This tech CEO thinks AI detection has an evidence problem

Undetectable AI rewrites machine-generated text so that it reads like a person wrote it. That is the whole product, and text is the only thing it touches. TruthScan, which Christian Perry started afterward, looks for fraud in images and documents: doctored photographs, synthetic images passed off as real, fabricated receipts, forged identity paperwork.

Stated that way it sounds like Perry is arguing with himself. He says the division is deliberate, and that it comes down to a question about evidence.

The problem with text detection

AI text detectors work by measuring statistical properties of writing. How predictable a given sentence is, which words the author reached for, how the rhythm falls. None of those properties belong exclusively to machines. Plenty of people write predictably, and some of them have been punished for it.

In 2023 a group of Stanford researchers ran seven widely used detectors against essays by non-native English speakers and published the results in the journal Patterns. The detectors called the essays AI-generated 61 percent of the time. One of them flagged nearly 98 percent. Handed essays by American eighth-graders, the same tools got it right more than nine times out of ten. James Zou, the paper's senior author, said schools should avoid the detectors wherever they can.

OpenAI had already reached a version of the same conclusion about its own work. The company released an AI text classifier in January 2023 and conceded at launch that it caught 26 percent of AI-written text while mislabeling human writing 9 percent of the time. It withdrew the tool that July, citing the accuracy, and said it would look into provenance techniques instead.

Underneath those percentages are people. Students accused of cheating on essays they wrote themselves. Applicants whose cover letters went in the bin. Freelancers who lost clients over a score that came with no appeal attached.

"There is no evidence chain in text," Perry says. "A number appears on a screen, and someone's academic record or job offer turns on it. We decided we were not going to build that."

Undetectable AI works on text and nothing else, and Perry says that scope was a decision rather than an accident. He drew the line there because text was where he could see the false flags landing on people.

His companies also sell an AI text detector, which he knows reads as a contradiction. The distinction he makes is about what the number is allowed to do rather than whether it should exist at all. A detector can tell writers how their own draft scores before someone else runs it. It can flag a document as worth a closer look. What it cannot do, in his account, is stand in as proof, and treating it as proof is where he thinks the field went wrong. The Stanford researchers landed somewhere similar, recommending the tools be used as aids and self-checks instead of as assessments.

Images and documents behave differently. A forged invoice leaves artifacts. A face-swapped video carries an edit history. A cloned voice on a call about a wire transfer tends to arrive attached to an actual attempted theft, with a dollar figure on the other end of it.

Undetectable AI now has more than 20 million registered users and sits among the 50 most-used AI tools in the world. Traffic across the company's sites has passed a billion visits. Perry has upwards of 70 people working for him, and Forbes put him on its 2026 30 Under 30 list in artificial intelligence.

Popsicles at 1 percent

Perry was nine when he borrowed $40 from his father at 1 percent interest, bought popsicles, and sold them on the beach. His father ran his own company and made him account for the cooler and the ice before calling any of it profit. By 13 Perry had a video game services business with two employees, both twice his age. He later studied at Boise State University.

ChatterQuant Inc. came in February 2021, a social media tracking data platform aimed at quantitative finance and built on large language models. Perry raised venture capital and ran sales while overseeing the product, until one of ChatterQuant's own customers bought the company. He went from that exit straight into the idea that became Undetectable AI.

Building Undetectable AI

The product launched with one capability. Proprietary algorithms rewrite AI-generated text so that it reads naturally. It found an audience quickly and expanded internationally under three additional brands, and the growth came off inbound leads, LinkedIn campaigns, and cold outreach rather than any real sales operation.

Perry had raised venture money once already at ChatterQuant. He funded this one out of its own revenue and put the attention on monetization.

He has also lined the company up with where the regulation is heading. Undetectable AI has adopted C2PA provenance credentials, the cross-industry standard for cryptographically signing content with its origin and edit history, and has built toward the disclosure obligations in Article 50 of the EU AI Act. Perry's argument for provenance is an evidentiary one, the same argument he makes about detection generally. A signed record of where a file came from can be checked. A statistical guess cannot. He notes that OpenAI said roughly the same thing on its way out of the detection business.

Perry sits on the Forbes Technology Council, where he wrote in June 2026 that companies should treat deepfakes as a primary vector of fraud. He has commented on synthetic media for Business Insider, NBC, and Fox.

TruthScan

TruthScan concentrates on the media where detection produces something checkable. Its core is image and document fraud: manipulated and synthetic images, altered receipts, forged identity documents, with additional coverage across video, audio, and text. Enterprises run the API inside identity verification, payment approval, claims review, and customer onboarding, which is where AI fraud tends to get into a business in the first place. Anyone else can drop a suspicious image into the public tools and get an answer in seconds.

The threat has a paper trail by now. Reuters reported in September 2025 that international gangs were using AI to make online romance scams more convincing. In a separate case, a finance employee at the engineering firm Arup was talked into transferring roughly $25 million after a video call on which every other participant turned out to be a deepfake. "If executives working at companies making millions of dollars can be duped, then anyone can," Perry says.

What he points to as the company's advantage is the adversarial work that preceded the product. Before building anything, his team ran red-team case studies on how detection systems get bypassed and where they fail, drawing on published research, public benchmarks, and its own attack testing against detectors already on the market. Those findings shape the model design. Perry is specific that this is a methodology and not a data pipeline. Undetectable AI's customer content is walled off, it is not used to train TruthScan's models, and the two products run as separate systems with separate data handling.

The company reports 99 percent accuracy on its internal image benchmark suite, documented by model, content type, and generation tool. Perry treats that figure as a measurement and says the output is a risk signal for human review.

TruthScan bought the detection site imagedetector.com in February 2026, a deal SecurityWeek included in its roundup of cybersecurity acquisitions for the month alongside moves by Check Point, Palo Alto Networks, and Zscaler. A partnership with the image platform DeepAI puts TruthScan's analysis in front of DeepAI users.

Some of the work lands in newsrooms. In March 2026, after a German public broadcaster mistakenly aired an AI-generated clip in a report, the fact-checking team at Euronews used TruthScan data to document more than 200 viral fake videos of US immigration agents. TruthScan also narrowed down where they were coming from. Every clip ran exactly 10 or 15 seconds, which matched the output options of one particular video generation model. It is the case Perry brings up when he wants to explain what he means about evidence. Nobody had to guess at a probability. The tell was sitting in the files.

How he works

Perry puts his own schedule at 70 to 80 hours a week for the past few years, holidays excepted, and tells founders to expect the early stage of a company to be demanding rather than balanced. At around 70 people he still writes copy and runs QA, on the theory that a chief executive either finds the person for the job or does the job. He works out of Las Vegas.

He turns down most of what comes across his desk, including plenty that would make money. His rule is that the only distraction worth taking is one with a chance of outgrowing the core business, and everything under that bar waits.

None of this makes him an opponent of the technology. Perry supports artistic and creative uses of AI, and describes what TruthScan sells as transparency, so that people know when they are looking at generated content and when they are looking at a photograph. New generation models keep arriving, which means the detection models keep getting retrained, which means the work does not have an end date.