I research AI authenticity - detection tools, humanizers, deepfakes, content provenance. I test the tools, publish the numbers, and question the claims vendors won't. If you're trying to figure out what's real in AI-generated content, you're in the right place.
I ran 847 samples through every major AI detector on the market. Here's what actually works, what doesn't, and why the accuracy claims on most landing pages are misleading.
AI humanizers rewrite machine-generated text to bypass detection tools. I tested the top 8 against GPTZero, Originality, and Turnitin to see which ones actually deliver.
Deepfake detection is a moving target. I looked at the current state of forensic tools, what the research says about their limits, and where the real vulnerabilities are.
C2PA, invisible watermarks, metadata provenance - the industry is building infrastructure to prove what's real. Here's what's actually shipping and what's still vaporware.
Perplexity scoring, classifier thresholds, token probability distributions - the mechanics behind every AI detector, explained without the marketing spin.
I test AI detection tools, humanizers, deepfake detectors, and content provenance systems. Not as a vendor selling one of them, but as an independent researcher trying to figure out what actually works. The AI authenticity space is full of bold accuracy claims, cherry-picked benchmarks, and marketing that doesn't survive contact with real-world text. My job is to run the tests that vendors won't run on themselves, publish the numbers, and let people make informed decisions. I've put over 4,000 writing samples through detection pipelines in the past 18 months, spanning everything from student essays to professional journalism to machine-generated content across GPT-4o, Claude, Gemini, and Llama outputs. The results don't always match what the product pages promise.
My methodology is straightforward. I build controlled test sets with known provenance, so I know exactly what's human-written and what isn't before any tool touches it. I run every sample through multiple detectors under identical conditions, record the raw scores, and calculate accuracy, false positive rates, and false negative rates myself rather than relying on vendor-reported numbers. When I test humanizers, I measure whether the rewritten text actually passes detection and whether it still reads like something a human would write. I published the "State of AI Detection 2026" benchmark using this approach, covering 12 major detectors across 847 controlled samples. That study found that the gap between claimed and observed accuracy averaged 14 percentage points across the tools tested.
My background is in computational linguistics. I spent years working on NLP problems before the current wave of generative AI made detection a mainstream concern. That training shapes how I approach this work. I think about token distributions, classifier thresholds, and perplexity scoring rather than treating these tools as black boxes. It also means I'm skeptical of simple narratives. Detection isn't a solved problem. Humanization isn't going away. Deepfakes are getting better faster than forensic tools can keep up. The honest answer to most questions in this space is "it's complicated," and I'd rather give you the complicated truth than a clean headline that falls apart under scrutiny.
If you're a journalist, educator, researcher, or policy person trying to navigate AI-generated content, this site is where I publish my findings. Expect tool reviews grounded in data, explainers that don't skip the hard parts, and occasional sharp opinions about an industry that sometimes prioritizes marketing over accuracy. I don't take money from detection companies or humanizer vendors. The work stands on its own.