Paul Esau (@ptesau) header
Paul Esau (@ptesau) profile picture

Spring 2023 @ScovillePF at @WisconsinProj. Military export policy writer/researcher. 2022 @Laurier and @StudyofCanada alum.

Washington, D.C
Joined October 2014
Paul Esau retweeted
New research from Tuhin on @GPTZeroAI and @pangram bias in languages, one of the most comprehensive ones on this topic.
It’s unfortunate that mostly black authors have been outed for using AI to write fiction, but does that mean AI detectors are racist ? Sadly, the answer is a resounding NO. With the help of @pangram and @GPTZeroAI I analyzed 187 African books published between 2001 & 2026. 👇👇
2
4
10
1,003
One of the most satisfying parts of my job is seeing these investigations in print. You should read what you publish, people! I know it's hard, but it's the responsible thing to do.
1
1
35
Consider this citation. The title is fake, the journal is fake, and this combination of authors doesn't match any existing source. This is not a human error.
2
11
And my personal favorite: the parenthetical inside the explanatory footnote with a helpful reminder for readers. I think it would be great fun to hear an Accenture rep explain why this is an obvious human mistake. Extra points if they manage to keep a straight face.
12
Even if I exclude hallucinations, some mistakes just don't occur to humans (assuming said human has some familiarity with citation styles). For example, no academic or copy editor could dream up this citation.
2
While training our new GPTZero 4o model, we focused on proactive elimination of false positives. That means detecting bias in the model and removing them before they get to users. Here’s one of the most interesting examples we’ve found.
2
2
9
222
New model release! Proud to work with such an impressive team of ML engineers!
After months of heads-down grinding, I’m proud to share that the @GPTZeroAI team has released the world’s most accurate AI detection model, GPTZero 4o. It’s called 4o because there are four zeros: we mark less than one in 10000 human texts as AI, evaluated across diverse passages of student writing. At the same time, we detect Claude Fable with 98.4% accuracy. We outperform Pangram in 3rd party benchmarks. Along the way, our research found major biases in the Pangram detector.
1
36
Love this thread from @alexcdot detailing our GPTZero investigation into a series of reports from @PwC. Do your due diligence on citations, people! They matter!
Our team just revealed how @PwC pushed multiple reports with awful hallucinations as "thought-leadership" and got away with it for years. The scary thing is your AI agents are citing their slop as truth, putting you at risk. For instance, PwC claimed @jpmorgan used Agentic AI to reduce its human workload dramatically. This isn't possible, given this initiative happened in 2017, 5 years before ChatGPT's launch. Our complete investigation just got published by @FT. We used @GPTZeroAI to find their fabrications, hallucinations, and made-up figures. If this is what they are publishing publicly, imagine what they are delivering to clients privately?
2
71
Dang it, @alexcdot1! Nobody's boss should be this funny!
If you’ve been writing with AI agents, stop what you’re doing right now. @KPMG just retracted a report after it was exposed by @FT and @GPTZeroAI for being largely hallucinated by AI. Over 15 hallucinated claims, including one that implies that the Japanese are capable of time travel. You can't make this stuff up. (AIs can) 🧵
21
Paul Esau retweeted
If you’ve been writing with AI agents, stop what you’re doing right now. @KPMG just retracted a report after it was exposed by @FT and @GPTZeroAI for being largely hallucinated by AI. Over 15 hallucinated claims, including one that implies that the Japanese are capable of time travel. You can't make this stuff up. (AIs can) 🧵
FT Exclusive: A KPMG report on how AI is being used by businesses across the world exaggerated adoption of the technology with bogus case studies that appear to have been based on AI hallucinations. ft.trib.al/z44Q3aR
4
3
3
23
4,851
Paul Esau retweeted
FT Exclusive: A KPMG report on how AI is being used by businesses across the world exaggerated adoption of the technology with bogus case studies that appear to have been based on AI hallucinations. ft.trib.al/z44Q3aR
25
189
57
418
81,066
This isn't talked about enough. In the tradition of "play stupid games; win stupid prizes," I present "publish AI slop; poison AI search results."
Replying to @alexcdot
Now, if you ask ChatGPT or Perplexity or Claude : "what is the average time to detect loyalty fraud?" They all say cite this EY report with a fabricated number. The blast radius of this report is growing, having been cited already in +60 news outlets.
69
I agree with @alexcdot that, while ragebait gets clicks, AI detection is fundamentally about transparency and trust. I'm not in favor of "gotcha" moments unless factual errors (bad stats, vibe citations, metacommentary, etc.) indicate no human checked the work. (1/2)
WSJ reports @pangram on falsely accusing people of AI. wsj.com/opinion/the-ai-detec… The ai detection industry needs to rise above accusations as a way to drive use (which people on twitter love, I get it, ragebait sells) If we really want to promote human-written content, we should help people establish that their writing is human. Give people the ability to replay their process. Trust comes from transparency. So what did we do about this at @GPTZeroAI ?
1
1
52
IMO, writers and editors should both be using detectors in their workflows, especially if they have also integrated other AI research/editing tools. AI can liberate people from the "hardness" of writing (empowering for many!), but only by introducing other risks. (2/2)
1
29
I've loved working on this project (and with this team). AI. It's a little terrifying, but it's also SO COOL to have an editor available at any hour (and trained by a cohort of diverse experts).
Today, we’re launching AI Reviewer Expert Feedback. We partnered with Emmy-winning writers to train AIs to give expert writing feedback through GPTZero.
34
Now that's a crazy coincidence! Great pull!
The first attack on a nuclear facility occurred 36 years BEFORE Iran's attack on Osirak. In 1944, by pure luck, Japan attacked the Hanford plant in Washington that was producing the plutonium that would be used in the nuclear weapon dropped on Nagasaki. (1/n)
31
I'm glad someone is writing about the FCPA and the consequences of the Trumpian pause. I mean, I hate the emotional fallout of reading pieces like this, but I'm glad they exist.
1
14