One AI detector flagged the US Constitution as AI-written. Another insisted The Da Vinci Code was 100% AI-generated.
And this week, a startup building AI detectors just raised $9 million.
Run two detectors, get two different answers
Let's start with how much of the internet is actually AI now. As of May 2025, 35.3% of newly created websites had AI assistance, and 17.6% were almost entirely AI-generated. Researchers from Imperial College London, Stanford, and the Internet Archive pulled these numbers by combing through the Wayback Machine.
So naturally, a wave of tools showed up to answer "did AI write this?" Here's the thing: these tools flatly contradict each other. In a head-to-head benchmark by Fritz AI, GPTZero flagged the US Constitution as AI-written, while Originality.ai rated the actual text of The Da Vinci Code as 100% AI-generated.
The gap gets worse with newer models. On the same benchmark, GPTZero caught 100% of GPT-5 text, while Originality.ai caught just 31.7%. For GPT-5-mini, the split was 94.9% versus 7.3%. Which tool you pick can literally flip the verdict.
| Tool | Strength | Known weakness |
|---|---|---|
| GPTZero | 100% on GPT-5, free tier (10K words/mo) | Weak on paraphrased text, flagged the Constitution |
| Originality.ai | Strong on paraphrase evasion, bundled plagiarism check | Trained mostly on older models — only 31.7% on GPT-5 |
| Pangram | 99%+ accuracy, ~1 in 10,000 error rate, now covers images too | No free tier ($20/mo+), fewer track records outside the US |
This isn't abstract — it's already hurting real people. Students who wrote their own essays get falsely flagged, while AI-written submissions slip through. Even Originality.ai's own meta-analysis of 15 independent studies warns that "AI detection scores should not be used as the sole indicator of academic misconduct".
So Pangram took a different approach
Pangram, the New York startup behind this week's $9M round, is led by Menlo Ventures with Haystack, ScOp, Script Capital, and Cadenza joining in. Including a $2.7M seed in June 2025, the company has now raised about $13M total. Co-founder Max Spero, a Stanford AI grad, says he started the company after watching LLM-powered Russian disinformation campaigns.
The technique itself is clever. Instead of watermarks or metadata, Pangram builds a "synthetic mirror" for every document — a twin written by a top LLM, matched on topic, length, and tone. It then learns the stylistic gap between the two: the consistent word-choice patterns that give AI away.
This round also funds a new product, Pangram Image. It works off pixel-level statistical differences rather than watermarks, so it can catch output from FLUX, Midjourney, and Grok Imagine — none of which embed watermarks. That matters because humans are bad at this: average accuracy at spotting AI images is 63.7%, and it drops to 29% for FLUX specifically — worse than a coin flip. Substack and Quora already run content through Pangram, and Substack started labeling posts over 100 words with a human / AI-assisted / AI-generated split on July 21.
Korea is a different story, though. According to an interview with Muhayu's CEO, running Korean text through foreign detectors was basically a coin toss. So the company built a Korean-specific detector, GPT Killer, and in August 2025 added GPT-5 detection with a claimed 98% accuracy.
Here's how to actually use detection results safely
Taking one detector's number at face value and rescinding a job offer or failing a student is how you end up with a real mess. Use this order instead.
- Never treat a score as final proof
It's circumstantial, not a verdict. Give the person a chance to explain before any decision sticks. - Cross-check with at least two tools
Start with GPTZero's free tier (10K words/mo), then run anything borderline through Pangram or Originality.ai. If the results flatly disagree, that disagreement is itself a signal. - Check paragraph-level scores, not the document average
Detectors compute probabilities per paragraph. Find which paragraph triggered the flag instead of trusting the overall number. - Pair with a language-specific tool for non-English content
Foreign detectors can misfire badly outside English. For Korean, run a local tool like GPT Killer alongside. - Write down your appeals process before you need it
If no one owns what happens after a false positive, wrongly flagged people just get buried.
| Trust one score | Cross-check first | |
|---|---|---|
| False-positive risk | Constitution/Da Vinci Code-level misfires | Two tools rarely fail the same way |
| Basis for a decision | One number, taken at face value | Circumstantial evidence + a chance to explain |
| Cost | One subscription | Doable on free tiers alone |
Want to dig deeper?
Pangram's $9M funding, straight from the source The most detailed writeup on the synthetic mirror technique and the founding story techcrunch.com
Pangram's image detection launch Cumulative funding, image-detection accuracy, and the human-accuracy numbers siliconangle.com
GPTZero vs Originality.ai benchmark The actual test that caught the Constitution and Da Vinci Code false positives fritz.ai
Meta-analysis of 15 AI-detection accuracy studies The source of the "don't use this as your only signal" warning originality.ai
The study behind Dead Internet Theory numbers Imperial College London, Stanford, and the Internet Archive actually measured how much of the internet is AI gizmodo.com
Muhayu's GPT Killer hits 98% claimed accuracy The latest update from Korea's dedicated Korean-language detector hellot.net
