Everyone is obsessed with smarter AI models, but the uncomfortable truth is that the next wave of wins will come from something less flashy: trusted data. I’m joined by Daniela Braga, founder and CEO of Defined.ai. We dig into why the AI race is turning into a data race, and why closed source, consented datasets are becoming the real competitive edge. We also get honest about synthetic data risks, human accountability, and what “responsible AI” must mean when models start acting in the real world.
• Why AI labs did not consolidate and how open source and geopolitics reshaped competition
• Why closed source data is the differentiator once the open web is exhausted
• How synthetic data helps and how it can trigger model collapse
• Why CMOs and business leaders should treat data quality as brand risk
• How specialisation is driving more human curation in healthcare, legal, finance and physical AI
• What to measure to prove better data improves a model, including test splits and red teaming
• Why metadata quality determines whether AI powered search actually works
• What trusted data means in practice, including consent, payment and accurate metadata
• Where human accountability must stay as agents take actions
• How companies can earn real trust as misinformation rises and sources degrade
Why Closed Source Data Is The New Advantage In AI – with Daniela Braga
The AI conversation has shifted from “Who has the biggest model?” to “Who has the most trusted data?” On Web3 CMO Stories, Daniela Braga, CEO of Defined.ai, explains why ethically sourced training data is becoming the real moat as agentic AI spreads across industries. Frontier labs may crawl the open web daily, but the competitive edge increasingly sits in closed source data: private, proprietary, licensed, consented, and paid datasets with clear usage rights. For CMOs and founders, this is not an engineering detail. It is a brand reputation issue, a compliance issue, and a product differentiation issue that will shape how customers judge responsible AI.
A major theme is that better models do not reduce the need for better data, they amplify it. As companies build AI agents on top of large foundation models, fine tuning quality depends on the relevance, cleanliness, and governance of the underlying dataset. Daniela also flags synthetic data as both powerful and risky. Synthetic data can unlock hard to capture scenarios, but it must be balanced against the rising volume of AI generated content already polluting the web. When models train on synthetic outputs of other models, you risk model collapse, degraded accuracy, and hidden bias loops, especially if you cannot trace provenance and verify that your training pipeline is not eating its own exhaust.
The episode also makes a practical point about measurement: organisations need model evaluation, not vibes. Teams should keep a clean test split, probe models with structured challenges, and red team outputs at scale to see whether new data actually improves performance. This is where “human in the loop” remains essential, not as a talking point but as an operating system for quality. Human review can be synchronous or asynchronous, and it gets more important as data requests become specialised: healthcare, finance, legal, robotics, physical AI, and other domains where errors have real world consequences and where data must be curated, anonymised, and machine learning ready.
Another insight is that search itself is changing. Defined.ai uses an LLM assistant to help customers describe needs and discover datasets across massive catalogues, but Daniela notes that the quality of results depends heavily on metadata quality. Messy, missing, or inconsistent metadata is one of the most common failure points in data marketplaces and enterprise AI pipelines. The opportunity is faster discovery and better matching; the risk is false confidence when an AI assistant surfaces data that looks right but lacks consent, licensing clarity, or accurate labels. Responsible AI governance therefore needs to include not only the dataset, but also the layer that helps users select it.
Finally, the conversation turns blunt on accountability. As AI systems move from generating content to taking actions autonomously, Daniela argues that human accountability must remain with the companies that deploy models and agents, not just the downstream builders. She pushes back on simplistic indemnification thresholds and highlights the ethical problem of pricing harm in “number of lives.” Trust, in this framing, is earned through traceability, paid rights, consent, strong evaluation, and clear limits on where autonomy is acceptable. For leaders, the takeaway is simple: invest in trusted, high quality data and provenance now, because the next competitive era of AI will be decided less by who can generate, and more by who can prove.
Want to see more of my articles in Google? Add webdrie.net as a preferred source.








