Former OpenAI employees launch tool to see if AI has permanently memorized your identity

A newly launched website called In the Weights has captured the attention of the tech industry by offering individuals a direct way to see if their names and identities have been permanently memorized by the world's most powerful artificial intelligence systems. Developed by Joey Flynn and Thomas Dimson, two former OpenAI employees, the interactive tool assigns a strength score to queried names, illustrating how deeply a person's identity is baked into the neural networks of leading generative models[1][2]. While historical giants and modern global superstars like Wolfgang Amadeus Mozart, William Shakespeare, and Taylor Swift predictably command the maximum possible strength score of 996[3][4], the project is sparking a much broader conversation about data privacy, model transparency, and the fundamental mechanics of how AI systems store knowledge[5].

To understand the significance of appearing on the website, one must first grasp what model weights represent in modern deep learning[6]. When a large language model is trained on massive datasets scraped from the internet, it does not store a literal copy of the text. Instead, it adjusts billions of internal numerical parameters, known as weights, which dictate how signals flow through the artificial neural network[7][8]. These weights act like a vast system of dials, defining the model's core capabilities, reasoning patterns, and stored knowledge[7][9]. If a person's biographical details are embedded in these weights, the model can recall who they are instantly and natively, without needing to perform a real-time web search or rely on retrieval-augmented generation[2]. In essence, they have become a permanent fixture of the AI's frozen, post-training memory[9].

The methodology behind the platform relies on cross-model comparisons across thirteen prominent artificial intelligence models[4][10]. The site queries each model with a prompt asking for a list of individuals matching a specific name, along with short descriptions and the system's self-reported confidence[10]. The platform then clusters the resulting descriptions to resolve ambiguities, especially for common names, and calculates an overall strength score[10]. This score represents an average of how strongly each individual model recognized the subject, supplemented by a bonus for being identified across a higher number of different systems[10]. The participating models evaluated by the tool reflect the current cutting-edge landscape of generative technology, including frontier architectures such as OpenAI's GPT-5.5, Anthropic's Claude Opus 4.8, DeepSeek V4, Grok 4.20, and Meta's Llama series[4].

Model size plays a critical role in determining who makes it into these neural pathways[10]. The developers point out that smaller models have a much harder time memorizing specific biographical details because their storage capacity is heavily constrained[2][10]. For instance, Meta's Llama 3.2 1B model possesses only about one billion parameters, which roughly equates to a single gigabyte of memory[4][10]. For an individual to register in a model of that size, they must have a highly prominent digital footprint, as the network is forced to prioritize only the most frequent and culturally significant information from its training corpora[10]. Conversely, larger models with hundreds of billions or even trillions of parameters have the capacity to retain highly obscure, niche figures, making them far more likely to recognize average professionals, software developers, and localized public figures.

The creators of the website bring a wealth of practical experience in product design and machine learning to the project[5]. Joey Flynn and Thomas Dimson joined OpenAI following the acquisition of their startup, Global Illumination, which was OpenAI's first public acquisition[11][12]. Prior to that, Flynn spent years as a product designer at Facebook and Instagram, helping shape massive user experiences like Facebook's Timeline and Instagram's camera[13][12]. Dimson is widely recognized in the tech world as the engineer who built Instagram's core algorithmic recommendation engine, and at OpenAI, he contributed directly to the development of the Sora video generation model[14][15]. Their background as hands-on builders is evident in the platform's gamified, retro-pixel design, which transforms a highly complex technical topic into an accessible public leaderboard[16][17].

Despite its engaging interface, the tool highlights several inherent limitations of modern large language models, which the creators are quick to point out[2]. Because generative models are probabilistic rather than deterministic, they are prone to hallucinating biographical facts, inventing awards, or merging the lives of different individuals with similar names[2][10]. Additionally, minor typographical errors or variations in spelling can dramatically lower a person's strength score because the AI might fail to link the misspelled prompt to its trained parameters[2][10]. Common names also present a significant challenge, as the models frequently struggle with disambiguation, diluting the overall recognition score compared to individuals with highly unique names[2][10]. The developers emphasize that the site is not a rigorous scientific study, but rather an exploratory visualization of how information is distributed inside artificial minds[16][5].

The broader implications of this project for the artificial intelligence industry are profound, particularly concerning user privacy and data governance[5]. For years, major AI developers have maintained a high level of secrecy regarding the specific web pages, books, and articles used to pre-train their models[18]. By allowing the public to easily search for names, the website acts as a decentralized audit of these closed-source datasets. It provides concrete evidence of whose personal data was ingested, processed, and ultimately memorized without explicit consent. Under robust privacy frameworks like Europe's General Data Protection Regulation, citizens possess a "right to be forgotten," which legally mandates the deletion of their personal data upon request. However, because a trained model's weights are permanently fixed and highly interconnected[9], surgically removing a single person's identity from a neural network remains an unsolved computer science challenge, typically requiring either a complete and prohibitively expensive retraining process or still-experimental machine unlearning techniques.

Beyond legal and privacy debates, the platform also sheds light on an ongoing architectural dispute within the AI research community regarding the future of knowledge storage[19][20]. Many researchers argue that the industry's current practice of forcing models to memorize factual information directly within their weights is highly inefficient, leading to bloated, expensive systems that quickly become outdated[18][20]. An emerging school of thought suggests that future models should be designed primarily as reasoning engines with minimal factual memorization, relying instead on external databases, knowledge graphs, and dynamic search tools to retrieve specific facts as needed[19][20]. By visualizing just how much biographical data is currently hardcoded into the weights of systems like GPT-5.5 or Claude Opus 4.8[4], the website illustrates the sheer scale of the brute-force memorization that defines the current generation of generative AI[19][20].

Ultimately, the website serves as a compelling mirror for our increasingly digitized society, transforming abstract discussions about data scraping and machine learning into a deeply personal realization. It demonstrates that the digital footprints we leave behind are no longer just passive entries in search engine indexes, but are actively being synthesized into the literal cognitive structures of the next generation of computing. As the boundary between human identity and artificial intelligence continues to blur, tools that demystify the black-box nature of these models will become increasingly vital[21][5]. By revealing who is stored in the weights, the project not only redefines our understanding of digital fame[22], but also prompts a necessary reckoning over who owns our digital legacies in the age of artificial intelligence.