The Center for Data Innovation recently spoke with Anatoly Kvitnitsky, CEO of AI or Not, a California-based company using AI models to detect and recognize images, speech, videos, and other factors created by AI tools. Kvitnitsky explained how the company uses automated systems to scan online content and quickly update its detection models in response to emerging AI tools and trends.
David Kertai: How does your tool detect AI in various media?
Anatoly Kvitnisky: Our detection system is tailored to each type of media. For images, we analyze pixel patterns and compare them to patterns seen in real photographs, such as those taken by iPhones or Canon cameras. AI-generated images tend to have distinct visual traits that real-world optics don’t replicate.
For text, we examine sentence structure and word usage. Certain phrases and constructions appear far more frequently in AI-generated writing, especially before tools like ChatGPT became more advanced in late 2022. These linguistic fingerprints help us flag synthetic text.
With audio, we split the analysis between voice and music. We generate a spectrogram, a visual representation of sound, and compare its structure to that of human-generated audio. AI still struggles with certain phonetic nuances, like the harsh consonants in Russian, which can be visually identified.
For video, we analyze it frame by frame to detect synthetic visuals and deepfake alterations. We also study motion patterns; AI-generated video often features unnatural smoothness or gravity-defying movement. Audio cues are also integrated to strengthen the overall analysis.
Kertai: How does AI or Not detect deepfakes like face swaps or altered expressions?
Kvitnisky: We detect deepfakes by training our models to recognize the subtle digital artifacts these techniques leave behind. While AI-generated images are created from scratch, deepfakes manipulate real content, often swapping faces or syncing lips to change meaning or context. These alterations create inconsistencies in pixel structures or transitions that our models are trained to spot.
Each deepfake method, such as face swaps or lip syncing, leaves behind different clues. For example, after the assassination attempt on Donald Trump, a widely circulated deepfake altered the expressions of Secret Service agents to make them appear to be smiling. Our system can catch these fabricated details by identifying irregularities in the image.
Kertai: How often are your detection tools updated to keep up with new AI models?
Kvitnisky: We update our detection models several times per week to keep pace with new AI tools and trends. We monitor the internet using bots that flag potential drops in detection accuracy. If something changes in the data patterns, like those triggered by a new model, we investigate and adjust accordingly.
We also actively track the release of prominent models such as ChatGPT 5.0, as well as less-publicized models from countries like China. New models are often followed by waves of synthetic content, particularly during high-profile global events or elections. These surges can rapidly increase the spread of misinformation, so it’s critical that our tools remain current and adaptive.
Kertai: Does AI or Not explain its results or show how confident it is?
Kvitnisky: Yes, we provide a confidence score indicating how likely the content is to be AI-generated, and we also estimate which model may have created it. We aggregate results from multiple detection models to give users a reliable signal. This matters because not all flagged content has the same consequences, for example, distinguishing a fake movie poster from a forged financial document is critical in different ways.
Kertai: Who uses AI or Not, and how is it applied in the real-world?
Kvitnisky: Our platform is used across a wide range of industries, including media, journalism, fintech, and government. Organizations rely on us for identity verification, fraud detection, and content authentication. For example, we help identify fake checks, detect synthetic images in breaking news, and flag AI-generated audio used in scams or impersonation attempts.
As AI-generated content becomes more common, and more convincing, it’s being weaponized to impersonate people, rewrite historical events, or manufacture visual evidence. Our goal is to give professionals the tools they need to quickly and confidently assess whether something is real or fake.
