Ai Model Watermarking | AI Final Year Project Report

Complete Technical Paper Content — Ai Model Watermarking

1. Executive Abstract, Problem Formulation & Background

AI watermarking is a technique that modifies the output of generative AI models, such as large language models (LLMs), allowing it to be later identified as AI-generated. Watermarks are one of several techniques used for AI content detection, and include both traditional digital watermarking techniques on media files and metadata, and text watermarking techniques specific to LLMs that have been developed during the 2020s AI boom. Recent research in text watermarking has focused on statistical watermarking techniques, which were first introduced in 2022, with the first production deployments in 2024. Most deployed text watermarking algorithms use a secret only known to the specific model provider, and the detection algorithm requires knowing the secret, as the text is not distinguished in a generally human or machine-readable way. As a result, determining if text is watermarked requires access to the secret, or sending the text to an API provided by the model provider, and there is no general-purpose way to determine if text is watermarked without knowing what model produced it and access to the corresponding detector. Watermarking has been one focus of the regulation of artificial intelligence, examples including the European Union's Artificial Intelligence Act and the California AI Transparency Act, requiring text, images, audio, and video output by new providers' AI models to be watermarked and detectable from 2 August 2026, with existing providers being affected from 2 December 2026. Analyst criticism includes how these laws and resultant global industry policies not only establish mass scanning infrastructure readily applicable to citizens' personal communications, but also actively allow and drive this usage. Technical realities mean all content in scope is affected, AI and non-AI, because approved detection methods require sending both away for scanning. Effectiveness-wise, watermarking has been criticized for being incapable of detecting artificial content from motivated groups and individuals.

2. Related Work, Literature Survey & Algorithmic Foundations

Sora is a discontinued text-to-video model and social media app developed by OpenAI. Using artificial intelligence, the model generated short video clips based on prompts, and could also extend existing short videos. In February 2024, OpenAI previewed examples of its output to the public, with the first generation of Sora released publicly for ChatGPT Plus and ChatGPT Pro users in the United States and Canada in December 2024. The second generation of Sora was released to select users in the US and Canada at the end of September 2025. Sora 2 integrated social media features into the app. The app was shut down on April 26, 2026 and the application programming interface (API) was discontinued on September 24, 2026, marking the end of the Sora AI brand as a whole. By default, the generator used copyrighted material in its videos, unless copyright holders actively opt out of having their content included. Videos contained a visible, moving digital watermark to prevent misuse, but a week after Sora 2's release, third-party programs became available which could remove the watermark.

3. Dataset Ingestion, Preprocessing & Stochastic Augmentations

Several other models capable of generating video from text had been created prior to Sora, including Meta's Make‑A‑Video, Runway's Gen‑2 and Google Veo. OpenAI, the company behind Sora, had released DALL·E 3, the third of its DALL-E text-to-image models, in September 2023.

4. Neural Network Backbone, Layer Specifications & Loss Formulations

Sora 2 was unveiled on September 30, 2025, with an iOS app at the same time, as well as an Android app two months later. All videos generated by the model feature a visible, moving watermark to prevent misuse of the tool. The previous version of Sora also added a safety watermark to allow viewers to distinguish between real and fictional content. On October 7, 404 Media reported that third-party programs that could remove the watermark from Sora 2 videos had become prevalent. Many outlets, such as Wired magazine, have noted that the Sora 2 app is overtly similar to TikTok in style and features.

5. Training Protocol, Convergence Behavior & Experimental Evaluation

The technology behind Sora is an adaptation of the technology behind DALL-E 3. According to OpenAI, Sora is a diffusion transformer, a denoising latent diffusion model with one transformer as its denoiser. A video is generated in latent space by denoising 3D "patches", then transformed to standard space by a video decompressor. Recaptioning is employed to augment training data by using a video-to-text model to create detailed captions for videos. OpenAI trained the model using publicly available videos as well as copyrighted videos licensed for the purpose, but did not reveal the number or the exact source of the videos. Upon its release, OpenAI acknowledged some of Sora's shortcomings, including its limited capacity to simulate complex physics, to understand causality and to differentiate left from right. OpenAI also stated that, in adherence to the company's existing safety practices, Sora will restrict text prompts for sexual, violent, hateful or celebrity imagery, as well as content featuring existing intellectual property. Sora researcher Tim Brooks stated that the model learned how to create 3D graphics from its dataset alone, while fellow Sora researcher Bill Peebles said that the model automatically created different video angles without being prompted. According to OpenAI, Sora-generated videos are also tagged with C2PA metadata to indicate that they are AI-processed.