Have you ever thought about the information you deleted years ago, but suddenly that information appears in your search result while searching for anything about yourself?
That information can be anything, whether it is your social media post or an old photograph, or your old written blogs and articles and/or any personal details that you once shared publicly.
You may even have successfully asked Google to remove the result from Search. Yet the information may have already travelled somewhere else, including into datasets used to develop artificial intelligence.
This raises a question that the internet was not designed to answer: Can a person really be forgotten by a machine?
The idea sounds simple because human forgetting is simple, like we remove our photos from the album and think it goes from the album forever. Also, think about it: we delete a file, and it goes to the trash folder, and then it is permanently deleted. But artificial intelligence does not work like this. Once data has been used to train a machine learning model, its influence may be distributed across millions or billions of learned parameters.
This is where the traditional right to be forgotten meets one of the hardest problems in modern technology.
The Right to Be Forgotten Was Never About Erasing the Internet
In 2014, the Court of Justice of the European Union decided “The Right to be Forgotten”. This decision specified that people may ask the Google search engine to remove those links that show their name in certain cases. Article 17 of the General Data Protection Regulation later established a broader right to erasure, although that right is not absolute. Google itself explains that the right generally concerns removing certain search results, rather than deleting information from the entire internet.
That distinction matters even more in the age of generative AI.
Google has introduced a tool called “Results About You”. This tool lets people ask Google to remove their information from search results. The information includes things like addresses, phone numbers, and email addresses. Google made it clear that taking a result out of search does not mean the data is gone from the original website. The data stays where it was posted.
So there are already two different questions: Can Google stop showing information about me? And can Google make its AI systems forget that information ever existed?
The second question is much harder.
Deleting Data Is Not the Same as Making AI Forget
A conventional database stores information in recognisable records. If a person’s record needs to be deleted, the system can locate that record and remove it.
An AI model is different.
During training, a model reads a lot of data. Changes its internal settings to recognize patterns. Personal information can sometimes be. Repeated, even though the original training material is no longer inside the model as a plain document.
This creates a difficult legal and technical problem. A 2023 study on the relationship between machine learning and the right to erasure found that because the connection between individual training examples and a model’s internal parameters is difficult to determine, enforcing the right to erasure in machine learning systems can be technically challenging.
The problem has become important enough to create an entire research field called machine unlearning.
The basic idea is straightforward: if an individual asks an AI system to forget their data, researchers want to remove the influence of that particular data without having to rebuild the entire model from scratch.
But “straightforward” is not the same as “solved”.
The New Science of Forgetting
Machine unlearning is now one of the fastest-growing areas of privacy research. A 2025 survey described it as an emerging approach for removing the contribution and influence of individual data from trained models, specifically in response to privacy concerns and regulations such as the GDPR.
Research published in Nature Machine Intelligence in 2025 went further by examining unlearning specifically in large language models. The researchers argued that successful unlearning should remove unwanted data influence while preserving unrelated knowledge and useful capabilities. They also identified major challenges involving the scope of unlearning, the relationship between data and model behaviour, and how effectiveness should actually be measured.
The difficulty is that nobody wants an AI model to forget everything.
Suppose a model was trained on millions of books and websites, and one person asks for their information to be removed. Ideally, the model would forget that person’s private information without suddenly becoming worse at answering unrelated questions.
That is the central challenge: how do we remove one person’s influence without damaging everything around it?
Recent research suggests that even this may be harder than it sounds. In 2025, a study showed that even after texts were said to be “unlearned” from language models, those texts could still be found with certainty. In words, when you ask a model to forget something, it does not automatically erase every sign of that something.
This should make us cautious about using the word “delete” when discussing AI.
What Does Google Actually Know About You?
Google says that its AI models are primarily trained using publicly available information from the open internet. It also says that it applies privacy protections such as data cleaning and removing duplicate information before training, with the aim of reducing the possibility that models reproduce personal information.
But “publicly available” does not automatically mean “fair to use forever.”
A person may have published something publicly when they were 18 and changed their mind at 28. A photograph may have been uploaded by someone else. A news article may contain a person’s name even though that person no longer wants the association. A website may disappear while copies remain elsewhere.
The digital world was built around persistence. AI is now adding another layer to that persistence.
The UK Information Commissioner’s Office has acknowledged this problem directly. Its guidance states that individual data protection rights apply when personal data is used to train and fine-tune generative AI models. It also notes that organisations may need to consider requests involving data contained in training datasets and that, in some situations, fulfilling an erasure request could require retraining a model.
The guidance makes an important distinction, however. Removing a person’s data from a training dataset does not automatically mean that every model trained on that dataset must be deleted. The real question is whether the model itself contains the data or if the personal data can be inferred from the model.
This distinction may become one of the important privacy questions of the AI era.
The Law Is Moving, but Technology Is Moving Faster
European regulators are beginning to address this problem more directly.
In December 2024, the European Data Protection Board gave an opinion that looks at how personal data can be used when building and using AI models. I read the opinion. It asks when an AI model can be called anonymous, what legal basis is needed to process personal data, and what happens if an AI model is built with personal data that was processed illegally.
The EU AI Act also reinforces the idea that privacy and data protection should be considered throughout the lifecycle of AI systems. The lifecycle of AI systems should include practices such as data minimisation and privacy-by-design principles. The focus on privacy and data protection remains strong throughout the lifecycle of AI systems.
The direction is becoming clear: people should not lose their existing privacy rights simply because their information has entered an AI pipeline.
But legislation cannot solve a technical problem by itself.
A law can say that personal data must be erased. It cannot instantly tell a neural network which mathematical changes need to be reversed.
So, Can We Really Delete Our AI Identity?
At present, the honest answer is sometimes, but not reliably enough to promise complete digital forgetting.
A person can request the removal of certain personal information from Google Search. Also, data protection laws have given people the right to erasure, objection, and other forms of control over their information. Organisations can remove data from training datasets, retrain models, or use machine-unlearning methods. Nevertheless, none of these methods ensure that people’s information has been deleted permanently from every model, dataset, backup, derivative system, or future training process.
This creates a strange contradiction.
The law increasingly recognises that people should have some control over their digital identity, while AI systems are being built to learn from enormous quantities of information and retain patterns that are difficult to trace back to their sources.
Perhaps the real challenge is therefore not teaching machines how to forget.
It is deciding what forgetting should mean in the first place.
If removing your name from Google Search does not remove the underlying webpage, and removing your data from a training dataset does not necessarily remove its influence from a trained model, then the phrase “right to be forgotten” may promise more than technology can currently deliver.
The next stage of digital privacy will depend on closing that gap.
People should not have to choose between living openly online and losing control over their identities forever. If AI is going to learn from our digital lives, then the ability to say “this is mine, and I want it removed” must become more than a legal principle.
It must become a technical possibility.
Until then, the right to be forgotten remains less like a delete button and more like a negotiation with a machine that has already learned too much.
If you want to submit your articles and/or research papers, please visit the Submissions page.
To stay updated with the latest jobs, CSS news, internships, scholarships, and current affairs articles, join our Community Forum!
The views and opinions expressed in this article/paper are the author’s own and do not necessarily reflect the editorial position of Paradigm Shift.
Hooria Akbar is an independent researcher and content writer with a strong interest in contemporary issues across AI, technology, society, and emerging trends. She writes research-informed articles, thought pieces, and blogs on a wide range of topics, aiming to present complex ideas clearly and engagingly for diverse audiences.
