Is your online crush an AI?

Hi friends 👋🏻  - Mariena here. I appreciate all of you supporting Mellonhead.

I recently studied AI content detection tools and research and wanted to share what I learned. Read on below.

But first – an update on my workshops.

AI Workshops

I recently finished teaching the second and third cohorts of my Basics of Generative AI workshop series. The feedback has been positive—the workshops have helped both those who have never written a prompt before and those who were using tools like ChatGPT occasionally. After six weeks, everyone uses Generative AI multiple times a day or week and has gained experience using multiple frameworks and best practices.

Some of the feedback I’m hearing is that people have a hard time developing the muscle memory to use AI and that 6-weeks is hard to commit to.

Based on that feedback, I’m currently working through curriculum changes to meet more frequently over a shorter period (e.g., twice a week over a two or three-week period).

I’ve also successfully broken out the curriculum into two small sessions that I am considering offering as one-off sessions:

1. Basic AI for those that have little to no experience with Generative AI
2. Better AI is for those who know some best practices and use AI regularly but have that nagging sense that they’re not using it to its full potential.

I’m working on setting this up in Maven, a teaching platform that will make it easier for me to gauge interest. Stay tuned as I share signups for each of these next week.

Is Your Online Crush an AI?

This was a real conversation I had recently. Well, almost. More like, “I think this scammer is a bot.” I asked my friend to share the messages she was sent on a Dating App so I could try out various AI detection tools.

The FTC has reported that romance scammers swindled $1.1 billion from daters in 2023 - more than double the $547 million reported in 2021.[1][2]. To no one’s surprise, Bloomberg confirmed that dating sites are seeing AI-generated fake profiles.[3]. With that in mind - her concerns seem valid.

So, how can we tell if our new online romantic interest is a charlatan hidden behind proficient prompts or if a profile is fake?

The market is full of tools for detecting AI-generated text and images. This is just a subset.

AI Text Detection

• Originality.ai ($14.95/mo OR $30 flat to scan 300k)
• Turnitin.com (EdTech software, no self-serve signup)
• Winston AI ($12/mo, first 2k words free)
• Copyleaks - ($10/mo, limited free version)

AI Image Detection

• Is It AI? (free)
• AI or not (free)
• Content at Scale (part of broader package)
• Illuminarty (free)

What Research Says

To understand the state of AI content detection, I found multiple relevant research papers

• First, published in January 2024, the first survey to comprehensively cover existing research on detecting multimedia created by LAIMs, and introduces a novel taxonomy for detection methods, categorized by media modality, and aligned with two perspectives: pure detection and beyond detection. See Lin, Li et al. “Detecting Multimedia Generated by Large AI Models: A Survey.” ArXiv abs/2402.00045 (2024): n. pag.
• Second, published in July 2023, is research on AI-generated text detection tools showing they struggle to accurately distinguish between human and AI-generated content, with a bias towards classifying AI-generated text as human. The study evaluates 12 public tools and two commercial systems used in academia, finding they are neither accurate nor reliable.
• A third paper published in January 2024 describes a methodology to detect AI-generated content by prompting LLMs to rewrite text and calculating the editing distance of the output, which significantly improves the F1 detection scores of existing AI content detection models across various domains. See Mao, Chengzhi et al. “Raidar: geneRative AI Detection viA Rewriting.” ArXiv abs/2401.12970 (2024): n. pag.
• Fourth, a paper published in April 2024 explores the trends and patterns observed in real, deepfake and synthetic facial images shows that synthetic, deepfake and real face images are indeed three different classes. Naeem, Shahzeb et al. “Real, fake and synthetic faces - does the coin have three sides?” (2024).
• This NYTimes author experimented with several AI detection tools for images and has a thorough writeup on detecting AI-generated images.

The takeaway is that there’s a lot of research happening in this area, but current AI detection methodologies have limited reliability and are unfair. from Lin et al, “It has been revealed that current detectors, though with high detection accuracy, result in unfair performance disparities among demographic groups, such as race and gender. This can lead to particular groups facing unfair targeting or exclusion from detection, potentially allowing misclassified LAIM-generated multimedia to manipulate public opinion and undermine trust in the model.”

My Experience

I tested Originality.ai, Copy Leaks, and Winston using a variety of my thought leadership pieces: (1) fully written by me using Grammarly to polish, (2) drafted by AI but revised by me, and (3) unpublished drafts fully written by AI. I also provided a few dating messages, which were of unknown authorship.

The results:

• Originality seemed to do the best, in part because it gave a score and was able to show that “some parts may be generated by AI” rather than depicting the text as all or nothing.
• Copy Leaks wasn’t very reliable - it flagged my rewritten drafts as AI, which surprised me, given I’m pretty confident most of the final product was my own. You get what you pay for with a free tool.

And drumroll, please… although the dating app messages definitely give off scammer vibes, they do not seem to be AI-generated.

Thanks for reading! If you know someone interested in leveraging AI in their organization or training their teams, share my info or make an introduction. www.mellonhead.co