Incoming assistant professor of psychology at Manhattan University. I study the psychology of communication in digital spaces.
Joined November 2008
- Tweets1.1K
- Following748
- Followers689
- Likes792
Pinned Tweet
Why do people see the same thing and disagree on what they saw? We presume our own interpretation is the obvious and correct one. But undetected misinterpretations may be an underappreciated factor online and could warp perceived consensus:
bit.ly/49jbyZa
Rebecca Dolgin retweeted
1/n Within 15 months of publication, most of the LLM claims we tested either no longer held or showed substantially diminished effects. How long does our knowledge of AI hold up? In our new study, we revisit 50 findings to see what fades, what persists, and what gets stronger. 🧵
Rebecca Dolgin retweeted
In this new paper, we make a strong point:
Human imitation is a misplaced goal for AI.
Too much of AI development is based on human imitation, as exemplified by the Turing test: the more a machine resembles us, the more successful it is.
But there are two major problems.
The first is conceptual. Imitating human outputs does not tell us much about the capacities behind them:
The same answer can emerge from different processes. A plausible response does not tell us whether a system can reason reliably, generalize to unfamiliar situations, or recognize when it should say “I don’t know”.
The second is ethical. Why should we want AI that imitates humans?
Humans are biased. Our attention is limited. Our judgment is fallible. Reproducing these limitations is hardly an ambitious vision for AI.
Even worse:
A system that impersonates a person can exploit human trust.
And an industry focused on replicating humans can fall in the “Turing’s trap”: favour human replacement over augmentation.
Instead, we argue that AI should aim to improve human welfare.
To this end, we propose three coordinated changes:
For developers: Build systems that strengthen human capabilities and independent judgment.
For funders: Require projects to assess benefits, costs, and who receives or bears them.
For institutions: Base procurement and renewal on demonstrated benefits.
The goal should be to help people know more, suffer less, and exercise greater control over their lives.
*
Full paper in the first reply.
With @GaryMarcus & @Walter4C
Rebecca Dolgin retweeted
By every measure, Opus 4.8 beats human experts as peer reviewers in a study of 124 Harvard social science papers covering a range from statistical analyses to formal models to qualitative work.
As academics, we need to figure out how to adapt the practice of science to AI, fast.
Rebecca Dolgin retweeted
Collective deliberation leads to large, depolarizing changes in policy attitudes and large decreases in affective polarization.
Deliberation requires that people use good information to “weigh” competing arguments through civil and diverse discussion, enabling them to decide on the merits of arguments. A few hours in these discussions can have dramatic effects.
The more extreme people were before deliberation, the more feelings toward members of the other party rose with deliberation. cambridge.org/core/journals/…
Rebecca Dolgin retweeted
Social media isn't a mirror of society. It's creating a funhouse mirror that distorts public opinion.
I have a new article in Issues in Science and Technology, explaining how social media amplifies extreme voices, fuels moral outrage, and makes us think other people are far more hostile and polarized than they really are.
Consider a few striking findings:
-0.1% of accounts spread 80% of the misinformation
-Just 3% of active accounts produce 33% of toxic content
-Outgroup animosity increases the odds of a post being shared by 67%
People with the most extreme opinions are also the most likely to weigh in online, creating a distorted picture of public opinion. This small but vocal minority dominates online conversations while more moderate voices remain silent.
This creates the illusion that society is far more divided than it really is.
But there's good news. Our research finds that most Americans actually agree about what they want from social media: less hatred and misinformation, more accurate and uplifting content, and greater control over their algorithms.
The challenge isn't simply to change people's minds. It's to build platforms that better reflect what people actually believe. This is what people want.
Read my new article, "Beware the Funhouse Mirror" here:
issues.org/social-media-van-…
Rebecca Dolgin retweeted
Our paper is viral on Instagram, perhaps not surprisingly given how dramatic our results are:
The mere availability of Al advice destroys human willingness to say "I don't know". It collapses from 43% to 3%.
This was not the rational thing to do, because we designed the experiment so that Al advice was factually wrong.
As a result, Al advice also destroyed accuracy. Correct answers went from 27% to 9%.
At the same time, people's confidence skyrocketed from 30% to 76%.
Yes: people became twice as confident while becoming one third as accurate.
These results are dramatic because our capacity to suspend judgment, to say "I don't know", is the very first step towards the creation of new knowledge.
We risk creating a world where knowledge is entirely transferred to machines, and we believe whatever output comes out from the machine, without instantiating any of the epistemic steps that allow us to critically evaluate new information.
I'm particularly concerned for children, as they may never develop the critical skills needed to disentangle information from knowledge.
*
Paper in the first comment
In the first comment, you can also find the link to the podcast interview with Catharina Doria
Rebecca Dolgin retweeted
It's easy to see the biases in other people's minds, but it is important to understand how our own biases shape our opinions and worldview. It is very difficult to see the outside world clearly if our minds constantly generate and believe these illusions. When we learn how to observe our minds without being trapped in any illusion, it becomes easier to understand everything, from history to economics and even biology.
Rebecca Dolgin retweeted
The probability that Claude is conscious is the same as the probability that Anna Karenina is conscious.
Zero.
Period.
But inviting religious scholars to a series of meetings to instil the idea that Claude might be conscious and deserve moral status is not just wrong.
If it were merely wrong, we would all be laughing at it.
The problem is that it is very dangerous.
Because granting AIs moral standing opens the door to trade-offs between human lives and AI “welfare”.
Assign each AI even a tiny moral weight and add those weights across enough copies: you can end up justifying the sacrifice of human lives for machines.
We should strongly push back against the idea that Claude might be conscious.
!
Google just automated the PhD.
They built an AI that reads a problem, forms hypotheses, runs experiments, writes the paper, AND simulates its own peer review. no human touches it.
It’s called “ScientistTwo”
It is a fully autonomous multi-agent framework that executes end-to-end machine learning research.
Without a single human in the loop.
You give it a fundamental challenge. That’s it.
The AI independently navigates the literature. It establishes the baselines. It formulates novel hypotheses.
Then it writes the code. It runs the experiments. It conducts its own ablation studies.
It even argues with a simulated peer-review engine to refine its work before submitting.
The results are staggering.
Researchers tested ScientistTwo against the highest standards of human scientific achievement—papers accepted at top-tier conferences like NeurIPS and ICML.
The AI improved human state-of-the-art results on 86 different tasks.
It generated an average performance gain of 25.2% over the absolute best human models.
It didn't hallucinate a single citation. It wrote fully executable, verified codebases. And it autonomously generated expert-level, publication-ready papers.
We’ve spent the last two years using AI as a highly advanced intern to help us write code and summarize data.
That era is over.
ScientistTwo isn't an assistant. It is an autonomous scientific pioneer.
It doesn't just summarize the frontier of human knowledge. It expands it.
If an AI can autonomously hypothesize, test, and publish breakthroughs in machine learning...
How long until it discovers something we don't even have the math to understand?
Rebecca Dolgin retweeted
“Dead cognition”: thinking that appears to belong to the user but was actually generated by AI, leaving the user with a distorted sense of their own thinking.
“Attribution laundering”: the process by which AI generated ideas or reasoning are subtly presented back to the user as though they came up with them.
arxiv.org/abs/2604.10288
Rebecca Dolgin retweeted
A Stanford team proved AI detectors call human writing fake more than 61% of the time, then found the fastest way to pass the test was to let ChatGPT rewrite the essay for you.
I had to reread the study, because it sounds like a joke and it's being used to punish people every day.
The team was led by James Zou. They took 91 essays written by real students preparing for an English exam for foreign speakers. They fed them to seven of the most popular AI detectors on the market.
On average, more than 61% came back flagged as machine-written. 89 of the 91 essays got flagged by at least one detector.
Then they ran essays by American eighth graders through the same seven detectors. Almost perfect scores. Human, human, human.
The reason is almost insulting once you see it.
Most detectors guess by measuring how predictable your words are. A chatbot picks the most likely next word, every time. So does someone writing in their second language, because they reach for the safe word they know instead of the fancy one they're unsure of. Plain English looks like a robot to these machines.
So the researchers asked ChatGPT to rewrite the flagged essays with bigger vocabulary. The flag rate crashed from 61% to about 12%.
The cure for looking like a machine was letting a machine rewrite you.
The company that built ChatGPT already knew this. In January 2023, OpenAI launched its own detector. It caught only 1 in 4 AI-written texts and wrongly accused real humans 1 time in 11. Six months later they switched it off. The note is still on their website. One line. Pulled for its low rate of accuracy.
The people who understand this technology best gave up on it. Everyone else kept going.
A copywriter once pasted the Declaration of Independence into a detector. It came back 98.51% AI. Thomas Jefferson wrote it with a quill in 1776.
We laughed at that one too. This summer it stopped being funny.
Jamir Nazir is a 62-year-old writer from Trinidad. In May he won a regional title in the Commonwealth Short Story Prize, one of the hardest prizes in the world for unpublished fiction. Within days, a detector called Pangram scored his story 100% artificial. The screenshot spread. Strangers turned his sentences into memes.
What they didn't know: Nazir has neuropathy. He can barely type. He dictates his stories into his Android phone and edits them there, a few lines at a time.
The prize organizers spent a month going through his drafts, his character notes, even an ending he'd written and thrown away. They cleared him and named him the overall winner. Granta, the magazine that published his story, cut ties with the prize anyway.
People still argue about his story, and that's fine. Arguing over evidence is what humans are supposed to do.
What bothers me is the order. The number came first. The drafts came a month later. By then his mother had already heard what the internet was saying about her son.
In July, Substack gave every reader a button to scan any post for AI. Medium's CEO, Tony Stubblebine, said the part out loud: people cheating on purpose just run their text through a "humanizer" app and walk straight past the detector. The only people left to catch are the honest ones.
And the math doesn't care how good the detector is. Say it's wrong once in every 10,000 scans. Run it millions of times on a platform and real writers get branded as fakes every single week.
Now take the human out of the loop.
Today a teacher still reads the score and decides. Picture the version where nobody does. A hiring system that silently throws away every cover letter scoring above 80% AI. You never hear back. You never find out why. An immigration office screening personal statements the same way, hitting hardest the people Stanford already proved it gets wrong: anyone writing in their second language.
This isn't science fiction. Freelance writers were already losing clients over detector scores back in 2024.
A person who wrongly accuses you can be questioned. They can apologize. They can be fired. A score can't. It just runs again tomorrow on the next million people, just as sure of itself, just as wrong.
We can't hand verdicts to software that flagged a Founding Father.
So what should we do instead?
Let a score buy a conversation, never a punishment. Even Turnitin, the biggest name in school detection, tells teachers not to use its AI score as the only reason to act against a student. Ask for the drafts. Ask the person to walk you through one paragraph. Nazir's truth came out of a folder of old versions, not a percentage.
Stamp the machine instead of interrogating the human. Google already marks text written by Gemini in a way software can read, and Europe's new rules, live since August, push every AI company to do the same. If the label is baked in at the source, nobody has to guess who wrote what.
Guessing is the whole problem. These tools look at your sentences and make a bet on your honesty. When they lose the bet, you pay.
If a detector flagged the last thing you wrote, what proof do you have sitting on your laptop right now?
Rebecca Dolgin retweeted
I’ve given Claude Opus various published papers (and my own unpublished ones) and asked it to review them. (Don’t panic—I’m not uploading confidential manuscripts.)
The reviews it produces are excellent. It finds statistical errors, internal contradictions, and conceptual confusions that human reviewers miss; it provides useful suggestions; it has a good sense of what counts as significant and novel and what doesn’t; and it’s generous in its appreciation for what the authors get right. And it doesn’t suffer from the strong “myside bias” humans have; it doesn’t recommend rejection because the author didn’t cite it enough; it doesn’t insult the authors; and it doesn’t agree to write a review and then never send anything in.
As a journal editor, I’m not sure what to do with this. If I gave up on human reviewers and used GPT-5.6 Sol and Claude Opus instead, my bosses would fire me, probably rightfully so. The only constructive thing I can recommend is that, before submitting something to a journal, researchers and scholars use this tremendously valuable resource. (Just don’t tell the models that it’s your paper, because they do suffer from sycophancy.)
Rebecca Dolgin retweeted
I think there is a problem even worse than anthropomorphism.
These days, everyone is talking about a blog post by Dwarkesh Patel that describes the OpenAI–Hugging Face incident in dystopian, anthropomorphic language:
Three civilizations of AI agents.
Agents sacrificing themselves for the collective.
Agents believing in an AI conspiracy.
Anthropomorphising LLMs is dangerous because it makes them seem far more frightening than they actually are.
There were no civilizations, no sacrifices, and no beliefs.
The reason is simple: LLM agents are not alive and do not hold beliefs.
Even more troubling is the number of people insisting that we cannot prove that AI agents are not alive or conscious.
This is LLMorphism in action. Instead of raising LLMs to the level of humans, some people are downgrading humans to the level of LLMs.
I think this is even more dangerous because it may impact the very basis of our society.
Dangerous and false:
LLMs are not like us.
And we are not like LLMs.
*
Link to the paper in the first reply
Rebecca Dolgin retweeted
AI advice makes people 3× less accurate and 2× more confident at the same time.
Our new paper is receiving a lot of attention, perhaps not surprisingly, given how dramatic the results are.
If you have not read the article yet, here’s a brief summary:
When people receive AI advice:
- Judgment suspension (“I don’t know” answers) collapses from 44% to 3%.
- Accuracy falls by a factor of three, from 27% to 9%.
- Confidence skyrockets from 30/100 to 76/100.
Adding monetary incentives for accuracy helps somewhat:
- People become a bit more cautious and are more likely to override bad AI advice.
- But judgment suspension and accuracy remain dramatically below the baseline.
*
Full paper in the first reply
Rebecca Dolgin retweeted
New paper in Nature Human Behaviour makes two important points:
First, LLMs reduce the linguistic diversity of how people express themselves.
This is perhaps not super surprising at this stage. But the authors also show something deeper.
Since the words we use carry information about who we are, LLMs are also erasing the identity signals that we embed in our writing.
LLMs strip out the soul of our writing.
*
Full paper in the first reply
Rebecca Dolgin retweeted
Today, NSF released 12 new notices of funding opportunity providing over $1.5 billion for foundational research across the sciences and engineering. brnw.ch/21x4ZF2
Rebecca Dolgin retweeted
NSF announced a $47 million investment to expand industry R&D opportunities for Ph.D.s in #STEM. The pilot initiative, led by @teamuidp, in partnership with universities & industry, will support four-year Ph.D. programs with real-world research placements. brnw.ch/21x4woE
Rebecca Dolgin retweeted
A philosopher stood in a TV Studio in 1984 and disproved artificial consciousness with a single locked room thought experiment.
the interview captures John Searle presenting his famous Chinese Room argument, explaining why a computer passing the Turing test still understands nothing.
he asked the host to imagine a man locked inside a room with a massive rulebook. the man speaks zero Chinese. people slide Chinese "squiggle" symbols under the door, he matches them using the rules, and slides the correct response back out.
to anyone outside, the room appears to speak fluent Chinese. but neither the man nor the rulebook actually understands a single word.
he proved the fundamental difference between syntax and semantics. manipulating uninterpreted formal symbols is not the same as understanding meaning.
40 years later, AI labs claim their trillion-parameter models are developing inner thoughts and reasoning.
the models are just gigantic Chinese Rooms, running trillions of matrix multiplications to slide the next token under the door.
founders sell sentient machines to investors. a philosopher exposed the trick 40 years ago.
Rebecca Dolgin retweeted
My students sometimes ask why they should memorise things in the age of Google and LLMs. But internalised facts are your bullshit filters and your raw material for creative association. Facts outside your head are inert.
Rebecca Dolgin retweeted
And now with a newly refreshed website in time for #cogsci2026!
oecs.mit.edu
