Jul 10, 2026 · 54 min · 11 segments
On the seventh anniversary episode of Events Demystified Podcast, Anca Platon Trifan explains how AI bias and stereotyping are amplified through generative and predictive systems, starting with her…
Anca TrifanHost
This conversation has been on my mind for a long time because my work sits so deeply both inside the AV and the AI worlds.
In AV, I have experienced bias personally, and I have talked at large about it both on the podcast and in articles and in blogs and on LinkedIn and anyone and everywhere where I had a platform.
In AI, I can see how those same assumptions are being absorbed, they're being automated and then redistributed sometimes at such a larger scale than what it happened before.
That does not release technology companies from responsibility, but it does not excuse harmful models or suggest that nothing can be done.
It just means that we need to understand the actual source of the problem before pretending that a technical patch will just solve it.
They learn from our books and our websites and images and records and decisions and language and institutions and history.
It also contains discrimination and exclusion and stereotypes and incomplete information and unequal access and entire groups of people whose experiences were either poorly documented or totally ignored.
When we train AI on historical data, we are teaching it what happened before, right? We are not automatically teaching it what should happen next.
Now imagine a company that has spent 30 years hiring men into most of its technical leadership roles.
The company decides to use AI to identify which current employees have the characteristics of successful past leaders.
Perhaps the previous leaders had similar career paths, similar communication styles, attended similar schools, worked similar hours, or participated in similar professional networks.
The AI may find those patterns accurately, right? So the problem here is that this historical record was shaped by who received those opportunities, who was mentored, who could work those hours, who was perceived as leadership material and who was allowed into the room in the first place.
So then the model will accurately produce the company's past while completely failing to identify the leaders the company needs for its future.
which is a mix of male and female, not just one kind, right? So one of the most frequently discussed examples comes actually from Amazon.
The company developed an experimental recruiting tool trained using roughly 10 years of employment applications.
This conversation has been on my mind for a long time because my work sits so deeply both inside the AV and the AI worlds.
In AV, I have experienced bias personally, and I have talked at large about it both on the podcast and in articles and in blogs and on LinkedIn and anyone and everywhere where I had a platform.
In AI, I can see how those same assumptions are being absorbed, they're being automated and then redistributed sometimes at such a larger scale than what it happened before.
That does not release technology companies from responsibility, but it does not excuse harmful models or suggest that nothing can be done.
It just means that we need to understand the actual source of the problem before pretending that a technical patch will just solve it.
They learn from our books and our websites and images and records and decisions and language and institutions and history.
It also contains discrimination and exclusion and stereotypes and incomplete information and unequal access and entire groups of people whose experiences were either poorly documented or totally ignored.
When we train AI on historical data, we are teaching it what happened before, right? We are not automatically teaching it what should happen next.
Now imagine a company that has spent 30 years hiring men into most of its technical leadership roles.
The company decides to use AI to identify which current employees have the characteristics of successful past leaders.
Perhaps the previous leaders had similar career paths, similar communication styles, attended similar schools, worked similar hours, or participated in similar professional networks.
The AI may find those patterns accurately, right? So the problem here is that this historical record was shaped by who received those opportunities, who was mentored, who could work those hours, who was perceived as leadership material and who was allowed into the room in the first place.
So then the model will accurately produce the company's past while completely failing to identify the leaders the company needs for its future.
which is a mix of male and female, not just one kind, right? So one of the most frequently discussed examples comes actually from Amazon.
The company developed an experimental recruiting tool trained using roughly 10 years of employment applications.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
