Pangram Labs

Machine Learning conferences have an AI problem. Since the release of ChatGPT, an increasing number of submissions are AI-generated. We set out to answer the question: are AI researchers using ChatGPT to write their own research? We examined conference submissions from 2018 to 2024 at two of the largest AI conferences: ICLR and NeurIPS, and found that AI-written papers skyrocketed. From 2022, where AI was not used at all in writing papers, to 2024 where nearly 5% of submissions were AI-written to some degree. Read our blog post for a full analysis on how AI use has grown in the research community and how it correlates with paper rejection: https://lnkd.in/epV2eePf

I’m happy to share that I’m starting a new position as a founding Full Stack Engineer in Pangram Labs in new york.

Thrilled to announce my company's name change from Checkfor.ai to Pangram Labs! Six months ago when we started out we quickly chose the name Checkfor.ai to reflect our value proposition as AI detection. It worked for a while, but as we continued our journey we wanted a name that would better convey what we aspire to become. A pangram is a sentence or piece of text that uses all letters of the alphabet. For us, this is an apt metaphor for how we are developing our product. We're building machine learning models that analyze all facets of human language to determine what is human and what is AI. Most importantly, we're doubling down on our mission to keep the internet free of AI-generated spam. Nothing under the hood is changing, except maybe our emails. We've also got a new website - Check it out at pangramlabs.com!

Today we're releasing a 14 page technical report on the Checkfor.ai AI text classifier! We believe open research is key to our goal of seriously solving LLM generated text detection. https://lnkd.in/e3ZW4mHW Some of the hits: - We outperform two leading commercial competitors and DetectGPT by nearly an order of magnitude on a benchmark of 8 SOTA LLMs and 10 domains - The performance gap widens substantially when looking at only GPT-4 - We show our model has learned to generalize well enough to achieve 99.5%+ accuracy on several open source models and the latest OpenAI updated models that do not even have a single example in the training set!

Excited to be joining two rockstars, Bradley Emi and Max Spero, in building Checkfor.ai - the leading technology in identifying AI-generated content found online. We're on a mission to develop the best-in-class AI detection technology and protect the internet from an influx of inauthentic, fraudulent, and deceptive information that will result from the abuse of Gen AI tools. Check out our analysis of fake, AI-generated reviews left on Reddit: