Upol Ehsan | hiring PhD students Fall'26

gpt5 writes like it considers mayonnaise is spicy.

and considers salt and pepper as spices.

Homies, you aren't getting it.

The math isn't wrong.

It's just GPT-5 being "PhD level"

[Pls repost] Hi folks, do you have any templates for reasonable accommodations letters written by your provider/doc for Long Covid?

I'm collecting samples to help the broader community, esp those who are underserved.

If you feel comfortable sharing, it will help a lot of people.

DMs open.

The Onion used to make jokes.
Now it does academic ethnography.

2️⃣ Audit for agency leakage:
Implement checks that flag when language oversells cognition and feed those signals into eval and procurement pipelines.

3️⃣ Localize the “mind”:
In Majority World settings, pair cautious tone with localized knowledge grounding so fluency does not paper over epistemic gaps

Takeaways for Responsible AI:

1️⃣ Design out miscues:
Prevent C-expressions in sensitive applications.
Use tone constraints and capability disclaimers to fight unearned trust.

Reflections:
⚠️ We are building interfaces that invite over-trust. “C-expressions” + anthropomorphic branding = safety debt waiting to mature.

⚖️ The real risk isn’t whether LLMs are sentient; it’s that people think they are, and then absolve responsibility when things go wrong.

3️⃣ Theory-of-Mind misfires at scale
When LLMs speak fluently, our brains project beliefs and goals.
But LLMs do not “believe”... they autocomplete.

🚨 That delta between performance and cognition creates false agency perception, creating room for irresponsible AI.

2️⃣ Verbal uncanny valley is real
Unsettling mismatch: humanlike language + inhuman reasoning = a new flavor of uncanny valley (no creepy avatars needed). This has UI, trust calibration, and regulatory implications.

Findings (worthy of my time)

1️⃣ “C-expressions” ≠ consciousness
Phrases like “I panicked” or “I feel” trigger reflexive mind-ascription, but they’re linguistic hallucinations, not signs of sentience. These cues silently build user trust without grounding in system capacity.

The psychology of LLM interactions: the uncanny valley and other minds Through their ability to converse in a human-like fashion, large language models (LLMs) have underscored the need to revisit our definitions of consciousness, and how we know if someone who claims ...

🧠 LLMs don’t think; they 𝘱𝘦𝘳𝘧𝘰𝘳𝘮 thinking. Treating performance as proof is how we fall into the 𝐯𝐞𝐫𝐛𝐚𝐥 𝐮𝐧𝐜𝐚𝐧𝐧𝐲 𝐯𝐚𝐥𝐥𝐞𝐲.

Enjoyed reading this paper: www.tandfonline.com/doi/full/10....

Findings from the paper that were worth my time, reflections, and takeaways ↓

Imagine this being your opening line. Reads like someone asked ChatGPT: “Write a DM to an expert but make it weird.” what a time to be alive.

“So you’d rather segregate all of these people from society instead of mildly inconveniencing yourself?”🫠

The fight to get AI right in Bangladesh, the world’s eighth-most populous country  - Khoury College of Computer Sciences When Khoury College Assistant Professor Upol Ehsan sat for an AI-focused interview with Channel24, one of the largest TV channels in his native Bangladesh, he did not anticipate what would […]

Backstory: “AI in Bangladesh is a side quest,” they dismissed. I disagreed. 170 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝗳𝘂𝘁𝘂𝗿𝗲𝘀 𝙖𝙧𝙚 𝘁𝗵𝗲 𝗺𝗮𝗶𝗻 𝗾𝘂𝗲𝘀𝘁.

Glad to see Northeastern University spotlighting our work on Majority-World #AI governance and why we must act now"

#academicsky #RAI

www.khoury.northeastern.edu/the-fight-to...

The fight to get AI right in Bangladesh, the world’s eighth-most populous country  - Khoury College of Computer Sciences When Khoury College Assistant Professor Upol Ehsan sat for an AI-focused interview with Channel24, one of the largest TV channels in his native Bangladesh, he did not anticipate what would […]

Backstory: “AI in Bangladesh is a side quest,” they dismissed. I disagreed. 170 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝗳𝘂𝘁𝘂𝗿𝗲𝘀 𝙖𝙧𝙚 𝘁𝗵𝗲 𝗺𝗮𝗶𝗻 𝗾𝘂𝗲𝘀𝘁.

Glad to see Northeastern University spotlighting our work on Majority-World #AI governance and why we must act now"

#academicsky #RAI

www.khoury.northeastern.edu/the-fight-to...

And if you have a Bangladeshi colleague or friend, check in on them. The grief is not confined to one place. It has rippled across the world and deep into the diaspora.

Bangladesh jet crash: What we know about the deadly crash into a Dhaka school The crash, which killed 27, marks the deadliest aviation disaster the country has seen in decades.

Having a hard time processing this: A fighter jet crashed into a school in Bangladesh. Most of the injured and deceased are children.

This tragedy has left me speechless. Please keep these children and their families in your thoughts and prayers.

www.bbc.com/news/article...

Lol it me

The problem isn’t that ChatGPT uses em dashes. It’s that so many humans still don’t

Or just sheer panic

Indeed. Good times with good friends. Zero work done.

When you and your friends finally schedule that writing retreat...and write nothing. Just riding on vibes.

tell me how you are a pro at social media without telling me you are a pro at social media.

lol what a gigachad move.

just heard someone compare AI-assisted work similar to a (wait for it) a calculator.

give me your spiciest clapbacks to this horrible analogy.

winner gets a cutest cat pic.

Where CS/NSF/Big Tech went wrong is that we decided focusing on the future of work -- not the future of WORKERS -- was the right way to go.

It normalized a view of nowhere.

A reality where "I fired all my employees and replaced them with AI" is a flex not a shame.

Not sure if his immigration was "legal" either.
He is def an alien.
Yet, he is one of the most American superheroes.
Therein perhaps lies the point.

Best explanations of why N95 masks work. Hands down.
#covid19 #cleanair

“No rush on this” is academic for:
I needed it yesterday, I am panicking, and I am trying to sound chill.

Can we expense it under professional development funds? 😅

Academic mentorship is 90% helping someone else navigate a system you barely survived yourself.