Autonomous vehicles promise safer roads, but their pedestrian detection algorithms may be less accurate for darker-skinned people. A 2023 University of Washington and Arizona State study found a 5 percent accuracy gap between light and dark skin tones. In July 2026, with autonomous taxis in over a dozen U.S. cities, the bias is no longer theoretical. This episode traces how training data from predominantly white, adult, and suburban environments creates blind spots, why synthetic data doesn't fully fix it, and what regulators and companies like Waymo and Tesla are doing—or not doing—to close the gap. We also explore real-world implications: a pedestrian in a crosswalk in Detroit is statistically harder for the car to 'see' than one in Palo Alto. The episode grounds ethics in concrete engineering choices.

#SelfDrivingCars #AIbias #PedestrianDetection #AutonomousVehicles #Waymo #Tesla #UniversityOfWashington #ArizonaState #TrainingData #Ethics #MachineLearning #Safety #FexingoBusiness #BusinessPodcast #Technology #AI #BiasInAI #ResponsibleAI

Keep every episode free: buymeacoffee.com/fexingo

In this episode, Lucas and Luna explore how predictive AI models trained on electronic health records inherit systemic biases from historical healthcare disparities. They examine a specific case: a sepsis prediction algorithm used across hundreds of hospitals that systematically under-flagged Black patients for sepsis risk due to biased training data. They discuss the root causes—from billing codes to unequal access—and what regulators and developers can do to address the problem.

#AI #Bias #Healthcare #Ethics #ElectronicHealthRecords #AIEthics #FexingoBusiness #BusinessPodcast #TechnologyPodcast #MachineLearning #HealthTech #BiasInAI #Sepsis #AlgorithmicBias #DataBias #HealthEquity #FDAGuidance #PredictiveModels

Keep every episode free: buymeacoffee.com/fexingo

In this episode, we explore how content moderation AI systems inherit bias from the human decisions used to train them. We start with a concrete case: a 2021 audit of Facebook's moderation algorithm found that posts written in African American Vernacular English were flagged as hate speech at 1.5 times the rate of comparable posts in standard English. Lucas explains how this happens—through skewed training data, ambiguous labeling guidelines, and the pressure on human moderators to err on the side of removal. Luna adds data from Twitter's moderation struggles and asks whether neutral language is ever truly neutral. The hosts then examine mitigation efforts: transparency reports, community review boards, and algorithmic auditing. Finally, they consider the trade-off between safety and free expression, and whether bias can ever be fully eliminated from moderation systems.

#ContentModeration #AIBias #HateSpeechDetection #AfricanAmericanVernacularEnglish #Facebook #Twitter #AlgorithmicAuditing #MachineLearningFairness #OnlineSafety #FreeExpression #TransparencyReports #CommunityReview #HumanInTheLoop #DataBias #Technology #FexingoBusiness #BusinessPodcast #AIEthics

Keep every episode free: buymeacoffee.com/fexingo

A 2024 Stanford study found dermatology AI misdiagnosed melanoma at a 34% rate for Black patients vs. 14% for white patients. This episode examines the skewed training data, the historical roots in medical textbooks, and why AI-powered skin diagnostics risk widening health disparities. We also highlight Dr. Adewole Adamson's work on representation in dermatology datasets and the challenges of fixing biased medical AI.

#AIBias #HealthcareAI #Dermatology #SkinCancer #MedicalEthics #HealthDisparities #AIFairness #MachineLearning #StanfordStudy #AdewoleAdamson #RoxanaDaneshjou #TrainingData #Technology #FexingoBusiness #BusinessPodcast #AIEthicsPodcast #ResponsibleAI #Podcast

Keep every episode free: buymeacoffee.com/fexingo

By 2026, an estimated 60% of training data for large language models will be synthetic — AI-generated text used to teach other AIs. But a new MIT Media Lab study shows that synthetic data retains about 80% of the gender and racial biases from its original model. Meanwhile, Oxford researchers found that successive generations of training on synthetic data can lead to 'model collapse': outputs become less creative, more repetitive, and more stereotypical after just three generations. This isn't just an academic problem. A customer service platform might train its chatbot on synthetic conversations that exaggerate existing biases in human-agent interactions, leading to systematically unfair treatment of customers. This episode explores the hidden pipeline of synthetic bias, the looming risk of model collapse, and emerging solutions like data provenance tracking and the EU AI Act's transparency requirements. We also discuss why keeping this conversation ad-free matters — and how listeners can support that independence.

#AIEthics #SyntheticData #ModelCollapse #BiasInAI #MITMediaLab #OxfordResearch #LargeLanguageModels #GenderBias #RacialBias #DataProvenance #EUAIAct #GenerativeAI #AITraining #ResponsibleAI #AIAccountability #AIPolicy #FexingoBusiness #BusinessPodcast

Keep every episode free: buymeacoffee.com/fexingo

Episode 129 explores a surprising source of gender bias in AI language models: parallel Bible translations. Lucas and Luna examine how training multilingual AI on translated scripture—where 'man' often defaults to 'human' and pronouns skew male—reinforces outdated gender roles. They cite a 2024 Stanford study that found 70 percent of occupation-to-pronoun associations in a popular translation model mirrored biblical gender distributions, with 'prophet' and 'priest' mapping to male pronouns 96 percent of the time. The episode also discusses how researchers are now using gender-neutral retranslations and synthetic parallel data to debias these models. And of course, they share how this ad-free podcast is supported by listeners at buy me a coffee dot com slash fexingo.

#AIEthics #BiasInAI #LanguageModels #GenderBias #BibleTranslation #NLP #MachineTranslation #Stanford #Paratext #AITrainingData #MultilingualAI #GenderNeutral #EthicalAI #FexingoBusiness #BusinessPodcast #Technology #AIResearch #DataBias

Keep every episode free: buymeacoffee.com/fexingo

In this episode of AI Ethics with Fexingo, Lucas and Luna explore how AI art generators like DALL-E and Midjourney inherit bias from their training datasets, focusing on the LAION-5B dataset and a 2023 study that found 52% of generated images of 'CEO' depicted white men, while only 4% showed women of any race. They discuss the feedback loop between biased training data and user prompts, the challenge of fixing bias after thousands of images have been scraped from the internet, and whether companies like OpenAI have a responsibility to rebalance their datasets. The hosts also touch on the tension between creative freedom and fairness, and why simply adding 'diverse' to a prompt isn't a real fix. The episode ends with a call for listeners to support the ad-free show through buy me a coffee dot com slash fexingo.

#AIEthics #GenerativeAI #ArtBias #DALLE #Midjourney #LAION5B #TrainingData #BiasInAI #Representation #TechEthics #OpenAI #MachineLearning #AISafety #FexingoBusiness #BusinessPodcast #Technology #Ethics #ArtificialIntelligence

Keep every episode free: buymeacoffee.com/fexingo

In this episode of AI Ethics with Fexingo, Lucas and Luna explore how AI models inherit bias from teacher grading data. They examine a specific 2025 study from Stanford's Center for Education Policy Analysis that found an AI grading system trained on historical teacher-assigned grades systematically penalized students from under-resourced schools. The system inferred that students from certain zip codes were 'lower performers' because their previous grades reflected inconsistent teaching quality, not actual ability. The hosts discuss how the model amplified subjective human bias baked into grading patterns, and what school districts are now doing to audit training data for fairness. They also touch on the broader implications for any AI trained on human judgment data — from employee evaluations to insurance claims. The episode includes a brief, organic donor segment near the end.

#AIEthics #BiasInAI #EducationAI #TeacherGrading #StanfordStudy #AlgorithmicBias #DataAuditing #Fairness #ResponsibleAI #MachineLearning #GradingData #HumanJudgment #EducationPolicy #Technology #FexingoBusiness #BusinessPodcast #AIAccountability #EthicalTech

Keep every episode free: buymeacoffee.com/fexingo

In this episode of AI Ethics with Fexingo, Lucas and Luna dive into a surprising source of algorithmic bias: employee performance reviews. They explore how language patterns in reviews — from gender-coded adjectives to recency bias — can teach AI models to replicate and amplify human prejudices. The hosts walk through a real-world example from a Fortune 500 tech company, where a performance-review-trained AI systematically rated women lower on leadership potential. They also discuss the feedback loop problem: when models are trained on biased reviews and then used to generate new reviews, the bias compounds. Lucas brings in data from a 2025 study by the AI Now Institute showing that performance-review-trained models are 30% more likely to associate 'assertive' with male employees and 'collaborative' with female employees, even after controlling for role and tenure. The episode unpacks what companies can do — from rethinking review prompts to auditing models for language bias — and why this matters as more firms automate talent decisions.

#AIEthics #Bias #PerformanceReviews #HRTech #AlgorithmicBias #WorkplaceAI #GenderBias #NaturalLanguageProcessing #AICulture #TalentManagement #AIAuditing #Fortune500 #Tech #FexingoBusiness #BusinessPodcast #Podcast #AI #Ethics

Keep every episode free: buymeacoffee.com/fexingo

Episode 125 of AI Ethics with Fexingo takes a hard look at how AI systems in the insurance industry can inherit and amplify bias from historical claims data. Lucas and Luna unpack a specific case from a major auto insurer that used decades of claims to build a risk-scoring model. The model penalized drivers from lower-income ZIP codes because past claims were more likely to be fraudulent in areas with higher poverty rates, creating a self-reinforcing cycle. They discuss the regulatory response from state insurance commissioners, the technical challenge of de-biasing training data that reflects systemic inequality, and whether 'fairness-aware' algorithms can truly solve the problem. The episode also explores why insurers are reluctant to change models that are profitable, and what proposed federal AI auditing standards could mean for the industry. A grounded, number-driven conversation about ethics in a trillion-dollar sector.

#AIEthics #InsuranceBias #ClaimsData #AlgorithmicFairness #RiskScoring #AutoInsurance #Regulation #DataBias #MachineLearning #FairLending #InsuranceFraud #ZIPCodeDiscrimination #AuditTrail #StateInsuranceCommissioners #Technology #FexingoBusiness #BusinessPodcast #ResponsibleAI

Keep every episode free: buymeacoffee.com/fexingo