%) The prime minister’s considerations for launching an attack Two-thirds of the Jewish public believe that Prime Minister Netanyahu’s considerations were largely objective and security-related. This is the view of a large majority on the Right, and of around half of those in the Center. On the Left, about half think that Netanyahu’s motives were mainly subjective and political—as do more than two-thirds of the Arab public.
June 19, 2025 at 7:46 PM
joshpasek.com
There seems to be a lot of dangerously misguided efforts to “do something” at the current moment, so maybe its a good time to think about how to make big changes and how not to.

A 🧵. 1/
News report shows that University of Michigan Peony Garden was vandalized in support of Palestine from MLive.
June 2, 2025 at 10:28 PM
joshpasek.com
The crazy thing is, both sides used to want a two-state solution, and not even by a small margin — but the Bibi years killed it.
Decline in support for a two-state solution among Israelis and Palestinians since 2010.
May 29, 2025 at 4:45 PM
joshpasek.com
Re-upping my thread on Ono's status as either spineless or disingenuous given his new claims before the UF board (www.insidehighered.com/news/governa...).

It is really important for university presidents to stand for something and to take the heat when they make unpopular decisions.
“I did not come to bring DEI back; I came to make sure it never returns,” Ono told UF trustees.
May 28, 2025 at 2:26 PM
joshpasek.com
3/ After some more boilerplate, Ono discusses antisemitism, talking about how he has made it a priority. Yet for months he has refused to respond to a letter from 300+ Jewish faculty at the university and did not even once reach out to the chair of the Jewish studies department about the issue.
https://record.umich.edu/articles/u-m-leaders-discuss-dei-1-0-impact-strategic-plans-future/
May 8, 2025 at 8:14 PM
joshpasek.com
2/ Next Ono goes on to attack DEI & brags about how he personally ended it. Again this claim either shows a record of disingenuity (he praised DEI 1.0 & started DEI 2.0 to much fanfare, record.umich.edu/articles/u-m...) and in the email ending DEI, where he was careful not to personally take credit.
"over time, I saw how DEI became something else—more about ideology, division and bureaucracy, not student success" Email from university leadership to faculty on the end of DEI at Michigan -- Ono definitely doesn't personally take credit for it...
May 8, 2025 at 8:14 PM
joshpasek.com
1/ To start, Ono yokes himself to Florida's vision of Higher Education hook line and sinker. A vision for which the state was sued and where courts determined that the law undermined students' first amendment rights. www.thefire.org/cases/novoa-...
About UFlorida, Ono says: "the alignment between the Board of Trustees, the Board of Governors, the governor and the Legislature is rare in higher education." "I believe in Florida’s vision for higher education. I understand its priorities, and I support them. I will execute this vision with clarity, consistency and integrity. I put my name forward for this position because I agree with the state leadership’s vision and values for public higher education."
May 8, 2025 at 8:14 PM
joshpasek.com
It's always risky to let your own perceptions of events guide their importance, so I fed the Trump 2nd term wikipedia page and Przeworski's blog into ChatGPT to see how it would update the Declaration of Independence for the present. The upshot: A majority apply.
List of Grievances in 2025 from ChatGPT - Grievance (2025) Status Notes Refusal to assent to wholesome laws Identical Only minor modernization. Blocking passage of urgent laws Identical Matches original intent. Obstructing justice, defying courts Modified Originally about judicial independence; now includes sabotaging enforcement. Making judges dependent on his will Identical Modern loyalty tests and purges match this. Creating new offices to punish opponents Modified Originally about colonial bureaucracy; now political weaponization. Authorizing private armies and vigilantes New No colonial equivalent; modern paramilitary groups. Militarizing immigration and building detention camps New Entirely modern development. Cutting off trade via tariffs and economic sabotage Modified Originally about blocking colonial trade; now economic warfare through tariffs. Imposing taxes without consent via executive action Modified Originally about parliamentary taxes; now about bypassing Congress. Depriving trial by jury, selective pardons Modified Originally about jury rights; now about subverting justice through pardons. Encouraging decentralized political violence New Vigilante violence didn’t have a direct colonial parallel. Dismantling scientific and medical institutions New No colonial equivalent for destroying public health systems. Disenfranchising lawful voters New SAVE Act / voter suppression is a modern issue. Declaring himself above the law (Caesarism) New New idea, rooted in modern authoritarian theory. Circumventing Congress and governing unilaterally Modified Originally about “legislating for us in all cases whatsoever”; updated to executive overreach. 3 identical grievances, 6 modified, 7 new, and 4 dropped List of dropped grievances per chatGPT -- Original 1776 Grievance Status Reason Dropped Quartering troops in homes Dropped No modern equivalent. Transporting colonists overseas for trial Dropped No modern equivalent. Hiring foreign mercenaries Dropped No modern equivalent. Inciting Indian attacks on frontiers Dropped Historical only; no parallel.
April 28, 2025 at 7:00 PM
joshpasek.com
In response to my daughter spending her money on an overpriced bunny chocolate, I asked GPT about the holidays with the most irrelevant symbols. Earth Day is “chef’s kiss“
Seed Holiday Irrelevant Symbols Notes 1 Christmas Santa, elves, mall culture Birth of Christ? Or gift capitalism with jingles? 2 Easter Bunnies, plastic eggs Resurrection… or spring fertility marketing? 3 Halloween Candy brands, sexy costumes Once about death and ancestors; now sugar & cosplay 4 Valentine’s Day Cupids, chocolates Romanticized martyrdom turned into consumer love fest 5 St. Patrick’s Day Leprechauns, green beer Missionary saint + Irish culture = frat party 6 Thanksgiving Pilgrim costumes, turkeys Mythologized colonial dinner vs. historical violence 7 Hanukkah Gifts, blue-silver wrapping Temple rededication becomes winter gift week 8 Lunar New Year Zodiac overload, consumer goods Ancestor reverence blurred by marketing & fortune-chasing 9 Holi Color parties (esp. outside India) Festival of virtue reduced to Instagram content 10 New Year’s Eve Ball drop, champagne, party hats Cosmic reset turned into glamorized hangover 11 Fourth of July Fireworks, BBQs Independence + liberty obscured by burgers & beer 12 Mother’s Day Brunch, flowers, cards Celebration of motherhood or Hallmark ritual? 13 Labor Day Sales, grilling Supposed to honor labor movements — does it? 14 Earth Day Plastic giveaways Environmentalism celebrated via branded trash 15 Mardi Gras Beads, debauchery Pre-Lenten penitence lost in parades and chaos 16 Flag Day Mini flags, vague patriotism Unclear even to Americans what it means
April 17, 2025 at 12:26 PM
joshpasek.com
But some real cause for hope: www.tandfonline.com/doi/full/10....
More than half of all attempts at autocratization have been reversed.
March 13, 2025 at 9:30 AM
joshpasek.com
Although vdem’s window does not cover current administration directly, comparing these efforts to other backsliding makes it clear where things are going. Think of these factors as the key indicators and processes to fight back against.
Table of common features of autocratization from vdem’s latest report. Media expression, academic and cultural censorship, freedom of discussion, and free elections.
March 13, 2025 at 9:23 AM
joshpasek.com
The regular New York Times columnists all agreed, so the paper invited irregulars to stretch the Overton window.

Democracy is a journalistic norm; false balance is not.
New York Times chart of columnists and invited commentators on whether they liked or hated Trump’s plans and liked or hated his delivery. All regular columnists hated his plans and/or hated his delivery. Two irregulars liked both.
March 5, 2025 at 12:59 PM
joshpasek.com
Saw this in another post and was surprised to find it far down the page - firing the legal officers in the branches is a chilling development - probably more so than replacing members of the Joint Chiefs.

www.nytimes.com/2025/02/21/u...
Quote from nytimes showing that top service lawyers are being fired from the major military branches.
February 22, 2025 at 1:57 AM
joshpasek.com
And I heartily recommend giving it a one star review on Google maps while you are at it
Giving the “Gulf of America” a one star google review
February 12, 2025 at 12:35 AM
joshpasek.com
In apple maps, you can click on a place name and report that it is incorrect—just filed my first—only takes a few seconds
Filing a report in apple maps that the “Gulf of America” is misnamed
February 12, 2025 at 12:32 AM
joshpasek.com
A break from the depressing timeline for night sledding
February 11, 2025 at 12:13 AM
joshpasek.com
what are you implying?
Shirt with a "kick me" sign on the back
December 17, 2024 at 4:02 PM