As machine-written text and images pile up across search results, social feeds, and even courtrooms, one detection-focused startup just secured new backing to scale a system designed to tell people what they're really looking at.

A New York-based startup focused on identifying AI-written and AI-generated content has closed a $9 million funding round, led by a prominent venture firm with several smaller funds joining in. The raise lands alongside two product launches: an upgraded text-detection model and a brand-new image-detection tool currently available only through early research access.

The company says its latest text model now clears the 99% accuracy mark when identifying content that was either fully written by AI or produced through a mix of human and machine effort. It also claims stronger performance against so-called "humanizer" tools, apps designed specifically to disguise AI writing so it slips past detectors.

The founders, both machine learning graduates from Stanford, started the company roughly two years ago, shortly after ChatGPT's debut triggered a wave of bot accounts, SEO spam, and coordinated influence campaigns online. One co-founder described watching state-linked disinformation networks exploit generative AI to flood social platforms almost overnight.

Their pitch is straightforward: readers deserve to know whether what's in front of them was crafted by a person or assembled by a model, because that knowledge changes how skeptically someone should approach the material, whether it's a news story, an academic paper, or a legal brief.

Rather than scanning for hidden watermarks or metadata, the system was trained on tens of millions of verified human-written documents. For each one, engineers generated a matching "synthetic twin," a version with the same topic, length, and tone, but produced entirely by a leading language model. Comparing the two teaches the system the subtle stylistic fingerprints AI models tend to leave behind.

The tool doesn't just flag fully AI-written material. It also tries to measure degrees of assistance, distinguishing between text a person wrote from scratch, text that was lightly polished by AI, and text where AI did the heavy lifting. The company's leadership has said this kind of partial AI use isn't inherently a problem, as long as it's disclosed.

The push for reliable AI detection is landing amid a string of public missteps. A Canadian official was recently caught reading an AI-generated prompt aloud during a legislative speech, and multiple attorneys have faced sanctions after submitting legal filings containing fabricated case citations produced by chatbots.

Academic institutions are tightening the screws too. The research repository arXiv rolled out a policy this year that can impose a one-year submission ban on authors whose papers show clear signs they never reviewed AI-generated output, such as leftover hallucinated citations or stray phrases like offers to make further edits.

Testing by a reporter found the tool consistently caught fully AI-written news articles, even after deliberate attempts to make the writing sound more human or to prompt chatbots into evading detection. But it wasn't flawless: some fully human-rewritten sentences were still incorrectly flagged as machine-made, and a full human-authored article scored as entirely human only when submitted in full rather than in fragments.

The image detector performed similarly well in hands-on trials, correctly spotting synthetic visuals across photorealistic and stylized styles, and even identifying an AI-generated image embedded within an otherwise real photograph. It wasn't perfect either, mislabeling at least one AI-generated image as authentic during testing.