I spent the better part of last week staring at a spreadsheet that shouldn't have made me angry. It was a simple data table โ 2,034 books, a percentage column, a category column. Religious texts, mostly. The number that kept pulling my eyes back was 63%. That's the share of recently published religious books on Amazon that Originality.ai's detection tools flagged as AI-generated. Nearly two-thirds. And of the verifiable factual claims in those books, roughly 53% contained errors.
I've been in this industry long enough to know when a number is doing more work than the methodology behind it. But I've also been in this industry long enough to recognize a systemic failure when I see one. The publishing industry is experiencing what DeFi went through in 2020 โ a flood of cheap, automated capital (in this case, content) overwhelming a system built on trust signals that no longer function. And the tools we're reaching for to solve it are the same tools we keep reaching for in crypto: centralized oracles that claim to verify truth but are themselves the weakest link in the chain.
Truth is immutable, unlike the price action. But in the publishing world, truth has become a probabilistic output of a statistical model that nobody has independently audited.
The Context: How We Got Here
Let me step back and give you the full picture, because the numbers only make sense when you understand the machinery behind them.
Amazon's Kindle Direct Publishing (KDP) platform has, since its inception, been the great democratizer of the written word. Anyone with a manuscript and a few dollars can publish a book and reach millions of readers. The barrier to entry was always the writing itself โ the months of labor, the craft, the editorial rigor. That barrier has now been effectively eliminated.
With tools like ChatGPT and Claude, generating a 200-page religious book costs approximately zero dollars in marginal terms. The generation is instant. The editing is optional. The publishing is a few clicks. And the economics are devastatingly simple: religious books have stable, predictable demand. People search for Bible studies, prayer guides, Wiccan rituals, and theological commentaries with consistent volume year after year. It's a long-tail market with guaranteed traffic.
So what happened? The same thing that happens in every market when the cost of production drops to zero: production explodes, quality collapses, and the platforms that take a cut of every transaction look the other way.
Amazon takes 30% to 70% of every KDP sale. AI-generated books are pure margin for the platform. The conflict of interest is so obvious that it's almost invisible โ the way oxygen is invisible. We don't see it because it's the air we breathe.
Originality.ai's study, released on August 24th, examined 2,034 recently published religious books on Amazon's platform. Their detection tools flagged 63% as AI-generated. The category breakdown is even more telling: witchcraft and occult books led the pack at 78% AI-generated, followed by other religious subgenres at similarly alarming rates.
But here's where I start to get uncomfortable with the narrative. Originality.ai is not a neutral research institution. It's a commercial AI detection tool provider. Its business model depends on the perception that AI-generated content is a pervasive threat. The more alarming the numbers, the more valuable the detection service. This is what we in crypto would call a classic oracle problem โ the entity providing the truth signal has a direct financial interest in the signal reading a particular way.
The Core: Why Detection Tools Are the Wrong Answer
Let me get technical for a moment, because this is where the analogy to blockchain infrastructure becomes unavoidable.
AI detection tools like Originality.ai, GPTZero, and Turnitin's AI detection feature all operate on a similar fundamental principle. They analyze statistical features of text โ perplexity (how surprised a language model is by the text), burstiness (the variation in sentence length and complexity), and other distributional properties โ to estimate the probability that a given text was machine-generated.
This is not deterministic verification. It is probabilistic inference. The output is not "this text is AI-generated" but rather "this text has features statistically consistent with AI generation." The distinction matters enormously, and it's the same distinction between a centralized oracle and a cryptographic proof.
A cryptographic proof is verifiable by anyone, anywhere, with mathematical certainty. A statistical inference is a guess with a confidence interval. And in adversarial settings โ which is exactly what we're dealing with here โ confidence intervals collapse.
Based on my audit experience in 2017, when I spent six months examining the Solidity code of the Tezos mainnet launch and identified 14 critical vulnerabilities in the consensus mechanism's implementation, I learned something that has stayed with me: any system that relies on a trusted third party to verify truth is a system that will eventually be gamed. The question is not whether it will be gamed, but when, and at what scale.
AI detection tools are being gamed right now. The academic literature on this is clear. Tools like GPTZero and DetectGPT achieve 70-90% accuracy under ideal conditions โ clean, unedited AI text. But under adversarial conditions โ text that has been rewritten, translated, mixed with human writing, or edited by a human โ accuracy drops below 50%. That's worse than a coin flip.
And here's the dirty secret that Originality.ai's study doesn't mention: false positive rates. The rate at which human-written text is flagged as AI-generated. For religious texts specifically, this is a critical blind spot. Religious writing is full of ritualistic language, repetitive prayer structures, formulaic expressions, and stylized patterns that are centuries old. These are precisely the features that AI detection tools are trained to identify as machine-generated. A detection tool might flag a traditional Catholic prayer book as AI-generated because the language is too regular, too patterned, too predictable.
The study reports that 53% of verifiable factual claims in the flagged books contain errors. But who verified those claims? What methodology was used? Were the verifiers experts in the relevant religious traditions? A claim about a historical event in a religious text might be "factually wrong" by one standard and "theologically interpretive" by another. The line between error and interpretation in religious texts is notoriously blurry.
I'm not saying the study is wrong. I'm saying the study is insufficiently rigorous to support the certainty with which it's being reported. And that's the pattern we keep seeing in this space โ a commercial entity publishing alarming statistics that happen to support its business model, with methodology that doesn't withstand independent scrutiny.
This is the same problem we face in DeFi with oracle feeds. Chainlink, the dominant oracle provider, is supposed to solve the problem of getting real-world data onto the blockchain. But the solution itself is centralized โ a network of nodes that, while distributed, is still a trusted intermediary. The entire premise of decentralization is undermined when the truth source is a centralized entity. The same logic applies to AI detection. We're building centralized detectors to solve a problem created by decentralized generation, and the detectors themselves are the new single point of failure.
The Economics of the Flood
Let me walk you through the unit economics, because this is where the story becomes truly uncomfortable.
A traditional religious book goes through a publishing pipeline that includes editing, proofreading, design, marketing, and distribution. The cost structure is real: editors need salaries, designers need fees, marketers need budgets. A traditional religious publisher might spend $20,000 to $50,000 bringing a single book to market.
An AI-generated religious book costs nothing to produce. The text is generated by a language model. The cover is generated by an image model. The formatting is automated. The only cost is the KDP platform fee, which is a percentage of sales. Even at a low price point of $2.99 to $9.99, the profit margin is nearly 100%.
This is not a sustainable competitive dynamic. It's a race to the bottom where the bottom is zero. And the bottom is winning.
The study found that 63% of recent religious books on Amazon are AI-generated. That means traditional authors and publishers are competing against an army of automated content factories that can produce books in minutes, price them at the floor, and flood the market with volume. The signal-to-noise ratio in the religious book category has collapsed.
And here's the part that keeps me up at night: the readers can't tell the difference. Or rather, they can't tell the difference until they encounter the errors. And by then, they've already paid their money and absorbed the misinformation.
Fifty-three percent of verifiable factual claims in these books contain errors. Let me put that in perspective. If you buy a religious book on Amazon today, you have better than even odds of encountering a factual error. For a category where readers are seeking spiritual guidance, historical accuracy, and ritual instruction, this is not a minor inconvenience. It's a spiritual hazard.
Consider the witchcraft and occult category, where 78% of books are AI-generated. These books often contain instructions for rituals, spellwork, and spiritual practices. If those instructions are wrong โ and with a 53% error rate, many of them are โ readers are not just getting bad information. They're getting potentially harmful guidance presented with the authority of a published book.
This is the same dynamic we saw in DeFi in 2020, when unaudited smart contracts were draining users' funds at an alarming rate. The technology had democratized access to financial services, but it had also democratized access to financial destruction. The difference is that in DeFi, at least the code was visible. You could audit the smart contract if you had the skills. In publishing, the reader has no way to audit the content before purchasing it.
The Platform's Complicity
Let me talk about Amazon's role in this, because it's the elephant in the room that nobody wants to address.
Amazon's KDP platform updated its AI content disclosure policy in 2023, requiring authors to declare whether their content is AI-generated. The policy exists on paper. In practice, enforcement is minimal to nonexistent. The platform has no meaningful mechanism for verifying AI content declarations, and there's no penalty structure that would deter non-compliance.
Why? Because Amazon makes money on every sale, regardless of whether the content is AI-generated or human-written. The platform takes a 30% to 70% cut of every transaction. AI-generated books are a revenue stream, and aggressive enforcement of AI disclosure policies would reduce that revenue stream.
This is not a conspiracy theory. It's basic incentive analysis. Amazon is a publicly traded company with a fiduciary duty to maximize shareholder value. AI-generated books are profitable. The platform has no financial incentive to crack down on them, and every incentive to maintain the status quo.
The same dynamic exists in the crypto ecosystem. Exchanges make money on trading volume, regardless of whether the tokens being traded are legitimate or scams. The incentives are misaligned, and the result is that bad actors thrive while the platform looks the other way.
I've seen this pattern before. In 2022, when the Terra-Luna collapse shattered my idealization of algorithmic stability, I retreated to a cabin in rural Virginia for six weeks, disconnecting from all digital devices. During that solitude, I drafted the manuscript for "The Soul of Sovereignty," a book arguing that blockchain must serve human dignity, not just capital efficiency. The lesson I took from that experience was simple: when the incentives are wrong, the technology doesn't matter. The system will find a way to fail.
Amazon's incentives are wrong. The platform is complicit in the AI content flood, not because of malice, but because of economics. And no amount of detection tooling will fix a problem that the platform has no incentive to solve.
The False Promise of Detection
Let me now address the detection tool market directly, because it's where the crypto analogy becomes most uncomfortable.
The AI detection market is currently a fragmented landscape of commercial tools (Originality.ai, GPTZero, Copyleaks, Turnitin), open-source tools (DetectGPT, Radar AI), and platform-native solutions (Google's SynthID, OpenAI's content classifiers). No single player has achieved market dominance, and the technical reliability of all of them is contested.
Originality.ai's strategy is to establish thought leadership through industry research. The religious books study is a classic example of research-driven marketing โ publish alarming statistics, get media coverage, drive demand for the detection product. It's a smart B2B marketing strategy, but it's also a conflict of interest that undermines the credibility of the research.
Here's the fundamental problem: AI detection is an arms race, and the detection tools are always one step behind the generation models. Every time a detection tool identifies a statistical pattern characteristic of AI generation, the next generation of language models is trained to avoid that pattern. The cat-and-mouse game is perpetual, and the detection tools are structurally disadvantaged because they're reactive.
This is the same problem we face with smart contract auditing. Auditors find vulnerabilities, developers fix them, and new vulnerabilities emerge. The audit is a point-in-time assessment, not a guarantee of future security. The same is true for AI detection โ a detection result is a point-in-time probability estimate, not a definitive determination.
And then there's the false positive problem, which the industry has been remarkably reluctant to address. In 2023, Turnitin's AI detection feature was involved in multiple controversies when it flagged student essays as AI-generated, leading to false accusations of academic dishonesty. The false positive rate for AI detection tools is not publicly disclosed by most vendors, but independent evaluations suggest it's significant โ potentially 1-5% for clean text, and much higher for text with unusual stylistic features.
For religious texts, the false positive problem is amplified. The ritualistic language, repetitive structures, and formulaic expressions common in religious writing are precisely the features that AI detection tools are trained to identify as machine-generated. A detection tool might flag a legitimate human-written prayer book as AI-generated because the language is too regular, too patterned, too predictable.
This creates a perverse outcome: the tools designed to protect human authors from AI competition might actually be harming human authors by falsely flagging their work. The cure is becoming part of the disease.
The Blockchain Alternative
So what's the actual solution? If detection tools are unreliable, and platforms have no incentive to enforce disclosure policies, what can we do?
The answer, I believe, lies in the same technology that underpins the crypto ecosystem: cryptographic provenance.
Instead of trying to detect AI-generated content after the fact, we should be building systems that cryptographically verify content provenance at the point of creation. This is not a new idea โ the C2PA (Coalition for Content Provenance and Authenticity) standard has been developing exactly this kind of infrastructure, and companies like Adobe and Microsoft have been integrating provenance metadata into their content creation tools.
The concept is simple: when a human writes a book, they cryptographically sign the manuscript with their private key. The signature is recorded on a blockchain, creating an immutable timestamp and proof of authorship. When a reader encounters a book, they can verify the cryptographic signature and confirm that it was created by a human, at a specific time, with a verifiable chain of custody.
This is not detection. It's provenance. It's not probabilistic inference. It's deterministic verification. It's the difference between a centralized oracle and a cryptographic proof.
I've been thinking about this problem since 2025, when I launched my "Human-Centric AI" initiative and collaborated with three key ethicists to draft the "Decentralized Trust Protocol," a set of guidelines for ensuring AI agents respect user sovereignty. The core insight from that work was simple: you can't detect authenticity after the fact. You have to establish it at the point of creation.
The same principle applies to publishing. We can't build reliable detectors for AI-generated content because the detection problem is fundamentally adversarial โ every detection method can be evaded by a sufficiently sophisticated generation model. But we can build provenance systems that make the origin of content cryptographically verifiable.
If a book is published with a cryptographic signature from a human author, readers can verify that it was human-written. If a book is published without such a signature, readers can reasonably assume it may be AI-generated. The system doesn't need to detect anything โ it just needs to make the provenance transparent.
This is the same logic that underpins Bitcoin. We don't need to detect counterfeit money because the blockchain makes the history of every bitcoin verifiable. The system doesn't rely on detection โ it relies on cryptographic proof.
The Contrarian Angle: Why Provenance Won't Work Either
Now let me play devil's advocate with myself, because I've been in this industry long enough to know that every elegant solution has a fatal flaw.
The provenance solution has a fundamental problem: incentives. Who benefits from cryptographic provenance?
Traditional authors benefit โ they can differentiate themselves from AI-generated content and build trust with readers. But AI-generated content producers don't benefit, and they're the ones flooding the market. Amazon doesn't benefit โ the platform makes money on volume, not on provenance verification. And readers, the people who would benefit most, don't have the technical literacy to verify cryptographic signatures.
The adoption problem is the same one we face in crypto. The technology works, but the incentives aren't aligned for mass adoption. We've been building decentralized infrastructure for over a decade, and the vast majority of the world still doesn't use it. The same will be true for content provenance.
And there's a deeper problem: the stigma problem. If we create a "human-authored" certification, it implicitly creates a "possibly AI-generated" category. Books without certification become suspect. This could harm legitimate authors who, for whatever reason, don't use the certification system โ authors who are less technically savvy, authors in developing countries, authors who publish through traditional publishers that haven't adopted the standard.
The certification system could create a two-tier publishing market, where certified human authors are privileged and everyone else is marginalized. This is the same dynamic we see in crypto, where "verified" accounts on exchanges get preferential treatment, and unverified users are locked out of basic services.
And then there's the question of whether provenance actually solves the problem. Even if we can verify that a book was written by a human, that doesn't mean the book is accurate. Humans write books with factual errors all the time. The 53% error rate in the study might be partially attributable to AI, but it's not exclusively an AI problem. Human authors make mistakes too.
Provenance solves the authenticity problem, but it doesn't solve the quality problem. And the quality problem is the one that actually matters for readers.
The Real Problem: Economic Incentives
Let me step back and think about this from a systems perspective, because I think we're all looking at the wrong layer of the problem.
The AI content flood in publishing is not a technology problem. It's an economics problem. The cost of content production has dropped to zero, and the market is responding exactly as economic theory predicts: production has exploded, quality has collapsed, and the platforms that benefit from volume are not intervening.
This is the same pattern we see in every industry when production costs drop to zero. When the cost of creating a website dropped to zero, we got a flood of spam websites. When the cost of creating a video dropped to zero, we got a flood of low-quality YouTube content. When the cost of creating a financial product dropped to zero, we got a flood of scam tokens.
The solution is not better detection. The solution is better incentives. We need to create economic structures that reward quality over volume, that reward human authorship over automated generation, and that align the platform's incentives with the readers' interests.
This is where blockchain actually has something to offer. Not as a detection tool, but as an incentive alignment mechanism.
Imagine a publishing platform built on blockchain technology, where content creators stake tokens to publish, and their stake is slashed if their content is found to be low quality. Imagine a reputation system where authors build verifiable track records of accuracy and quality, and readers can see those track records before purchasing. Imagine a governance system where the community, not a centralized platform, decides what content gets promoted and what gets buried.
This is the promise of decentralized publishing. It's not about detecting AI content. It's about creating economic structures that make low-quality content unprofitable.
I've been thinking about this since 2020, when I founded OpenLedger Lab, a non-profit educational initiative. I personally mentored 50 junior developers from underrepresented backgrounds, helping them deploy their first ERC-20 tokens. The experience taught me something important: technology is easy. Incentives are hard. You can build the most elegant technical system in the world, and it will fail if the incentives aren't aligned.
The same is true for publishing. We can build the most elegant detection tools, the most sophisticated provenance systems, and they will fail if the economic incentives don't support them.
The Cultural Dimension
Let me now address something that the study touches on but doesn't fully explore: the cultural dimension of AI-generated religious content.
The study found that 78% of books in the witchcraft and occult category are AI-generated. This is not just a quality problem. It's a cultural appropriation problem.
Witchcraft and occult traditions are not just collections of facts and instructions. They are living traditions with deep cultural roots, specific lineages, and contextual knowledge that cannot be captured by a language model. When an AI generates a book about Wiccan rituals, it's not just producing potentially inaccurate information โ it's appropriating and flattening a cultural tradition that has been developed over centuries by specific communities.
The same is true for other religious traditions. A language model trained on the internet has absorbed a vast amount of religious text, but it has no understanding of the cultural context, the interpretive traditions, or the lived experience of religious communities. It can produce text that looks like religious writing, but it cannot produce text that embodies religious understanding.
This is the deepest problem with AI-generated religious content: it's not just wrong. It's hollow. It mimics the form without the substance. And readers who don't have the cultural knowledge to distinguish between the two are being misled.
I saw this dynamic play out in the crypto space in 2021, when a flood of NFT projects appropriated indigenous art and cultural symbols without permission or understanding. The projects were technically functional โ the tokens worked, the smart contracts executed โ but they were culturally destructive. They took something sacred and turned it into a commodity.
AI-generated religious books are doing the same thing. They're taking something sacred and turning it into a commodity. And the platforms that host them are profiting from the commodification.
The Regulatory Vacuum
Let me talk about regulation, because it's the missing piece of this puzzle.
Current publishing regulations focus on content legality โ obscenity, hate speech, defamation. They don't address AI-generated content at all. There's no legal framework for requiring AI content disclosure, no standards for AI content quality, no liability framework for AI-generated misinformation.
The EU's AI Act, which is the most comprehensive AI regulation in the world, includes transparency requirements for AI-generated content. But the implementation is still in progress, and the enforcement mechanisms are unclear. The FTC in the United States has shown some interest in AI content disclosure, but no concrete rules have been issued.
The regulatory vacuum means that platforms like Amazon have no legal obligation to address AI-generated content. They can continue to profit from the flood while claiming they're just a neutral platform.
This is the same regulatory vacuum we had in crypto before the 2024 ETF approval. For years, the crypto industry operated in a gray area, with no clear rules and no clear enforcement. The result was a flood of scams and a loss of public trust. The ETF approval brought regulatory clarity, but it also brought institutionalization โ and with it, a centralization that many of us in the purist camp find deeply uncomfortable.
I published a controversial op-ed in 2024 titled "Institutionalization vs. Ideology," arguing that while regulatory clarity is necessary, the current framework risks centralizing power back into traditional finance. I analyzed the custody structures of the top five ETF providers, highlighting a 95% reliance on centralized third parties. The response was overwhelming โ 2,000 emails from people thanking me for articulating their silent doubts about the "normalized" crypto space.
The same tension exists in publishing. We need regulation to address the AI content flood, but the regulation we're likely to get will favor the incumbents โ the large publishers, the established platforms, the centralized authorities. It will not favor the independent authors, the small publishers, or the diverse voices that the democratization of publishing was supposed to empower.
The Investment Angle
Let me briefly address the investment implications, because this is a question I get asked constantly.
The AI detection market is real, but it's not the investment opportunity that the marketing suggests. The market is estimated at a few hundred million dollars, growing at 20-30% annually. But the fundamental problems โ technical reliability, false positives, adversarial evasion โ limit the market's potential. Detection tools are a defensive service in an arms race, and defensive services in arms races have historically been poor investments.
The more interesting investment opportunity is in the "AI governance" space โ the infrastructure for content provenance, authenticity verification, and content tracing. This is the equivalent of the cybersecurity industry, which grew from a niche market to a multi-billion-dollar industry as threats multiplied. The same trajectory is likely for AI governance.
But here's my caution: the AI governance market is still in its early stages, and the winners haven't emerged yet. The technology is evolving rapidly, the regulatory landscape is uncertain, and the competitive dynamics are unstable. Investing in this space requires a long-term perspective and a tolerance for volatility.
This is the same advice I give to people asking about crypto investments in a bear market: survival matters more than gains. Focus on the protocols that are actually building value, not the ones that are just pumping tokens. The same logic applies to AI governance โ focus on the companies that are building real infrastructure, not the ones that are just publishing alarming studies to drive demand for their products.
The Path Forward
So where do we go from here?
I've been writing about the intersection of AI and blockchain since 2025, when I launched my "Human-Centric AI" initiative. The core principle of that work is simple: technology must be a servant to human values, not an autonomous master. This principle applies directly to the AI content problem.
We cannot solve the AI content flood with more technology alone. We need a combination of technical infrastructure, economic incentives, and cultural norms.
Technically, we need to build provenance systems that make content origin cryptographically verifiable. The C2PA standard is a good start, but it needs to be integrated into publishing platforms and made accessible to ordinary readers.
Economically, we need to create incentives for quality over volume. This might mean platform policies that reward verified human authors, or it might mean new business models that make low-quality content unprofitable.
Culturally, we need to develop norms around AI content disclosure. Readers should expect transparency about whether content is AI-generated, and authors should be proud to disclose their use of AI tools rather than hiding it.
None of this is easy. All of it requires coordination among stakeholders with conflicting interests. But the alternative โ continuing down the current path โ leads to a publishing ecosystem where readers can't trust what they read, authors can't compete with automated content factories, and the cultural heritage of religious traditions is flattened into algorithmic approximations.
I've been in this industry long enough to know that the bear market builds the foundation. The current chaos in publishing is the bear market โ the period of disruption and uncertainty that precedes the construction of something better. The question is whether we'll build that something better, or whether we'll let the chaos become the new normal.
A Personal Reflection
Let me end with a personal reflection, because I think it's relevant to the broader question of how we navigate this transition.
In 2022, after the Terra-Luna collapse, I spent six weeks in a cabin in rural Virginia, disconnected from all digital devices. It was the most difficult period of my professional life. I had believed in the promise of algorithmic stability, and the collapse had shattered that belief. I spent those six weeks not writing, not analyzing, not building โ just thinking.
What emerged from that solitude was a manuscript called "The Soul of Sovereignty," a book arguing that blockchain must serve human dignity, not just capital efficiency. The book was not a technical manual. It was a philosophical argument about the relationship between technology and human values.
The same argument applies to AI-generated content. The technology is not the problem. The problem is the values that guide its use. If we use AI to flood the market with low-quality content, we're using the technology to undermine human dignity. If we use AI to augment human creativity, to help authors write better books, to make publishing more accessible โ we're using the technology to serve human values.
The distinction is not always clear. AI-assisted writing is a spectrum, from fully automated generation to subtle grammar correction. The study doesn't distinguish between these cases, and the detection tools can't either. But the distinction matters, because it determines whether we're talking about a threat or an opportunity.
I believe we can navigate this transition. I believe we can build systems that preserve the democratizing promise of AI while protecting against its destructive potential. But it will require us to think differently โ not just about technology, but about values, incentives, and the kind of world we want to build.
The Takeaway
I keep coming back to a question that has haunted me since I first saw the 63% number: what does it mean for a reader to pick up a religious book, seeking guidance and wisdom, and unknowingly receive the output of a statistical model that has no understanding of the sacred?
This is not a technical question. It's a human question. And it's the question we should be asking as we build the infrastructure for the next era of content creation.
The tools we build will shape the answer. If we build detection tools that are unreliable and biased, we'll get false accusations and eroded trust. If we build provenance systems that are inaccessible and complex, we'll get a two-tier publishing market. If we build incentive structures that reward quality and transparency, we might actually get a publishing ecosystem that serves readers, authors, and the cultural traditions that sustain us.
The choice is ours. The technology is neutral. The values are not.
Truth is immutable, unlike the price action. And in the end, the truth about AI-generated content is that it's neither inherently good nor inherently bad. It's a tool. And like all tools, its value depends on how we use it.
The question is whether we'll use it to build or to destroy. And that's a question that no algorithm can answer.