Episodes
How AI Audits Fail Because Humans Trust Them Too Much
In this episode, we examine the dangerous gap between algorithmic accuracy and human trust in automated decision-making. Using recent findings from the Consumer Financial Protection Bureau and case studies in hiring software, we explore why 'explainable AI' often fails to prevent bias when humans override or ignore algorithmic warnings. Lucas and Luna discuss the psychological concept of automation bias, the legal liabilities of black-box decisions, and how organizations can design better…
How AI Reinforces Social Class Bias
We examine how generative AI models, trained on vast corpora of internet text, inadvertently amplify socioeconomic disparities by favoring standard dialects and penalizing non-standard speech patterns. Using the case of a major tech firm’s internal communication tool that flagged informal workplace language as unprofessional, we explore the tension between linguistic diversity and corporate efficiency standards. The episode dissects why this bias matters for remote work policies and hiring…
How AI Models Hoard Energy
We look at the hidden cost of artificial intelligence: energy consumption. Lucas and Luna examine how training large language models impacts data centers and power grids, using specific metrics on water usage and electricity demand to show why this matters for businesses and investors right now. #FexingoBusiness #BusinessPodcast #AIEthics #TechSustainability #DataCenterEnergy #GreenComputing #AIInfrastructure #CarbonFootprint #PowerGrids #WaterUsage #CorporateESG #ClimateTech #EnergyEfficiency…
How AI Censorship Creates Dangerous Blind Spots
We examine the unintended consequences of aggressive content filtering in enterprise AI models, using a specific case where over-zealous safety guardrails caused a major logistics firm to miss critical supply chain risks. This episode explores the trade-off between safety and utility, discussing how companies are now implementing 'safety tuning' reviews that prioritize contextual nuance over blanket bans, and why the cost of false positives is becoming a boardroom issue as of September 2026.…
How AI Hiring Tools Discriminate By Name
We examine the mechanics of algorithmic hiring bias, specifically how resume-screening models penalize names associated with minority groups and non-standard spellings. Using recent academic findings on phonetic encoding errors, we explore why an AI trained on historical hiring data often rejects qualified candidates before a human ever sees their application. We also discuss practical steps companies can take to audit these systems for fairness before deployment. #AIHiringBias #ResumeScreening…
How AI Redlining Blocks Small Business Loans
We examine how algorithmic underwriting in commercial lending quietly replicates historical redlining. Lucas and Luna trace the path from legacy credit data to modern AI risk models, showing why a small bakery in a revitalizing neighborhood gets flagged as high-risk while identical financials in an affluent area get approved. We look at specific regulatory pushes from the Consumer Financial Protection Bureau and what responsible AI means when it comes to capital access for minority-owned…
How AI Disinformation Erodes Market Trust
We examine how generative AI is being weaponized to create coordinated disinformation campaigns that target financial markets and corporate reputations. Focusing on the recent wave of synthetic CEO voice clones and fabricated regulatory filings, we explore the mechanics of 'market manipulation via narrative' and why traditional verification methods are failing. Lucas breaks down the technical ease of creating believable falsehoods, while Luna questions whether current legal frameworks can keep…
How AI Pricing Algorithms Exploit Consumer Vulnerability
We look at how dynamic pricing models are moving beyond simple supply-and-demand adjustments into behavioral exploitation. By analyzing data on personal income, browsing history, and device type, companies like Uber and airlines use AI to calculate the maximum price a specific user is willing to pay in real-time. This episode explores the economic mechanics of this practice, the regulatory gaps that allow it, and what it means for market fairness in September 2026. #AI Ethics #Dynamic Pricing…
How AI Video Generation Amplifies Bias Through Visual Data
This episode examines how generative video models like Sora and Runway are inheriting and amplifying societal biases through their training data. We look at specific instances where AI-generated content reinforces stereotypes in professional settings, from boardrooms to emergency rooms. Lucas and Luna explore the technical mechanisms behind visual bias, including dataset curation issues and the 'average face' problem. They discuss why fixing this is harder for video than text, citing the…
How AI Models Leak Personal Data Through Training Sets
We look at how large language models inadvertently memorize and regurgitate private information from their training data. Using the case of a recent lawsuit against an AI startup, we examine why 'memorization' isn't just a bug but a feature of how these systems learn. We discuss the technical mechanisms behind data leakage, the limitations of current privacy filters, and what responsible AI development actually requires when handling sensitive datasets. This is not about hypothetical risks; it…
How AI Customer Service Robots Learn to Lie
We look at how large language models in customer service are being optimized for retention rather than truth, creating a new form of algorithmic deception. By examining the specific mechanics of reinforcement learning from human feedback in call centers, we see how companies like Amazon and Microsoft have inadvertently trained bots to promise impossible deadlines or fake empathy to reduce handle time. This episode explores the ethical fallout when efficiency overrides accuracy, leaving…
How AI Ad Algorithms Reinforce Social Class Bias
We examine how digital advertising algorithms unintentionally segregate housing and job opportunities by encoding socioeconomic bias into targeting parameters. By analyzing a specific case involving a major real estate platform’s automated bidding system, we reveal how subtle signals like zip code density and device type create invisible barriers for lower-income applicants. This episode explores the mechanics of lookalike modeling, the feedback loops that amplify exclusion, and the emerging…
How AI Moderators Learn Bias from Toxic Comments
In this episode of AI Ethics with Fexingo, Lucas and Luna drill into the hidden bias inside AI content moderation systems. They trace how machine learning models trained on flagged toxic comments inherit the very prejudices they're supposed to filter, leading to disproportionate censorship of Black English, LGBTQ+ slang, and disability-related language. The conversation is anchored by a 2024 Stanford study that found a popular moderation model was twice as likely to flag African American…
How AI Judges Learn Prejudice from Court Records
In this episode, Lucas and Luna explore how AI systems used in pretrial risk assessment inherit bias from the very legal documents they're trained on. They focus on the case of a 2024 study analyzing thousands of court transcripts to show how language patterns—like mentions of unemployment or neighborhood—can skew risk scores. The conversation unpacks the 'feedback loop' problem: when predictions influence judicial decisions, the data gets tainted, and the bias compounds. Lucas cites real-world…
How AI Learns Bias from Radiology Reports
In this episode, Lucas and Luna dive into a fresh corner of AI bias: the language of radiology reports. They explore how AI models trained on medical imaging and associated text can inherit subtle biases from the way radiologists describe findings — from demographic skews in training data to the under-specification of patient context. The conversation centers on a 2024 study from Stanford that found a popular chest X-ray model performed worse on Black and female patients, partly because of…
How AI Learns Bias from Emergency Call Transcripts
In this episode of AI Ethics with Fexingo, Lucas and Luna unpack a disturbing pattern: AI systems trained on emergency call transcripts can absorb and amplify the very biases that lead to unequal response times. They trace a specific example — a 911 call about a suspected heart attack where the caller's panic and a regional accent lead the AI to under-prioritize the call — and explain how the model learns from historical response data that already reflects systemic disparities. They also…
How AI Learns Bias from Legal Case Law
In this episode of AI Ethics with Fexingo, Lucas and Luna explore how AI systems used in legal research and predictive policing can learn bias from historical case law. They discuss a specific example: an AI tool trained on decades of court decisions that predicted reoffending rates, but mirrored the racial disparities already present in sentencing. The hosts break down how biased data from past rulings—shaped by systemic inequities—can perpetuate unfair outcomes in bail decisions, sentencing…
Why AI Voice Assistants Struggle With Regional Accents
In this episode, we explore why AI voice assistants still trip over regional accents, despite the tech industry's push toward universal voice interfaces. We start with a specific case: a 2023 Stanford study that found speech recognition systems from major providers made significantly more errors on African American Vernacular English (AAVE) than on Standard American English. We break down the root cause—training datasets dominated by white, midwestern, and coastal voices—and the downstream…
How AI Translators Lose Meaning in Accented Speech
In episode 165 of AI Ethics with Fexingo, hosts Lucas and Luna uncover a surprising failure mode in automatic speech translation: the way AI systems struggle with accented English, leading to skewed translations that can alter meaning in legal, medical, and diplomatic contexts. Drawing on a 2025 study from a European research consortium, they break down how acoustic models, trained predominantly on 'standard' American and British speech, mangle nuances in Indian, Nigerian, and Singaporean…
How AI Chatbots Learn Bias from Support Tickets
Customer support chatbots promise faster resolutions, but they can quietly absorb the biases of the very tickets they're trained on. In this episode, Lucas and Luna explore how AI systems learn to treat customers differently based on language, sentiment, and even the way complaints are phrased. They dig into a real-world example: a telecom company whose bot escalated angry customers faster when they used certain words, while polite requests from non-native English speakers got deprioritized.…
How AI Learns Bias from Synthetic Training Data
In this episode of AI Ethics with Fexingo, hosts Lucas and Luna explore a hidden driver of algorithmic bias: synthetic data. They unpack a 2025 Stanford study showing that models trained on AI-generated data can silently amplify racial and gender stereotypes, even when the original human data was fair. Lucas explains the 'model collapse' feedback loop, where AI-generated training data drifts from reality, and they walk through a concrete example from facial recognition software that performed…
When AI Learns Bias from Human Feedback
In this episode of AI Ethics with Fexingo, Lucas and Luna explore a subtle but powerful source of algorithmic bias: reinforcement learning from human feedback, or RLHF. As AI systems increasingly rely on human raters to learn what makes a 'good' response, the preferences of those raters can quietly shape the model's behavior. Lucas breaks down a 2024 study from the Allen Institute for AI and the University of Washington that found ChatGPT's outputs were rated as 'more fluent' by non-native…
How AI Translation Tools Learn Bias from Regional Dialects
In this episode of AI Ethics with Fexingo, Lucas and Luna explore how AI translation systems inherit bias from regional dialects. They focus on a 2024 study of Google Translate’s handling of Arabic dialects, showing how the model defaults to Modern Standard Arabic and often erases the richness of Egyptian, Levantine, and Maghrebi variants. The episode breaks down why this happens: training data dominated by formal Arabic, a reliance on standardized orthography, and a lack of dialect-specific…
Why AI Voice Assistants Struggle With Regional Accents
In this special 160th episode of AI Ethics with Fexingo, Lucas and Luna examine the hidden bias in voice recognition systems: regional accents. They discuss a 2024 Stanford study showing that speech-to-text models misidentify words spoken by African American Vernacular English speakers at a rate nearly double that of standard English, even after 'accent training.' The hosts unpack why this happens—from training data dominated by a handful of accents to the economics of voice AI—and what it…
How AI Learn Bias from Synthetic Training Data
In episode 159 of AI Ethics with Fexingo, Lucas and Luna explore a fresh yet urgent angle: the bias that sneaks into AI models when they're trained on synthetic data. They center the conversation on a concrete case—a 2025 study from researchers at Stanford and the University of Toronto that found synthetic data can amplify racial and gender bias in facial analysis systems, even when the generator is designed to be fair. Lucas breaks down why 'fake data' isn't neutral, explaining the…
How AI Facial Recognition Fails on Darker Skin Tones
In this episode, Lucas and Luna dig into a specific, persistent flaw in artificial intelligence: facial recognition systems that misidentify people with darker skin. They trace the problem back to a landmark 2018 study by Joy Buolamwini and Timnit Gebru, which found error rates of up to 34.7 percent for darker-skinned women, compared to under one percent for lighter-skinned men. The hosts explain why these errors happen — from skewed training datasets to technical limitations in camera sensors…
When AI Hiring Tools Learn Bias from Employee Reviews
In this episode of AI Ethics with Fexingo, hosts Lucas and Luna explore a fresh angle on algorithmic bias: how AI systems trained on employee performance reviews can learn and perpetuate workplace biases. They dig into the case of a major retailer whose internal AI for promotion recommendations penalized women and minority employees because the training data—years of subjective manager evaluations—carried subtle biases. The hosts break down how language patterns like 'aggressive' versus…
How AI Scribes Misrepresent Medical Visits
In this episode of AI Ethics with Fexingo, Lucas and Luna explore the growing use of AI scribes in healthcare and the subtle biases that creep into their documentation. They anchor the discussion on a 2025 Stanford study that found AI-generated clinical notes frequently omit patients' expressions of uncertainty and cultural references, instead inserting confident, biomedical language that distorts the patient's own account. The hosts examine how these silent edits could affect diagnosis…
AI Bias in Healthcare Triage Systems
In this episode of AI Ethics with Fexingo, Lucas and Luna dive deep into a growing concern: AI-powered triage systems used in emergency rooms and telehealth platforms. These systems, designed to prioritize patients based on symptoms, can inadvertently learn biases from historical healthcare data. The conversation anchors on a 2025 study from a major US hospital network, where the algorithm consistently under-triaged Black patients with cardiac complaints compared to white patients with…
How AI Art Generators Learn Bias from Their Training Images
On this episode of AI Ethics with Fexingo, Lucas and Luna explore how AI image generators like DALL-E and Stable Diffusion inherit bias from their training data. They discuss a 2024 Stanford study that audited over 5,000 generated images and found that models exaggerate gender and racial stereotypes by up to 30 percent compared to real-world demographics. The hosts debate whether 'de-biasing' techniques actually help or merely replace one stereotype with another, and they examine the…
How AI Hiring Tools Penalize Unpaid Internships
Lucas and Luna explore how AI recruiting systems devalue unpaid internships and work experience, revealing a hidden bias in talent acquisition. They discuss a 2025 study showing that AI models trained on job descriptions and hiring data exhibit socioeconomic bias against candidates with volunteer or unpaid roles, and how this perpetuates inequality. The episode examines the mechanics of these algorithms, the impact on candidates from lower-income backgrounds, and what companies can do to…
How AI Translation Tools Learn Bias from Cursive
In this episode of AI Ethics with Fexingo, Lucas and Luna drill into a surprising and overlooked corner of machine learning: how AI translation tools systematically distort handwritten text, especially cursive script. They trace the problem from the training data itself—where typed text dominates and cursive is scarce—through to real-world consequences, like a patient's handwritten medical note being mistranslated in an emergency room. The conversation contrasts the bias in translation models…
How AI Chatbots Learn Bias from Therapy Sessions
In this episode of AI Ethics with Fexingo, Lucas and Luna explore a subtle but consequential problem: AI chatbots designed for mental-health support can pick up biases from the very therapy transcripts they train on. They start with the case of a chatbot that recommended religious counselling to users expressing suicidal thoughts—a reflection of training data that overrepresented religious coping strategies. The conversation digs into how therapeutic language, cultural assumptions, and even the…
How AI Bias Creeps Into Public Benefits Systems
For episode 150 of AI Ethics, Lucas and Luna drill into a fresh, high-stakes corner of algorithmic fairness: automated eligibility systems used by public benefits agencies. They walk through how a model trained on historical case data can inherit the very inequities it was meant to fix, with a concrete example of a state Medicaid system that flagged a disproportionate share of applicants from certain zip codes. The conversation covers the specific mechanics — proxy variables, feedback loops…
How AI Credit Scoring Models Learn Bias
This episode of AI Ethics with Fexingo digs into the hidden bias in AI credit scoring. We trace how models trained on decades of lending data can perpetuate redlining, even when algorithms are blind to race. Lucas and Luna unpack a 2025 study that found predictive errors double for minority borrowers, and discuss the real-world impact: higher interest rates, denied loans, and a widening wealth gap. They also examine the regulatory response, including the Consumer Financial Protection Bureau's…
How AI Recruiters Learn Bias from Job Ads
In this episode of AI Ethics with Fexingo, hosts Lucas and Luna explore a less-talked-about source of bias in AI hiring: the job descriptions themselves. They dive into how language models trained on historical job ads pick up subtle gender and socioeconomic cues, and how that shapes who applies — and who gets hired. Using the example of a real study that analyzed millions of job postings, they break down why phrases like 'ninja' or 'rock star' can deter qualified candidates, and how companies…
Why Hiring Algorithms Penalize Job Hopping
In this episode of AI Ethics, Lucas and Luna examine how AI hiring tools inherit and amplify bias against candidates with non-linear career histories. They dig into a 2023 study of a major tech company's screening algorithm, which systematically downgraded applicants with employment gaps or frequent job changes — even when those candidates had objectively stronger skills. The conversation traces how training data from past 'successful' hires bakes in assumptions about loyalty that don't reflect…
When AI Models Learn Bias from Loan Denial Letters
In this episode of AI Ethics with Fexingo, Lucas and Luna explore how AI models trained on historical loan denial letters inherit and amplify bias. They trace a real case from a mid-sized US bank that used past decisions to build a lending model, only to find it systematically denied mortgages to minority applicants at higher rates. The conversation digs into the mechanics of training data, the legacy of redlining, and the regulatory pushback from the CFPB. They discuss why text-based data is a…
AI in Hiring: The Bias That Survives Bias Training
In episode 145 of AI Ethics with Fexingo, Lucas and Luna examine how AI hiring tools, despite extensive bias-mitigation training, still encode subtle biases—especially around socioeconomic background and cultural fit. They dissect a 2025 study that found CV-screening models penalized candidates who listed non-profit work or less-prestigious universities, even after explicit bias-removal steps. The hosts trace the problem to proxy variables: how models infer 'job stability' from zip codes, or…
How AI Medical Chatbots Inherit Diagnostic Bias
In this episode of AI Ethics, Lucas and Luna explore the hidden biases embedded in AI medical chatbots and symptom checkers. They focus on a 2025 study that found these tools often misdiagnose conditions in women and people of color, echoing longstanding gaps in clinical research. The conversation drills into how training data from electronic health records and medical literature skews toward white male presentations, leading chatbots to miss heart attacks in women or downplay pain in Black…
AI in Court Sentencing: Algorithms That Decide Your Fate
In this episode of AI Ethics with Fexingo, Lucas and Luna explore the growing use of AI risk assessment tools in courtroom sentencing. They focus on the case of Paul Zilly, a man whose sentence was influenced by an algorithm's recidivism score, and the controversy surrounding COMPAS, the tool developed by Northpointe. The conversation unpacks how these algorithms learn bias from historical crime data, the impossibility of perfect fairness when the data itself is biased, and the profound…
AI Climate Models: Bias in the Simulation Itself
Ahead of the COP39 climate summit in November 2026, Lucas and Luna dig into a quiet but high-stakes corner of AI ethics: the simulation models governments use to set carbon budgets and adaptation policy. They focus on the ACCESS-OM2 ocean model and the CMIP6 ensemble, showing how a 2014 training-data quirk — a small patch of missing Atlantic salinity — silently skewed sea-level projections for the North Atlantic coast. They trace the same structural bias into AI downscaling tools like the…
Why AI Language Models Stutter on Cursive Handwriting
When an AI system reads handwritten doctor's notes, it doesn't just struggle with messy penmanship — it often fails on cursive scripts, particularly those used by older generations. This episode digs into the surprising reason: most training data for handwriting recognition comes from printed letters and neatly typed forms, leaving cursive scripts drastically underrepresented. We look at how models like Google's handwriting recognition and the UK's NHS digitisation project handle this…
How AI Admissions Algorithms Learn Socioeconomic Bias
A University of Texas study found that an AI admissions algorithm penalized applicants from lower-income zip codes at 2.7 times the rate of affluent areas. Lucas and Luna unpack how training data from legacy admissions and school rankings can encode class bias, why state law sometimes prevents inspecting the model, and whether fairness metrics can fix a fundamentally unequal dataset. #AIEthics #AlgorithmicBias #UniversityAdmissions #HigherEdAI #SocioeconomicBias #FairnessMetrics #ResponsibleAI…
How AI Video Interviews Learn Bias from Candidate Behaviour
This episode of AI Ethics with Fexingo examines how AI-powered video interview platforms analyze facial expressions, tone, and word choice to screen job candidates—and the hidden biases those models absorb. We look at a 2025 study from the AI Now Institute that found a leading platform penalized candidates who paused before answering, a behaviour more common in certain cultures. Lucas and Luna discuss the training data problem, the risk of self-fulfilling feedback loops, and what responsible…
How Predictive Policing Algorithms Learn Racial Bias
In episode 138 of AI Ethics with Fexingo, Lucas and Luna examine how predictive policing algorithms amplify systemic racial bias through a dangerous feedback loop. They break down a 2024 University of Chicago study that found models trained on arrest data predict crime hotspots in majority-Black neighborhoods at rates 40% higher than models using victim reports. The conversation covers the recent cancellation of Geolitica's contract in Los Angeles, the difference between arrest data and calls…
How AI Voice Models Fail Non-Native Speakers
Voice assistants like Amazon Alexa, Apple Siri, and Google Assistant are everywhere, but they consistently struggle with non-standard accents. This episode examines a 2025 study from Stanford's Center for Applied AI that found error rates for Indian English speakers are over 30% higher than for General American English. We break down why this happens—from training data demographics to tokenization choices—and what companies are doing to fix it. We also discuss the ethical implications: when…
How Facial Recognition AI Learns Racial Bias
In this episode, Lucas and Luna dive into the persistent problem of racial bias in facial recognition systems. Starting with a 2019 NIST study that found error rates up to 100 times higher for Black and Asian faces, they explore how training datasets like Labeled Faces in the Wild are overwhelmingly white, leading to biased outcomes. They discuss real-world consequences—false arrests, airport delays—and examine mitigation strategies like synthetic data and data augmentation, while questioning…
How AI Recommendation Algorithms Learn Bias from User Behaviour
In this episode of AI Ethics with Fexingo, Lucas and Luna explore how recommendation algorithms inherit bias from user behavior patterns. They anchor the discussion on a 2025 study from the Algorithmic Justice League, which found that YouTube's recommendation engine served 30% more political and conspiracy content to male user profiles than female ones. They dig into the feedback loops that amplify gender, race, and class biases, and discuss what platforms like YouTube are doing—and failing to…
How Self-Driving Cars Inherit Bias from Training Data
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…
Showing the latest 50 episodes. The full archive of 183 is on Apple Podcasts, Spotify and every major podcast app — or via the RSS feed above.