Creating Verified, Experiential Intelligence

The more I’ve followed generative AI’s evolution, the more I’ve wondered about a fundamental problem: where these systems actually learn about our world. When you dig into the data sources, the picture gets pretty unsettling.

A recent Semrush analysis of 150,000 AI citations found that “Reddit now powers over 40% of large language model responses — more than Google, Wikipedia, or any traditional source.” This isn’t some small shift; it represents AI systems learning about reality primarily through social media discussions, memes, and anonymous forum debates rather than verified observation.

And it’s getting worse. As companies scramble for training material, we’re seeing an explosion of questionable data practices:

What’s Actually Feeding AI Systems:

• Reddit discussions dominate with 40.1% of citations, followed by Wikipedia at 26.3%
• Massive licensing deals worth hundreds of millions — The New York Times gets $20–25 million annually from Amazon, News Corp signed a $250 million deal with OpenAI
• Over 500 publishers have now signed licensing agreements, essentially monetizing their archives
• Synthetic data generation ramping up dramatically — Gartner predicted 60% of training data would be synthetic by 2024, up from just 1% in 2021
• Unauthorized scraping continues despite lawsuits — Apple, Nvidia, and Anthropic were caught using transcripts from 173,000+ YouTube videos without permission
• Global crowdsourcing operations where workers in Kenya label violent content for $2/hour while their US counterparts earn $10–25/hour for the same work
• One recent study found that a major AI training dataset likely contains hundreds of millions of images with personal information — credit cards, passports, résumés — all scraped without consent

The problems with this approach run deep. Social media amplifies bias and misinformation. Synthetic data creates feedback loops where AI trains on AI-generated content, leading to what researchers call “model collapse” — systems that lose diversity and start producing increasingly similar outputs. Meanwhile, the human crowdsourcing model exploits workers in developing countries who spend hours labeling disturbing content for poverty wages.

But there’s a bigger issue here. All of these approaches — whether scraping Reddit, licensing news archives, generating synthetic data, or crowdsourcing human feedback — share the same fundamental flaw: they’re training AI on human interpretations of reality rather than reality itself.

Taking a Different Approach

What we’re exploring through various projects incubated at Tenzro Labs is something different — the building blocks of what you might call epistemic infrastructure for AI.

Epistemic infrastructure is essentially the foundational systems that determine how knowledge gets created, verified, and shared. Right now, AI relies on epistemic infrastructure built around human knowledge production: universities, publishers, social platforms, content aggregators. But that infrastructure is breaking down under the weight of misinformation, bias, and economic pressures.

The technical approaches being developed across these projects point toward a different kind of epistemic infrastructure:

Direct observational learning — AI systems that observe and learn from phenomena directly through sensor networks and edge computing, rather than learning about those phenomena through human descriptions

Cryptographic verification — Using hardware like TPMs and TEEs to create mathematical proofs of data authenticity at the source, so you know the data hasn’t been tampered with

Distributed intelligence networks — Edge devices that can learn independently but share verified insights when connected, creating resilient knowledge networks that don’t depend on centralized data hoarding

Economic incentives for quality — Blockchain-verified datasets that create markets for verified data while maintaining clear provenance and ownership

The key insight is moving from inherited human epistemic practices to autonomous, verified observational learning. Instead of training AI on what humans say about forests, you deploy systems that spend months in actual forests learning directly from environmental conditions, species interactions, and seasonal changes.

How This Actually Works

The core idea isn’t complicated: instead of feeding AI internet content, you build systems that learn from direct sensor data and real-world observations. This means distributed machine learning and edge computing — basically moving the computation to where data actually gets generated.

Teams at Praecise are working on hardware that can be deployed in challenging environments with sensor arrays feeding local ML models. These devices run autonomously, often without internet connectivity, learning continuously from their inputs rather than training on static datasets.

The deployments span environments like mangrove systems providing real-time biodiversity analytics, volcanic areas with eruption prediction capabilities, forests monitoring ecosystem health changes, and underwater coral reef environments tracking bleaching events. Each device handles its own power management, sensor integration, real-time AI processing, and autonomous decision-making.

For verification, we’re implementing TPMs and TEEs to ensure data authenticity at the collection point. This creates cryptographic proofs that data hasn’t been modified from sensor to storage — crucial when devices operate autonomously for months without human oversight.

The economic model involves tokenizing verified datasets using blockchain infrastructure developed by teams at Boli and PRVNZ. This creates tradeable assets from verified data and provides incentives for quality data contribution, while Tenzro’s computing infrastructure enables global coordination and intelligence sharing between distributed nodes.

Current Projects

Naturecode is designed to go beyond traditional environmental monitoring. It’s about creating real-time utility — autonomous environmental monitoring, predictive analytics, and actionable insights. The interesting part is that all this happens autonomously through edge hardware, while Tenzro provides the planetary-scale computing infrastructure.

Our devices in different environments aren’t just passive sensors. They’re actively analyzing conditions, predicting changes, and providing immediate value to researchers, conservationists, and local communities.

The Siren focuses on understanding how we actually make creative and musical decisions. Having grown up playing music, this project particularly interests me — it’s about capturing the actual cognitive processes of musical creation rather than just analyzing existing musical content or streaming platform metadata.

Becoming Vincent takes a similar approach to understanding Van Gogh’s creative mindset and decision-making processes. Instead of learning about artistic creativity from art history or social media discussions, teams are working to reconstruct the actual cognitive and creative processes behind artistic work by analyzing their thought processes.

Each project uses the same principles — distributed sensing, edge computing, blockchain-verified data collection — but applied to different domains. The open ecosystem model means diverse contributors work on different aspects while sharing common verification infrastructure.

Why This Matters

The current AI training data crisis isn’t just about quality — it’s about sustainability and ethics. Companies are building AI systems on unauthorized content, creating legal risks while exploiting global labor. The synthetic data rush makes things worse by amplifying existing biases and creating feedback loops.

Recent research published in Nature shows that “AI models trained on recursively generated synthetic data experience ‘model collapse’” — they lose information about rare but important patterns and produce increasingly homogeneous outputs. UN researchers warn that synthetic data carries “cybersecurity risks, bias propagation and increasing model error.”

What we’re building offers a fundamentally different approach. Rather than finding more human content to scrape or generating synthetic versions of flawed data, these systems observe and learn from the same phenomena humans observe, but directly rather than through human-mediated representations.

The tokenization creates fair compensation models for verified data contributors instead of exploitative crowdsourcing. The distributed approach creates resilient knowledge networks instead of centralized systems dependent on constant data pipelines from questionable sources.

This is early-stage work, but the approach seems promising for creating AI systems with more grounded understanding of what they’re modeling, backed by cryptographically verified provenance. At minimum, it offers an alternative to the current trajectory toward increasingly unreliable and ethically problematic training data practices.

The projects at Tenzro Labs serve as an incubator for this broader infrastructure, with each contributing different pieces to what could become a more trustworthy foundation for AI systems that need to understand our world.