Introduction
The EU Court of Justice, on the 13th of May 2014, ruled that a data subject (someone whose data is being used), under Article 12(b) and subparagraph (a) of Article 14 of Directive 95/45/EC, has the right to order the operator of a search engine to remove links to web pages or indexes containing their personal information, following the request of Mr Mario Costeja Gonzalez. Mr Mario tendered a complaint with the Agencia Española de Protección de Datos (AEPD) against a newspaper publisher, Google Inc., and Google Spain SL, requesting the involved companies to remove links to information about him being in debt. He claimed that the debt had been resolved a long time ago, and the information was no longer relevant.
The ruling emphasised the basic right of an individual to privacy regarding the use of their data. However, the AEPD rejected Mr. Mario’s request against the newspaper publisher, La Vanguardia Ediciones SL, finding the grounds for their publication of the information to be legitimate.1 This case bridges the gap between an individual’s right to privacy (including their right to be forgotten) and the public’s right to access to information.
The issue is also extensible to AI models and not limited to search engines alone in today’s context. This is because, as we will see later in this article, AI models are trained on data sourced from the internet, an entity that search engines contribute vastly to. Trained AI models are employed to provide accurate information about relevant domains. This, in a way, reinforces our right to access accurate information. Conversely, these models require data —personal and honest ones —to function accurately. This article aims to provide an insight into what the Right To Be Forgotten (RTBF) in the GDPR is (because GDPR is the model for data laws worldwide), how AI models are trained, and how the immutability of AI models threatens the RTBF. Furthermore, we look at what Amnesiac Machine Learning is, its usability, feasibility, and how it reconciles RTBF and AI immutability.
The Right To Be Forgotten (RTBF)
Everyone is entitled to security against the illegitimate use of their privacy and control over the same. Giving an individual control over their privacy (in any form, including the use of their data) also provides insight into what the RTBF might mean. The right to be forgotten or the right to erasure, according to Article 17 of the GDPR, states that a data subject has the right to request from a data controller the erasure of their data with immediate effect due to one or more of the reasons stated by the law.
The data controller is then obligated to provide a swift response to this request.4 Of course, this is irrespective of where and why the data is being used, be it in the training of an AI model that aims at bettering the existence of humankind, or an ongoing research where this withdrawal might cause a massive disruption. The immutability of AI models poses a threat to this fundamental right to privacy, as we will explore in this article by examining the development of a highly popular and versatile AI model– the GPT-4.
Immutability In AI Models
AI models, during development, have to go through a pre-training phase where data is fed into them so that they can recognise patterns, draw inferences, make calculations, and even make future predictions, as in the case of the development of GPT-4. This data is what the models use to learn, hence it constitutes the building block of AI. Data is the major raw material in the development of these models. But how is data fed and stored into these models? Can data possibly be withdrawn from them? What happens to the model’s performance upon withdrawal of a data set? Let us walk through these questions by using the development of GPT-4 as a case study.
GPT-4 is capable of processing multiple data types to produce the necessary outputs, ranging from text to video, audio, and generative text. Thus, we can refer to the model as being multimodal. The model was trained through two distinct phases: the pre-training phase and the post-training alignment phase.5 These two phases describe how data is stored and used in the model.
The data used for the pre-training phase was sourced from publicly available data, including Reddit posts, as well as data from third-party data providers.5 OpenAI could boast of consent as per the use of the training data from its sources. However, as clearly stated in the GDPR, data subjects can demand the deletion of their data from data controllers. Considering this, OpenAI did not explicitly state in its privacy policy whether this deletion also applies to the learned patterns in its system’s architecture. In Article 17 of the GDPR, users can withdraw their consent from the agreement on which the use of their data is based, which means that data used in the training process of LLMs is liable to consistent alteration, and knowledge patterns in the models should not be static. This provides a framework for using personal data. Now, if deletion is guaranteed upon request, there definitely should be a way for the model to forget. Otherwise, the guarantee is only there for show.
In the pre-training phase, the model is built as a transformer-based model, primarily aiming at enabling the model’s predictability of the next unit of text in a document through a series of algorithms. It should be noted that the data cutoff used in this phase is September 2021, making the model oblivious to possible data alterations from after that period.5 Therefore, GPT-4 does not learn from experience, no matter how slowly transient the experience is. The technical report made available by OpenAI to describe the development process of GPT-4 did not specify explicitly the process of feeding data into the model in this phase, but it did specify that data in the model is not retrievable like traditional databases for alteration because of the way knowledge from the data set on which it was trained is encoded into it.
The pre-training phase feeds the model with data, and the model eats and digests the data into knowledge. Hence, the immutability of AI models.
The Post-training Alignment Phase: The Doppelganger of The RTBF
The post-training alignment phase is the second and final phase in the training of the GPT-4 model. This phase aims to align with the objectives of using the model, as well as refining its personality. Here, the model’s behaviour is preferably fine-tuned by Reinforcement Learning from Human Feedback (RLHF).5 In relatable English, this stage filters out some words and concepts from all the possible outputs, thereby improving its safety and adherence. However, the model raised some significant risks and limitations that challenge the RTBF even after this stage of training, especially privacy violations: the model, having learned from publicly available data, is capable of memorising and reproducing people’s data accidentally, raising privacy concerns.
Although OpenAI claims to have mitigation standards against the risks that may arise as far as an individual’s privacy is concerned , the pre-training stage still makes it clear that AI models, or in our case study, GPT-4, cannot forget due to the way data is represented in the model. Is the post-training phase some type of doppelganger for the RTBF? It is difficult to take a stance here because the model is prone to making hallucinations even after fine-tuning. There is no clear claim by OpenAI that this stage aims to adhere to the RTBF. Under the hood, the company expects us to settle for this development while the erasure problem persists in the industry because, of course, adhering immediately to a user’s data deletion request would mean the development of a new model. However, fine-tuning the model to not work with the data point in its memory is quite feasible economically.
RTBF Enforcement vs The Erasure Problem
Non-compliance with the RTBF by data controllers, according to the GDPR, imposes penalties on the defaulting data controllers. OpenAI was fined 15 million euros for illegally processing personal data. Amidst the erasure problem in the AI industry, the EU has not, in any form, compromised the RTBF, not even settling for the post-training alignment phase as a compliance with the RTBF. Can we, then, say that the GDPR is challenging AI industries to conduct research to tackle the problem? Affirmative. We could as well call it a challenge due to the present lack of scalable erasure of a user’s data from the memory of an AI model. Is the law, then, unfair to ask for what technology cannot yet provide? The law is probably embracing the claim that nothing is impossible in technology, holding technology companies to their word.
Immediate compliance with a user’s data deletion request in AI models would require a build of another model from the ground up, as stated earlier, which can be demanding, considering the cost of building a versatile model like GPT-4. On the flipside, these companies also risk huge fines should they not comply with the RTBF in the GDPR.11 This emphasises how AI companies are far from full compliance, and are not ready to bear the penalties that may come with non-compliance– a huge gap between innovation and law. It is then hard to tell whose responsibility it is to be fair: there needs to be a bridge to reconcile both worlds. Either machines are taught to forget completely, or the law is considerate enough to realise how expensive the RTBF is costing AI companies, whereas being considerate would only be to the detriment of the populace. Bridging this gap would require that we recognise what AI immutability means in context.
AI Immutability in Context
The law could be adamant from her perspective that AI mutability is not impossible in the real sense. Playing devil’s advocate, the law rather sees claims of immutability as AI companies not ready to rebuild their models immediately, as required by the law, upon a user’s deletion request. It is paramount to recognise that training data is different from knowledge representation in AI models, and that AI companies may want to hide under this distinction to prove themselves faultless. It is, however, also necessary to note that a single user’s data contributes to this pattern of knowledge, no matter how large the data used to train the model is.
Immutability might perhaps be more of an economic impracticality than a technical one, as we have seen that it is not impossible per se. If models can be rebuilt from the ground up to balance the models upon deleting a data set, why claim immutability as a technical issue? From the technical lens, immutability is portrayed as an architectural limitation, as we have seen in the development of GPT-4. This limitation has to do with how data is embedded into the system’s architecture during training. As stated earlier with a metaphor, the system digests the data such that it would take interrupting its digestive system to alter the data.
Contextually, the law insists on the possibility. Perhaps it would have been another case if it were not possible at all. Economics also presents a major setback in adherence to the law, but the GDPR won’t take any setback as an excuse for not protecting data privacy, not even economics. AI companies and researchers have been looking for a solution to contextual immutability, exploring the process of machine unlearning.
Machine Unlearning
Retraining a model once a data deletion occurs can be quite impractical in terms of time and memory architecture. Aside from the erasure problem, AI models are also prone to attacks such as Membership Inference Attacks (MIAs), where attackers can compromise specific training records embedded in a model’s training data. In an MIA, it is possible to tell with high accuracy if a user’s record was used in the training of an AI model, even when the data has been deleted from the model’s database. Machine unlearning is the cavalry for AI companies to fully embrace the law’s demand for data privacy.
Machine unlearning is simply trying to selectively remove the influence of data points from the knowledge representation patterns of a model. In simple words, it is making the model adjust itself when data is removed from it, thereby drafting new knowledge representation patterns innits architecture without having to retrain it. The study is still in its early stages, although there have been breakthroughs, and of course, challenges.
As we have seen from the technical lens that immutability in AI models is from the economic posture; machine unlearning is solving the problem by posing as being economically feasible and technically possible. The major aim of the solution is to keep AI companies in full compliance with the GDPR, thereby maintaining integrity on the path of AI companies.
AI companies are making efforts to make machine unlearning scalable. This, in a way, shows their readiness to be reconciled with the right to erasure, to be reconciled with the law, as AI giants will keep being at loggerheads with the law as long as the erasure problem persists. Google, in 2023, announced its first machine unlearning challenge, where applicants proposed possible solutions to the erasure problem. This demonstrates the company’s commitment to law adherence by employing possibilities from brilliant minds outside the company. Also, OpenAI announced the development and launch of Media Manager in 2024– a tool that will enable creators and data owners to specify how they want their content to be used in machine learning and research.
AI giants recognise the erasure problem as a setback in protecting private data, and are employing machine unlearning and other possible solutions to bridge the gap between technology and law, because the GDPR is not ready to compromise its standards and the digital security of those within its jurisdiction.
Conclusion
While AI companies are obligated to mitigate the erasure problem in their models as required by the law, they argue that the architecture of their products does not allow for the practicality of the RTBF, but as we have seen in this article by exploring the training process of GPT-4, it is more of an economic problem than a technical impossibility. Although AI companies also claim fine-tuning as a way of making their systems forget the influence of a user’s data on their knowledge patterns, these systems are still prone to “hallucinations”, as clearly stated in the technical report of GPT-4.
The GDPR’s unwillingness to accept AI immutability as an excuse for AI companies not to comply with the RTBF portrays the law’s integrity, which is to be commended. AI companies are also researching several ways to make their systems forget the influence of a user’s data without having to build their models by looking into machine unlearning/ amnesiac learning.
Machine unlearning is simply a study of ways to make AI models develop new patterns of knowledge upon data deletion, and it is in its early stages. Companies are making efforts to bridge the gap between the economic problem that is posing as immutability and the law, showing their readiness to comply with legal terms and requirements.
