Lawyer Lin Shanglun argues that as generative technology pushes the cost of producing text, code, and charts close to zero, the assets with real defensive value may be the ones algorithms cannot freely reproduce online.
In his commentary, Lin uses the recent strength in physical Pokemon card and vintage sports card auctions to make that case. He says those collectibles have not been displaced by digital technology. Their value has stood out because of scarcity, historical consensus, and offline grading and authentication.
Physical collectibles as an example of “AI-resistant” assets
Lin writes that after speaking with engineers in Silicon Valley and senior industry figures, including chairmen, the discussion kept returning to the same question: when generative technology drives the production cost of text, code, and charts close to zero, what kinds of assets and services still have real defensive strength, or in his words, what is most “resistant to artificial intelligence.”
He points to the market for physical collectibles. In recent years, auction prices for physical Pokemon cards and classic player cards have continued to climb rather than being pushed aside by digital tools. Lin says the logic is straightforward. Anything that can be infinitely copied and quickly generated in a digital environment can see its marginal value fall quickly. By contrast, assets with physical scarcity, historical consensus, and strict offline certification by specialist institutions stand out more clearly.
The article names PSA and BGS as examples of grading bodies. Lin writes that an algorithm can create a huge number of polished images in a second, but it cannot produce, in the real world, a first-generation card printed in 1999, marked by age, and authenticated by a third party, such as an original Charizard or Cloyster card.
Three natural barriers in traditional professional markets
Lin extends that observation to professional services and enterprise markets. He says the strongest resistance to replacement also comes from dimensions that algorithms cannot casually substitute online.
Data privacy, compliance, and cybersecurity certification
The first barrier, he writes, is strict data privacy rules and international cybersecurity verification. Many seemingly capable general-purpose assistants on the market are, in his view, little more than an interface wrapped around a public model, without rigorous data retention policies.
In high-value professional markets, clients are often less focused on generation speed than on who will be accountable if something goes wrong. Lin argues that a compliance system able to sign agreements against data leakage and obtain International Organization for Standardization security certification operates at a threshold that common generic tools cannot match.
Offline trust networks and professional business development
The second barrier is offline trust and professional relationship-building. According to Lin, leading professional service firms and enterprise software companies often spend heavily on expert teams that can work close to the front line.
He says clients in specialist fields are usually experts themselves and tend to care about details. Standardized demos are often not enough. Durable trust built face to face, and grounded in a real understanding of pain points, is what gives a product or service stronger resistance to substitution.
Non-public experience held by experts
The third barrier is non-public practical knowledge accumulated by experts. Lin notes that statutes, case law, and public documents on the internet can all be read by models, but the details that matter most in practice are often not online.
He gives examples including a particular judge’s preferred format for organizing disputed issues, liability nuances in California reproductive medical contracts, and special inspection procedures tied to building overseas factories. Those kinds of tacit knowledge, he says, depend on long-term accumulation by seasoned professionals and cannot simply be captured by general external models.
How AI products should embed “AI-resistant” elements
Lin argues that if “AI-resistant” assets are gaining value, then the most damaging mistake for AI product architects and founders is to build a tool that is purely AI-based, easy to copy, and lacks any physical anchor.
In his view, AI products with long-term pricing power need to package those physical and institutional elements directly into their architecture.
Turn verification into a product feature
The first step is to productize grading and certification, which he describes as Verification as a Feature. Just as Pokemon cards rely on PSA slabs to prove authenticity, Lin says future AI outputs, whether code, legal summaries, medical advice, or financial models, will face a core problem: output may be abundant, but trust may be scarce.
He argues that strong AI products should not compete only on speed. They should include expert verification workflows, source-tracing labels, and mechanisms for deterministic checks. The provider that can offer a form of certainty comparable to a grading case for a collectible card will be in a better position to charge a premium.
Build private feedback loops around tacit knowledge
The second step is to build closed private feedback loops. Lin says that if a model is trained on public internet data, the resulting output has little real defensibility.
He argues that product design should make it easy for domain experts to digitize offline experience during their normal work. When top lawyers, senior engineers, or doctors revise, annotate, and correct content inside a system, that non-public practical judgment becomes a proprietary fine-tuning asset held by the product itself.
Lin compares those assets to older physical cards, saying they become scarcer over time while remaining inaccessible to outside general-purpose models.
Make accountability and compliance part of the core stack
The third step is to treat accountability and compliance architecture as a core layer rather than an add-on. Lin writes that general AI sells probability, while professional clients buy certainty and clear responsibility.
He says products should integrate high-grade data isolation from the start, including options such as on-premises private deployment and zero-data-retention architecture, along with international cybersecurity compliance standards and contract-level accountability commitments. Those elements, he argues, should be part of the delivery package from the beginning.
When clients see not just a chat interface but a complete defense system capable of passing strict legal and security review, switching costs rise sharply.
Keep the final mile human-led
The fourth step is Human-in-the-Loop Delivery. Lin says the most successful professional AI products are often not the ones chasing 100% automation. Instead, they position AI as a force multiplier for elite experts.
That means embedding AI deeply into a client’s existing real-world workflow through high-quality offline consulting and business networks. In his framework, algorithms handle 80% of repetitive structured work, while the remaining 20% of critical decisions and interpersonal communication stay with human experts. That creates a double barrier built from technology and human trust.
Lin’s conclusion: the real moat sits outside the algorithm
Lin concludes that the spread of generative AI amounts to inflation driven by digital marginal costs falling toward zero. In an era of effectively unlimited output, the things most likely to see value expand are those tied to physical scarcity, real trust, private-domain knowledge, and verifiable responsibility.
On the surface, future AI competition may look like a race over model parameters and inference speed. At a deeper level, Lin argues, the deciding factor will be who can embed the most “AI-resistant” elements into products and business models.