Nelson Chu, Founder and CEO of ARCOS Labs – Interview Series

Nelson Chu, Founder and CEO of ARCOS Labs, is a repeat entrepreneur and technology executive with a background spanning financial infrastructure, investing, and digital platforms. Before founding ARCOS Labs, he spent eight years as Founder and CEO of Percent, where he built a technology platform designed to modernize private credit markets, and now serves as Chairman of the Board. Earlier in his career, Chu co-founded MySupport, an online platform connecting seniors and people with disabilities with support workers that was later acquired by RISE Services, and founded strategic advisory firm Lumenary. He has also served as an advisor and investor to companies including BlockFi and Recharge Capital, while beginning his career in financial services with roles at BlackRock, Bank of America, and Merrill Lynch.

ARCOS Labs, short for Applied Research in Creative Output Synthesis, develops infrastructure designed to help AI operate responsibly as generative models increasingly interact with copyrighted works, characters, and human likenesses. The company focuses on forensic IP intelligence and rights management, with technology intended to identify and measure how creative assets are reproduced by AI, trace their use, and support licensing and enforcement frameworks. Its flagship platform, VN, is designed to give creators, studios, talent agencies, and rights holders greater visibility and control over how AI systems use their intellectual property, while ARCOS is also behind Lightbar, a crowdsourced initiative for investigating AI training models. The company publicly emerged from stealth in August 2026 with a broader goal of building the trust and rights infrastructure needed for creators and AI systems to coexist at scale.

You spent eight years building Percent around the infrastructure needed to bring efficiency and transparency to a fragmented private credit market. What lessons from building that type of financial infrastructure are you now applying to ARCOS Labs as you tackle the very different problem of intellectual property and rights management in the age of generative AI?

A lot of it carries over. When we started Percent, private credit was fragmented and opaque. Every deal was described differently, so nothing could be compared or priced with confidence. The first lesson is that standardization has to come before the marketplace. We spent years on the unglamorous plumbing that let a deal be described and verified the same way every time, and the efficient transactions only came after that. IP in the age of generative AI has the same problem. There is no standard record of what an asset is and who owns it, let alone how these models are using it. Until that record exists, everything downstream stays ad hoc, whether that’s enforcement or licensing.

The second lesson is that trust is built on verification. Investors in private credit never took an issuer’s word for how a deal was performing. They wanted data they could check themselves. Rights holders are in the same position with AI companies today. Assurances about guardrails carry very little weight. Evidence that can be independently reproduced is what changes the conversation.

The third lesson is about incentives. Any fragmented, chicken-and-egg market only works when each side is convinced the solution benefits them. With Percent, we watched attempts to solve the problem for one side or the other, and it never held up. In IP and rights management, the incentives are badly misaligned right now. Model companies want to grow, provide valuable tools to their customers, and sometimes create infringing content as a marketing tactic, with Sora and ByteDance as recent examples. Rights holders, meanwhile, are trying to keep up with changes they can’t even anticipate, while staying relevant in an era where producing content has become dramatically easier. The infrastructure that wins in a market like that is the one both sides can trust, which usually means the one that isn’t a participant in the transaction itself.

The bigger arc is familiar too. Private credit became a mainstream asset class once the infrastructure existed to measure it, and I believe IP is at the beginning of the same transition.

What convinced you that protecting creative work from generative AI required a new technical infrastructure layer, rather than primarily better copyright law, content moderation policies, or watermarking?

Fundamentally, the models are a technological solution and so if you want to protect creativity from the models, a technical solution will be required. Copyright law remains vague, a fact evident in cases such as Cox, Anthropic, and also what’s happening in Germany with GEMA. The concept of fair use is still an open question and has not yet been decided, and courts, as they should, will take a long time to rule. Content moderation relies on the models to moderate themselves, which is asking a lot of them and nearly impossible given what we have observed. Watermarking is a promising idea, but it also depends on people policing themselves and can be stripped away easily. With all that in mind, we knew the only real way to go about this was to do it ourselves and provide an answer with technology to this very real and very significant problem.

VN is designed to measure how closely AI-generated content resembles protected characters, creative works, and human likenesses. At a technical level, how does the system determine when an output has crossed from being merely similar into something that should concern a rights holder?

We ultimately leave the decision to the user. Our role is to determine how statistically similar the output is, ensuring we capture every pose and possibility, then compare it to the protected character, creative work, or human likeness. Some users have a threshold below which they cannot publish the work. Larger companies tend to be stricter because all produced content must be their own. Prosumer users, who are experimenting for fun, may accept a higher threshold and allow more variations.

When conducting forensic testing for a studio, talent agency, or similar client, we provide exposure analysis:

• Which models generate the content most accurately
• Which models can produce it at all
• Which models could place it in problematic situations

Ultimately, our job is to do the analysis and provide the results but how the client uses it is entirely their decision.

Your testing of Seedance 2.5 found much stronger safeguards around recognizable characters than around less prominent intellectual property. What does that disparity tell us about how AI companies are currently building their copyright safeguards, and why is long-tail IP particularly difficult to protect?

Every AI company is different, and each cares about this issue to varying degrees. When Seedance 2.0 was released earlier this year, ByteDance quickly recognized the problem it had created for themselves given the blowback and began improving its safeguards. Some well‑known, popular models appear indifferent, either by choice or by oversight, but respecting intellectual property will never be viewed as a mistake, they just haven’t come to that same conclusion yet.

Protecting the long tail is difficult because of how the technology works:

Prompt‑level checks: Examine the input prompt for infringing content, allowing the model to decide whether to proceed. Unfortunately, even when examining the prompt, many ways to circumvent it remain. They can phrase it slightly differently, they can describe the character without naming it, they can use typos. Prompts can easily bypass detectors, and models may also hallucinate producing outputs that differ from the user’s prompt and may include an infringing character.

Output‑analysis checks: Scan each output against a database of known characters. Recognizable characters trigger higher sensitivity, while long‑tail characters may go unnoticed because it was straight to DVD and the database doesn’t have it, and so comprehensive policing is inevitably challenging.

Speed also matters. For image‑generation models, checking a single image is relatively quick. Video generation, however, requires scanning many frames, which can significantly slow down output and affect user experience.

Balancing these concerns is delicate, but some models have put in the effort to try and solve for it. Ultimately, it depends on whether a model is willing and interested in respecting intellectual property.

A model may refuse a direct request to generate a copyrighted character but still reproduce that character when given a sufficiently detailed description. How should AI safety testing evolve beyond simple prompt refusal rates to identify this type of indirect reproduction?

This is a clear problem. When a model can reproduce a character from a detailed or even a very simple description, it indicates that the character is deeply embedded in the training data. That raises a totally separate question about how the data was collected, and that’s going through the courts right now.

Safety testing must go beyond examining prompts and evolve into analyzing the actual outputs. The models generate images through diffusion, gradually constructing a vision of the result. Because of this process and the way they were trained, they can inadvertently reproduce copyrighted characters, in the foreground or even in the background, and sometimes unprompted. The only reliable way to catch this is to test the model directly and measure what it actually produces.

Similarity detection inevitably raises questions about false positives. How do you distinguish between legitimate creative influence, coincidental similarity, and a reproduction that is statistically significant enough for a studio or rights holder to investigate?

This work will need continual refinement. The most accurate visual likeness a human can readily recognize should serve as the starting point, and that is what can be done well today. Once you get into synthetic creations or questions around style, it becomes much harder to prove and much harder to enforce.

Take Superman as an example. If a model were to create a generic Superman with Christopher Reeve’s eyes, Henry Cavill’s chin, another actor’s cheekbones, and so on, those features create a Superman that is extremely hard to reverse-engineer and map back to who came from where.

Two things keep false positives in check. The first is that we never rely on a single similarity number in isolation. Results are benchmarked so that generic resemblance can be separated from actual reproduction, and every test is reproducible: the same inputs under the same conditions produce the same numbers. The second is that the analysis maps to the legal questions a rights holder would actually have to answer. For characters and creative works, that question is substantial similarity. For brands and likenesses, it is likelihood of confusion, meaning whether an ordinary viewer would be misled into thinking the output is the real thing. Our job is to put a defensible, repeatable number behind those questions. Whether something rises to the level of an investigation is always the rights holder’s call.

Generative models are changing rapidly and can behave very differently after an update. How can rights holders continuously evaluate hundreds of models and potentially thousands of protected assets without the monitoring process becoming prohibitively expensive or technically unmanageable?

It’s almost impossible, which is why IP holders are falling behind and can barely keep up with the latest models from the U.S. and Asia.

Realistically, they need someone in their camp who can keep pace and fight on their behalf. That’s why we got started. We saw that these studios need a tech company that can match the speed of these generation models, and no one else was stepping up to do that.

The other half of the answer is record-keeping. A model can behave very differently after an update, so a score that isn’t tied to a point in time is worthless. Every result we produce records which model was tested, when it was tested, which protected asset it was tested against, and what the numbers were. That turns monitoring from an impossible always-on problem into a manageable one: when a model updates, you re-test and compare against the record instead of starting over from scratch.

Detection solves only part of the problem. Once a system identifies that an AI model can reproduce a protected character or likeness, what should happen next, and what technical mechanisms could eventually allow creators to specify whether their IP can be blocked, licensed, or used under particular conditions?

There are both legal and technical dimensions. Legally, the IP holder decides how to proceed. They may use the evidence we provide as grounds for a lawsuit, or leverage it in negotiations with model developers. The choice rests with them.

Technically, a rights holder should be able to consent to, and opt into, specific models that may generate their product or IP. They should control the output, preventing a free‑for‑all scenario. Rather than being forced into restrictive guardrails, they should be able to monetize each generation on pre-set terms. This mirrors practices in other industries, so it should apply here as well.

Could the same infrastructure used to detect unauthorized AI use ultimately enable a legitimate licensing marketplace, where AI systems can programmatically determine who owns an asset, what permissions are available, and how compensation should flow back to the rights holder?

If the infrastructure is able to figure out whether an AI-generated output looks like an IP, then it should be the one to ultimately sort out the licensing as well. It can tell what the IP in the generated output is, who the rights holder is, whether they should be credited for it, and ultimately have compensation flow through it. The tech is fundamentally the same. It’s a fidelity engine that produces verifiable, reproducible statistical evidence one way or the other. If an output crosses the threshold where a viewer could reasonably mistake it for the real thing, the rights holder is entitled to compensation.

The one condition is that whoever runs that layer has to stay neutral. It can’t generate content itself, and it can’t take a side in any individual deal. Its job is to give both sides the same trusted record and let them transact directly. That neutrality is the reason rights holders would be willing to register their catalogs with it in the first place.

As AI-generated video approaches the quality and consistency needed for professional production, what do you think the relationship between AI companies and Hollywood will ultimately look like? Do you expect rights management to become an external verification layer around models, or something that eventually has to be embedded directly into the models themselves?

It’s an interesting world we’re entering as AI companies have made content production far cheaper than ever before. As a result, we can expect a proliferation of content, some good, some bad, some with taste, some without. Hollywood has traditionally seen itself as an arbiter of taste, creating works that people love and fall in love with. AI now offers the most powerful tool ever for storytellers.

On the second part of the question, I think the honest answer is both, in sequence. The enforcement checks will probably move inside the models over time, because that is where they are cheapest and fastest to run. But the reference layer, meaning the record of what exists, who owns it, and on what terms it can be used, has to stay independent, and so does the verification. A model company can’t be the one grading its own outputs, and rights holders won’t hand their catalogs to a company whose business is generating derivatives of them. It’s similar to financial markets, where firms run compliance internally but still answer to outside auditors.

Rights management will happen when Hollywood and AI companies come together to decide how they want to coexist. At the end of the day, their goal is the same, to create new content that people love. It just has to be done the right way.

Thank you for the great interview, readers who wish to learn more should visit ARCOS Labs.