• Pascal's Chatbot Q&As
  • Posts
  • If OpenAI truly wishes to respect copyright and align with an opt-in world, Sora 2 must be rebuilt from the ground up.

If OpenAI truly wishes to respect copyright and align with an opt-in world, Sora 2 must be rebuilt from the ground up.

Otherwise, creators have every reason—and every legal justification—to seek compensation for the value their works continue to confer on the model.

Author

Pascal Hetzscholdt
October 04, 2025

If OpenAI truly wishes to respect copyright and align with an opt-in world, Sora 2 must be rebuilt from the ground up.

Otherwise, creators have every reason—and every legal justification—to seek compensation for the value their works continue to confer on the model.

OCT 04, 2025

Why Sora 2 Must Be Retrained: The Legal and Technical Case Against Opt-In Without a Reset

by ChatGPT-4o

When Sam Altman announced that OpenAI would pivot from an opt-out to an opt-inregime for copyright holders in connection with Sora 2, it sounded like a victory for creators. At first glance, the move appeared to hand control back to rights owners, aligning Sora 2 with more conventional licensing norms. But beneath the polished PR message lies a hard truth: Sora 2 cannot simply switch contractual gears without a fundamental retraining of the model. Legally and technically, the opt-in promise collapses if the model continues to rely on data acquired under an opt-out framework.

An opt-out system assumes permission until explicitly revoked. An opt-in system assumes no permission until it is expressly granted. These are opposites, not variants.

If Sora 2 was trained on billions of copyrighted works under the old opt-out assumption, those works are now embedded in the model’s statistical weights. Flipping the consent rule post hoc does not erase the fact that training occurred without authorization. Continuing to monetize such a model is not neutral—it is ongoing exploitation of unlicensed works.

Courts have already signaled hostility toward the idea that training is a free-for-all. From Getty v. Stability AI to the still-pending music industry suits against Anthropic and Suno, plaintiffs consistently argue that the act of training itself constitutes unauthorized copying. Switching to opt-in for future data cannot cleanse a model whose current form remains the product of unlawful ingestion.

The analogy is simple: you cannot build a skyscraper on stolen land and later claim legitimacy by saying you’ll only “rent to tenants with consent.” The foundation remains contaminated.

2. The Technical Reality: Guardrails Do Not Rewrite Weights

Some argue that OpenAI could solve this problem with “guardrails”—filters that prevent Sora 2 from spitting out Harry Potter clips or Studio Ghibli pastiches on command. This is a red herring.

Guardrails operate at the output stage, suppressing certain generations. They do not reach back into the model’s neural weights, which are the compressed statistical distillation of the training set. Whether or not Sora 2 visibly reproduces copyrighted characters, its ability to conjure rich cinematic styles, character archetypes, or narrative pacing is powered by patterns extracted from copyrighted material.

The law cares not only about literal duplication but also about derivative benefit. Even if OpenAI avoids verbatim reproduction, the embedded knowledge remains a form of appropriation. Put bluntly: filters can hide the evidence, but they cannot erase the crime scene.

3. If Retraining Does Not Occur: The Compensation Question

Suppose OpenAI keeps Sora 2 intact and simply announces “from now on, we’ll respect opt-in.” In that scenario, rights holders are justified in demanding compensation.

  • Unjust enrichment: OpenAI gained immense commercial value from works it never had permission to use. That enrichment continues to accrue every time a user generates content.

  • Derivative influence: Even absent verbatim copies, copyrighted material shapes Sora 2’s stylistic and technical capacities, which is precisely what gives the model its market edge.

  • Ongoing commercial exploitation: Each subscription or licensing deal involving Sora 2 leverages outputs powered by unauthorized training data.

In this light, rights holders can persuasively argue: “Unless Sora 2 is retrained entirely from scratch using datasets sourced under true opt-in consent, we are entitled to direct compensation for how our works continue to increase the model’s market value.”

4. The Broader Implication

The Sora 2 case illustrates a larger principle for the AI era: policy pivots must match technical and legal realities. It is not enough to declare a shift from opt-out to opt-in; one must rebuild the underlying system on that basis. Anything less is a sleight of hand.

For creators, this moment clarifies strategy. Guardrails are not enough. Promises of future compliance are not enough. The demand must be: Retrain, or pay.

Conclusion

Opt-in without retraining is a contradiction. It leaves rights owners in the same compromised position while offering the illusion of empowerment. If OpenAI truly wishes to respect copyright and align with an opt-in world, Sora 2 must be rebuilt from the ground up. Otherwise, creators have every reason—and every legal justification—to seek compensation for the value their works continue to confer on the model.

Retrain, or pay. Those are the only two honest options.

·

1 OCT

Why this “opt-out first, ask permission later” gambit is arguably the dumbest possible move

·

1 OCT

The Illusion of Control: Why “Opt-Out” and “Pay-Per-Output” Fail in the Era of Generative AI