@nandoodles
@sreejan_kumar
@princeton_nlp
feel free to try the beta yourself at or it's still barebones right now, but will be spending the next few weeks improving the model and analysis.
@nandoodles
the motivation here is increasing AI plagiarism. think are high school teachers going to want students using ChatGPT to write their history essays? likely not.
@SydneyFreedberg
thanks Sydney! currently the app uses a few properties, perplexity (randomness of a text to a model, or how well a language model likes a text); and burstiness (machine written text exhibits more uniform and constant perplexity over time, while human written text varies
@nandoodles
the analysis is based on some ongoing research with and
@sreejan_kumar
and
@princeton_nlp
. hopefully we'll publish something empirical soon. but in the mean time this was a fun app to make :)
@nandoodles
@sreejan_kumar
@princeton_nlp
wow! woke up this morning, and did not expect so many ppl here! if the apps' not working rn it's likely because my free
@streamlit
hosting was not meant to handle so many ppl. going to reboot and try again :)
.
@bellingcat
I had the great fortune to work with
@IwriteOK
who's reporting and coverage of alt-right groups has been a real inspiration for many journalists this week. Here's a shoutout to his amazing work:
@SydneyFreedberg
we're still studying implicit bias in LM generated text right now, so hopefully will be adding a few more tests and factors to improve the model
Things you didn’t know about the QAnon conspiracy theories (that are relevant for 2020) based on my data investigations with
@bellingcat
A thread ...👇🏼
While you’re here, please check out Origin, our new chrome extension. Goal is not just AI detection, but to filter for toxicity, misinfo, fake news, verify source of info, and revolutionize media literacy in this AI world. We'd love feedback 🙏
For my New Years resolution, I'm starting a weekly newsletter on tech and journalism. The first issue features investigation tips and tools from an award winning
@bellingcat
and NYT investigator, and personal role model
@trbrtc
!
The new GPTZero model now handles mix AI + human text, and highlights portions of text that are most likely to be AI generated, a key feature that educators in our community have been requesting.
Pinch me.
I'm really honored to be on the
@BusinessInsider
AI 100 list this year.
Being the youngest and among the least experienced on the list is really humbling — especially alongside generational AI greats like Hinton and Urtasun.
also a reminder of the speed of
Said first demo was an invitation to present Origin to
@tglocer
who ran Reuters, and Mark Thompson, who led the NYT through its digital transition AND ran the BBC. Think it went well, here's our before and after pics with Mark!
Overall, we're really excited to continue improving and building GPTZero, and I'm committed to keeping the website interface free for individual teachers and educators. Try it yourself at
@SydneyFreedberg
hey Sydney, totally! perplexity is a metric of randomness, or the degree of uncertainty a model has in predicting (assigning probabilities) to a text. lower perplexity = more AI preferred. but just by itself, perplexity is a really imperfect indicator of ai generation
@VictorButoi
Hey Victor. just saw this. not marketing it at all. I agree with what you said, the current beta is exactly what I built casually over school break, shouldn't be used to make academic decisions. all I can say, is we're building a better version
Stoked to announce our new 10M Series A on TC, and to be working together with the amazing
@nbt
@jameskimEDU
and
@reachfund
in defining the future of AI adoption.
Same day, our team is releasing a new product — GPTZero docs, the future of transparent writing with built in AI
@_RobertSchultz
My roommates were two white ROTC dudes from New Jersey and Alaska, who never watched anime before in their lives. One of them saw me watching Dragon Ball Z, and joined, but the other one started making fun of him. Then we forced him to watch too. Now both are happy weebs.
What's 2020 without a plague of locusts?
We know what you're thinking, "It gets worse?"
Yep, it sure does!
But hey, with the right tools, you can track the locusts swarms!
Last year, we launched GPTZero to disrupt 'academic integrity' a space with a billion-dollar incumbent, Turn-it-in
(some differences, we're a team of AI-native researchers that philosophically believe the 'plagiarism model', and teachers punishing students, should go obsolete;
.
@Yaolri
is the best geolocation teacher! really fortunate to work and learn from such an incredible OSINT investigator; wrote out some of his advice in a post
We’ve been busy building the past weeks. Today, we’re finally sharing what we’ve created with the world: Origin by GPTZero - A chrome extension that enables you to check for AI-generated content as you browse the web!
#originbygptzero
On terminology, I’m working with
@billdthompson
to research the semantics of over four years of Qdrops ... incredibly interesting because QAnon followers has preferred to use hidden and code terms in their discussions
2/ The reason? Trump had misspelled a tweet, correcting consensual to consequential minutes after. 8chan (the platform at the time for QAnon posts) was bustling w discussions of the added letter “q”
@ChFrancavilla
@BardeyGoddie
@Ecostard
As a final note, I want to emphasize that these videos are not a reflection of Chinese people. There was plenty of push back when they first surfaced in China. I'm also ethnically Chinese, find these videos abhorrent, and have made lifelong friends in Africa to this day.
When I say 'we', I mean an incredible team of friends who’ve joined me to work on
#GPTZero
, not least
@Alexcdot
who dropped out of PHD to lead R&D,
@KusioOlivia
NY's most talented designer, and
@EdmondAtto
Uganda's first Obama scholar who flew 24h to NY for a demo this week..
3/ Later that day there was a “Qdrop,” the official posts from Q that make the bedrock of the theories, that solidified the belief that Trump was speaking directly to the community via Twitter
This was two days before the midterm elections. We really need to take a closer look at QAnon discussions in 2018, literally the same voter fraud playbook being reused by Trump today
Today our
#BBCAfricaEye
investigation
#ThreeKillingsinKampala
came out. During the months on this, I've been constantly thinking of my friends and brothers in Uganda, their health, safety and loved ones. Had a look back today in my diary on some entries in Uganda (a thread)
These victims were mothers and fathers, locals and passers-by, children and unborn babies...
...shot and killed by the same security forces whose job it is to protect the citizens of Uganda.
@billdthompson
and I are also currently training a machine learning model using
@TensorFlow
to analyze and visualize connections between Q posts and their meanings! like this
Yesterday night after we migrated our site to soft launch GPTZeroX (prior to sharing / fully launching today) we received 39.7k visitors and 2.2 million requests to GPTZero in one day ...
launched a new website today! last minute easter egg, added in things for people to read when our AI scan is running
so far its all my favorite journalists ... feel free to send quotes to add!
something else you might notice are the lightning fast processing speeds (esp. compared to our beta) thanks to our GPUs on AWS Lambda servers, which are pretty much just charging my credit card with each call!
For example, from early Q posts, “Alice and Wonderland” is code for Hillary Clinton and Saudi Arabia; and “Operation Mockingbird” refers to the belief of a program where the people in power who are child traffickers have controlled the CIA and fake news media to hide their crimes
13/ her passing removed this branch from popular Qanon discussion, but it illustrates how the conspiracy in its 48 month lifespan has taken ridiculous leaps to grow,
12/ QAnon groups believed that RBG was part of the occult and drinking the blood of children to stay alive. Why? because she was nominated by Bill Clinton, and thus connected to the Clintons.
when 30,000 people signed up for GPTZero as a school project, I already thought it was unbelievable
in the past year, we’ve continued to ship new products, delight new users, and grow to over 3M users
fairly proud we’re now up there on the list of fastest growing companies on
Oh i forgot to mention, thanks *a lot* to
@Pinperepette
,
@edward_the6
and
@chronic0ps
for the data that helped me baselining the model without introducing personal bias.
Robert's reporting — especially this piece the day before the breaching of the capital — are must reads for anyone interested in understanding the actors behind the events at the capital.
To ensure fair voting districts, five CS undergrads started , a free open-source online platform. It helps communities fill out questionnaires and draw themselves on a map, creating a database for the redistricting process.
@JaneFerguson5
and Yemense producer
@ahmedbaider1
join our Princeton journalism class, 1 day after winning the Emmys, to share how they smuggled Jane into Yemen and the inspiration behind their reporting ...
If this isn’t the coolest class in college I don’t know what is.
11/ imo the Qanon theories have always been molded by elections. The core focus of the theories in 2016 was Hillary Clinton (running a child trafficking ring) which has evolved to connect the Clinton family with numerous others characters
My profile on Dr. William Massey — trailblazing Black mathematician and mentor — and the many students he's inspired like
@NickAGJohnson
just won first place in the New Jersey Press Foundation's college journalism contest!
7/ Donald Trump’s statement on the “calm before the storm” in 2017 was interpret as supporting and enabling QAnon; “the storm” in the theory refers to a grand awakening demonstrating everything in QAnon is true
A new investigation from
@edward_the6
focuses on inauthentic social media campaigns targeting
@DrTedros
, the director of the World Health Organization. Edward also included his data sets for you to take a look at as well!
I’ve been using Perplexity for the last few months, but have noticed increasing numbers of AI-generated sources
This is important because even if Perplexity claims its AI is better than ChatGPT if the sources are AI hallucinations, the output is AI hallucinated 🧵1/n
for the longest time tech startups we're a monolith to me,
but as we were🎉🙌one of our Ugandan engineers becoming a father on father's day, struck me how awesome/diverse our team really is,
75% of our eng team is Black
50% of our ML team are women
100% international/non-US 😎
The end product, made and filmed by
@ChFrancavilla
@BardeyGoddie
@ecostard
Matthew Bradley, and rest of the phenomenal production team is nothing short of breathtaking to watch.
Feb 22. I'll be discussing the future of education with
@UNESCO
and our roadmap for building
#gptzero
from a viral app into an AI solution for educators everywhere. Thanks
@unesco_iite
institute for organizing.
Just published on
@bellingcat
! four months after we first dived into a conspiracy dataset, and my first encounter with sentiment analysis, I'm especially grateful to my advisor
@billdthompson
Our month-by-month analysis reveals the story of QAnon’s growth, showing how this notorious conspiracy theory has been able to evolve its narratives and adapt to contradictions — potentially allowing it to live on.
GPTZero is excited to partner with
@knightfdn
in sponsoring the Media Party’s journalism-tech Hackathon! as
@jeremygilbert
says instead of thinking about what gen ai can do, let’s focus on building what readers of news want.
1/ We’re thrilled to announce the launch of Vivid (), an AI wingman to build your web app. Just click a component on your screen, let AI generate all your boilerplate, and make in-browser edits that automatically sync with your source code.
back to the thread, 6/ Facebook (despite growing to be the host of many QAnon groups) is believed by the theory to be part of “operation mockingbird” and a tool used by the CIA for propaganda purposes
Recently there was a Washington Post piece on GPTZero. Frustrating journalism that we didn't receive any chance to comment before publication. Here's what I shared to the author after.
- emailed factual updates for
@washingtonpost
- wholeheartedly agree with the advice for
Analysis: AI detectors like Turnitin and GPTZero suffer from false positives that can accuse innocent students of cheating. Here’s the advice of academics, AI scientists and students on how to deal with it.
especially proud and so grateful to work with our team of four engineers now, who made sure this happened with zero crashes / network delays (as opposed to when GPTZero immediately crashed after going viral, when I was working on it myself)
"News is not 'content', readers are not 'users', stories are not 'syntheses' A truly free press controls its language from start to finish."
Powerful read, maybe a good one for tech people.