Here’s what’s on our plate today:

• 🧪 Why teaching AI to forget is becoming big business.

• 📰 Physical AI hits the factory floor, Claude goes to law school, Notion becomes agent HQ.

• 💡 Roko's Pro Tip: design for deletion from day one, retraining isn't a plan.

• 🗳️ Poll: How should the industry handle AI's permanence problem?

Let’s dive in. No floaties needed…

The Laboratory

• Legal meets mathematical stubbornness: AI models bake training data across billions of parameters, making true deletion nearly impossible, which puts them on a collision course with GDPR’s Right to Be Forgotten and similar laws.

• Bias gets locked in: France’s equality watchdog flagged Meta’s job ads for gendered targeting in 2025, illustrating how discriminatory patterns become permanently embedded once models train on flawed data.

• Unlearning is now investable: Startups like Hirundo ($8M seed) and Protect AI ($100M+ raised) are building tools to selectively erase data influence without the cost of full retraining, while global AI startup funding topped $200B in 2025.

• Forget or fall behind: Companies that master machine unlearning gain a real competitive edge in privacy and compliance, while those that can’t risk regulatory exposure, ballooning legal costs, and models that calcify around their worst historical assumptions.

Why teaching AI to forget is becoming big business

For centuries, philosophers and learning theorists have argued that true understanding requires the ability to unlearn: to question outdated assumptions, let go of flawed beliefs, and make space for new knowledge. At a societal level, this process has helped civilizations correct historical injustices and move beyond systems shaped by misinformation or obsolete thinking. Today, as artificial intelligence increasingly influences decisions across finance, healthcare, education, and governance, this principle is becoming increasingly critical.

Systems that cannot revise or forget outdated information risk reinforcing flawed assumptions at an unprecedented scale, and that is precisely the challenge modern AI is beginning to confront.

When learning systems cannot let go

Because AI models are trained on enormous volumes of internet data, removing harmful, inaccurate, or sensitive information after training has become extraordinarily difficult. Unlike traditional databases, AI systems do not store information in neat, removable records. Once data contributes to training, its influence spreads across millions or even billions of neural network parameters, becoming embedded deep within the model’s internal structure.

This creates a growing challenge as AI systems become integrated into organizational workflows and decision-making. The modern internet contains vast amounts of personal information from social media platforms, forums, blogs, and digital archives, much of which is scraped for AI training. Once absorbed into a model, tracing and removing the influence of specific individuals’ data becomes immensely complicated.

The issue is becoming increasingly significant as legal frameworks such as the Right to Be Forgotten (RTBF) gain traction across multiple jurisdictions. While RTBF assumes personal data can be deleted upon request, AI systems complicate that principle because neural networks do not retain information the way conventional databases do. The result is a growing conflict between legal expectations of erasure and the mathematical permanence of machine learning.

When legal rights collide with mathematical memory

Under Article 17 of the European Union’s General Data Protection Regulation, citizens have the right to request that their personal data be erased. In traditional computing systems, fulfilling such a request is relatively straightforward because the data is stored in identifiable locations within structured databases. This allows organizations to locate the relevant records, remove them, and demonstrate compliance with the request.

However, in modern AI systems, once personal data is used to train a machine learning model, its influence spreads throughout the network’s internal parameters. A user’s shopping behavior may shape an e-commerce recommendation engine, while medical records may influence a diagnostic model or insurance risk assessment system. By the time training is complete, the data no longer exists as an isolated entry that can simply be deleted. Instead, its influence becomes embedded in the model’s mathematical structure.

As AI systems move deeper into hiring, lending, healthcare, and governance, this technical limitation is increasingly colliding with legal frameworks such as the Right to Be Forgotten, which assume personal data can be meaningfully erased upon request.

The cost of irreversible learning

The practical implications of this challenge are already emerging across industries.

In 2025, France’s equality watchdog ruled that Meta’s job-advertisement algorithms displayed certain employment opportunities disproportionately to men or women, reinforcing discriminatory exposure patterns before candidates even applied. Cases like these illustrate how difficult it can become to isolate and remove biased historical assumptions once they are embedded in large-scale AI systems.

For enterprises, the operational consequences can become severe. Retraining large-scale AI systems from the ground up can consume enormous computational resources, require extensive engineering labor, and temporarily degrade system performance for millions of users. At the scale of frontier AI systems, retraining costs can easily climb into $100k or even millions of dollars, depending on the size and complexity of the model.

From technical debt to regulatory exposure

The growing permanence of AI memory is increasingly transforming into a legal and financial liability. Gartner projects that illegal AI-informed decision-making will contribute to a significant rise in legal disputes involving technology companies by 2028 as regulators intensify scrutiny around automated systems and algorithmic accountability.

In response, companies are rapidly investing in explainability frameworks intended to make AI decisions more transparent and auditable. Yet transparency alone does not solve the deeper problem. Understanding how a model arrived at a harmful outcome still leaves unresolved the question of how to remove the underlying influence that produced it.

As a result, the industry is increasingly shifting its focus to machine unlearning: the development of technical methods that selectively remove learned information from trained AI systems while preserving overall model functionality.

The business of teaching machines to forget

The rise of machine unlearning is also creating a new market for AI trust, safety, and governance infrastructure. Companies are developing systems that audit training datasets, monitor bias, detect harmful model behavior, and automate selective forgetting.

Platforms such as IBM’s AI Fairness 360 toolkit are evolving beyond passive bias detection toward more active model-intervention systems, while startups focused specifically on unlearning are attracting growing investment.

Beyond IBM, investment is rapidly flowing into the emerging AI trust, governance, and machine-unlearning sector as enterprises prepare for stricter regulation around automated decision-making. In 2025, machine-unlearning startup Hirundo raised an $8M seed round to develop systems that can remove hallucinations, bias, poisoned data, and sensitive information from trained models without full retraining. Meanwhile, AI security and governance firm Protect AI has raised more than $100M as demand grows for infrastructure that can monitor, secure, and govern AI systems at scale. The broader market is also expanding rapidly alongside the AI boom itself, with global AI startup funding surpassing $200B in 2025, driving demand for tools focused on auditing, explainability, compliance, and machine unlearning.

Organizations that successfully develop reliable unlearning systems may gain a significant competitive advantage by offering stronger guarantees of privacy, portability, and user control. In an increasingly regulated AI environment, the ability to remove data influence efficiently may become as commercially important as the ability to train advanced models in the first place.

The ancient philosophy returning through machines

What makes this emerging segment of the AI industry particularly striking is how closely it echoes philosophical ideas that long predate modern computing. Traditions centered on unlearning long recognized that intelligence depends on the ability to let go of outdated assumptions and create space for new understanding.

AI is now confronting the same principle at a technological scale. As AI systems become deeply embedded in employment, finance, healthcare, and education, their ability to forget grows important. Systems that endlessly accumulate information without mechanisms for revision or erasure risk becoming rigid, biased, and trapped by historical patterns.

Machine unlearning, therefore, represents more than a technical or regulatory challenge. It reflects a broader realization that enduring intelligence, whether human or artificial, requires the capacity to adapt, revise, and let go.

Roko Pro Tip

Monday Poll

Bite-Sized Brains

• Physical AI hits the factory floor: Humanoid robots are moving from demo videos to real factory deployments, marking the start of physical AI's commercial era.

• Claude goes to law school: Anthropic is expanding Claude's tools for law firms and lawyers, doubling down on legal as one of its most profitable verticals.

• Notion becomes agent HQ: Notion just turned its workspace into a hub for AI agents, positioning itself as the operating layer for how teams actually use AI.

Meme Of The Day

The Toolkit

• Assembly AI: Speech-to-text API that handles transcription, speaker detection, and audio intelligence for production apps. 

• Chroma: Open-source vector database built for AI apps, fast to set up and easy to scale for RAG and embeddings. 

• Continue: Open-source AI code assistant that plugs into VS Code and JetBrains with full control over models and context.

Rate This Edition