Investigating AI’s Decision-Making: Insights from Anthropic

On a recent episode of the Redemptive Marketing Podcast, we unpacked Anthropic’s Tracing the Thoughts of a Large Language Model, a research paper released in April 2025. Instead of treating large language models (LLMs) simply as unfathomable “black boxes,” the research begins to map the steps by which LLMs generate content. Rather than relying on opaque output, the authors compared LLM logic to a kind of human-like trial and error, assigning different weights to possible continuations during text generation.

As Parker discussed, the research highlighted how, when prompted to write a poem, the model contemplates branching sets of possibilities—such as whether to rhyme “carrot” with “rabbit” or “habit.” This provides a limited but intriguing glimpse into decision making inside the AI. While not offering a complete solution to questions of transparency, it represents a first, incremental step toward making these processes more observable, a theme that resonates with Snapmarket’s commitment to process clarity and thoughtful craftsmanship.

We recognized, however, that interpretability is only part of the picture. In practice, questions of copyright, attribution, and provenance continue to complicate the work of digital creators and marketers. The challenge is not just technical, but foundational to authorship in the modern age.

Web3 Publishing: Traceable, Attributable Ownership

Shifting from AI to Web3, the conversation drew on Parker’s experience in blockchain-enabled publishing. Web3 represents a notable change in how digital works are created, stored, and attributed. Rather than limiting publishing to simple access or editing rights, Web3 enables traceable, process-based ownership—content can be assigned to a unique identity, permanently timestamped, and theoretically linked to its creator on an immutable ledger.

We referenced platforms like Mirror.xyz, which allow writers to publish entire posts directly on chain. This approach both documents and creates a non-fungible record or token (NFT) of when and by whom something was published. As explored in the podcast, this opens a possibility: if authorship can be proven and linked to a known on-chain published work, establishing provenance for AI-generated digital works could become less ambiguous and more systematic.

Moreover, this model aligns with the broader vision of moving from a transactional to a transformational digital landscape, where content isn't just shared, but owned, with clear lines of origin and value distributed among contributors.

It’s important, as we emphasized, to recognize that these are theoretical possibilities, not guarantees. As Parker speculated, if future improvements in AI allow us to trace a model’s reasoning back to its original sources—and those sources are published on chain—we could see the emergence of systems where creators are automatically compensated for the use of their work, potentially through micro-payments or royalties. That scenario remains some distance from reality—but it provides a framework for what process-based, automated attribution might look like.

This speculative future raises questions: How might this change the way content is created and shared? Could this lead to a more equitable digital economy? The answers are unclear, yet they invite creators to think critically about future-proofing their work in an increasingly digital realm.

Provenance and Origin in the Age of AI

We spent time on the podcast discussing what’s at stake for writers, publishers, and brands. A primary concern is that AI, by drawing on extensive datasets, often reduces the need for audiences to visit or engage with original sources. This dynamic threatens well-established revenue streams for publishers—whether they’re major news outlets or independent writers.

Direct publishing platforms, such as paid newsletters, have offered some relief, helping creators maintain closer ties to their audience. Yet even with such systems, content can make its way into larger AI training datasets, raising the question: can original authors prove their claim if their intellectual property resurfaces elsewhere?

Publishing work on chain, as envisioned in our discussion, may offer a theoretical way to create a permanent, timestamped record—potentially enabling someone to demonstrate that they were the first to introduce an idea, phrase, or format. As both Parker and Brandon observed, this method is not a failsafe. Effective authorship verification would require alignment and transparency across technical, legal, and platform boundaries.

Considering the implications of this potential shift, content creators might rethink their strategies, exploring ways to integrate blockchain-based strategies into their daily operations. Could this become a standard for all creative works? How might this influence the relationship between creators and consumers?

We also noted the increasing sophistication of generative tools—capable of producing deepfakes, imitation books in the style of canonical writers like Mark Twain, or visual works channeling established brands such as Studio Ghibli. Here, the potential for on-chain provenance may serve as a process-oriented check, rather than a sure protection.

And beyond individual creators, there's a broader conversation to be had about the impact on collaborative works, where multiple creators contribute to a single piece. The ability to track contributions could redefine collaboration, ensuring each contributor receives fair acknowledgment and compensation.

AI, Web3, and the Realities of Resource Use

Toward the episode’s conclusion, we addressed a key limitation: powering AI and maintaining blockchain ledgers comes with significant energy demands. Even simple queries to advanced AI models, or maintaining blockchain operations, can draw as much power as a home, impacting environmental resources. We noted this with some humor, referencing reports about the unexpected costs of being polite to ChatGPT, yet regarded the broader point as a genuine concern.

This evaluation raises strategic opportunities. How can businesses leverage AI and blockchain responsibly? What are the implications for future infrastructure, and how can organizations design systems that consider long-term environmental stewardship in concert with business goals?

What's next?

In our discussion, we underscored that these ideas—AI interpretability, blockchain authorship records, and systematic attribution—are still developing. Most of the benefits remain speculative, dependent on future advances and the collective adoption of new standards. As we look to the future, we’re focused on building systems for authenticity and stewardship in digital authorship. This ongoing journey reflects our core belief: that thoughtful human-centric innovation offers the best path to redemptive marketing.