AI detection is a fool’s errand.
The basic math of generative AI simply doesn’t favor the detective. For every update to a classifier, there is a thousand-fold increase in the number of ways to bypass it. You can prompt a model to “write like a tired journalist from the 90s” or run the output through a simple paraphraser, and suddenly the statistical markers that detectors rely on vanish. Yet, the venture capital continues to flow into this void.
The latest example is Pangram, which recently raised $9 million (which is a staggering amount of capital for a tool that is essentially a statistical guess-engine) to scale its detection software. Along with the funding, they have rolled out Pangram 4 and a research preview for image detection. According to TechCrunch, the goal is to help publishers and educators keep the “AI flood” at bay.
The problem is that we are treating a linguistic shift as a security breach. We are trying to identify “bot-ness” by looking for patterns—perplexity and burstiness—that are already being smoothed out by better sampling methods and RLHF. It is like trying to catch a ghost with a butterfly net. You might snag something occasionally, but you aren’t actually controlling the environment.
It is a losing game.
The real danger here isn’t that the detectors fail to catch the AI; it’s that they confidently flag humans. We have already seen the carnage in academia, where students have been accused of cheating because they write with a level of clarity and structure that happens to mimic a well-tuned LLM. When a tool like Pangram 4 is marketed to institutions, the “detection” becomes a proxy for truth.
The friction is palpable. Imagine the cost of a false positive in a professional setting—a journalist losing their job or a developer being fired because a tool decided their commit messages were “too synthetic.” These tools don’t provide a proof; they provide a probability. But in the hands of a middle manager or a dean, a 70% probability is often treated as a conviction.
We are essentially building a massive, expensive infrastructure to police the way people write. But the more we use these tools, the more we force humans to write “weirdly” just to avoid being flagged. We are incentivizing a regression in prose to satisfy a classifier.
The irony is that the only way to truly “detect” AI is to have a cryptographic watermark embedded at the point of generation. But that requires the labs—the ones making the money—to cooperate. Why would they? There is far more value in a seamless, invisible integration into the human workflow than in a “Made by AI” sticker on every paragraph.
The market is currently obsessed with the “detector” side of the equation because it feels like a solution. It feels like we can go back to a world where we know who wrote what. But that world is gone. We are moving toward a reality where the content itself is irrelevant, and only the identity of the sender matters.
By Q1 2027, the market for standalone detectors will collapse in favor of cryptographic provenance and authenticated content streams. We will stop asking “was this written by a bot?” and start asking “is this signed by a verified human?”
Until then, we will keep seeing these funding rounds and these new model versions. We will see a cycle of “Pangram 5” and “Pangram 6,” each claiming to be more accurate, while the LLMs they are chasing simply move the goalposts further down the field. It’s an expensive treadmill, and we are all just watching the numbers climb without actually getting anywhere.
