Trustworthy AI Starts Before the Model Runs
Five AI takeaways from HumanX conference plus AI ethics and data sourcing discussion.
Original episode title: 5 AI Takeaways from HumanX + AI Ethics w/ Defined.AI
Guest
Daniela Braga
AI data and ethics leader
Daniela Braga is the founder and CEO of Defined.ai, a company that supplies training data and data services for artificial-intelligence development. She has focused on data quality, provenance, diversity, and ethical sourcing, arguing that model performance and responsible development depend on how datasets are assembled and governed.
Moves trustworthy AI upstream to training-data provenance, consent, bias, labor conditions, and responsibility for the inputs beneath a model.
What this conversation is really about
After a host-reported tour of HumanX themes, Daniela Braga pulls the story upstream to training data. Provenance, copyright, representation, consent, labeling labor, and liability shape the system before a user sees an output. That matters because output controls cannot repair every problem embedded in the inputs. The episode also connects model quality to hidden human work: who collects and labels data, under what conditions, and with what right to the material? The claims are valuable industry testimony, not an independent audit of any dataset or company practice.
From the conversation
The argument in focus
Daniela Braga
AI data and ethics leader
Daniela Braga is the founder and CEO of Defined.ai, a company that supplies training data and data services for artificial-intelligence development. She has focused on data quality, provenance, diversity, and ethical sourcing, arguing that model performance and responsible development depend on how datasets are assembled and governed.
Moves trustworthy AI upstream to training-data provenance, consent, bias, labor conditions, and responsibility for the inputs beneath a model.
Evidence status
Vendor-reported case
A data-company leader describes provenance, bias, labor, and commercial practices from inside the industry. Conference examples and company claims require primary and independent corroboration.
Boundary map
Where the system stops
- What the system handles
- Large-scale data collection, labeling, curation, model training inputs, and automated workflows built on those inputs.
- What remains human
- Consent, contracting, labor standards, source documentation, bias review, liability, and the decision about which data may be used at all.
- What remains open
- Can organizations trace a consequential output back through model behavior to lawful, representative, and responsibly produced training inputs?
Ideas worth carrying forward
- Ask where data came from before asking only how a model performs.
- Treat labeling labor and consent as quality and governance questions.
- Require provenance records that can survive procurement and incident review.
- Separate conference forecasts from demonstrated deployments.
What this changes Monday
Select one consequential AI system and request its data lineage: source categories, permissions, collection method, labeling workforce, representativeness checks, exclusions, and known gaps. If a vendor cannot supply enough information for risk review, record that absence as evidence rather than filling it with confidence language.
Original episode notes
We’re reporting live from HumanX, one of the world’s largest AI conferences, where we’ve spent three intense days immersed in AI discussions, moderating panels, and speaking with some of the most influential voices in the industry.
In this episode, we break down five takeaways from the conference, covering everything from AI’s growing role in politics to the future of AI-driven customer service and automation.
Then, we sit down with Daniela Braga, CEO of Defined AI, a leader in ethically sourced AI training data. We discuss the urgent need for transparency in AI data sourcing, the dangers of bias in large language models, and the growing demand for differentiated AI solutions.
5 takeaways from HumanX:
•AI is becoming political: Kamala Harris’s speech and the future of AI policy
•The death of human customer support and the rise of AI customer support (hot take: this is a good thing!)
•AI agents are finally getting real—how companies are deploying them today
•Businesses are scrambling to implement AI automation at scale
•Leaders are more AI-curious than ever—how executives are integrating AI into decision-making
Interview with Daniela Braga (CEO of Defined AI):
•The problem with AI training data: bias, legal concerns, and digital exploitation
•Why most AI models sound the same—and how that could change
•The myth that we’re “running out of data” and the future of AI training sets
•What’s next? Predictions for AI regulation, ethical sourcing, and industry differentiation
It’s a packed episode full of insights straight from the conference floor.
Defined.AI:
https://defined.ai/
🎧 Subscribe to AI-Curious:
• Apple Podcasts
https://podcasts.apple.com/us/podcast/ai-curious-with-jeff-wilser/id1703130308
• Spotify
https://open.spotify.com/show/70a9Xbhu5XQ47YOgVTE44Q?si=c31e2c02d8b64f1b
• YouTube
https://www.youtube.com/@jeffwilser