What Now? With Trevor NoahHilke Schellmann: Is the Algorithm Hiring the Wrong People?
CHAPTERS
- 0:04 – 1:08
Confidence vs. competence: why interviews reward the wrong traits
Hilke Schellmann explains how traditional interviews often select for confidence rather than job ability. The group connects this to gendered dynamics, where confidence is socially rewarded more in men, creating systematic mis-hires and inequities.
- •Interviews can filter for people who talk well rather than do well
- •"Competence versus confidence" as a structural hiring flaw
- •Gender dynamics: confidence is often read as competence, especially for men
- •Bias is embedded in familiar, "normal" hiring rituals
- 1:08 – 3:27
Greenpoint, gentrification, and the odd etiquette of learning 'I don’t speak your language'
Trevor, Eugene, and Hilke riff on where Hilke lives and how neighborhoods change over time. The conversation detours into why people learn the phrase “I don’t speak your language” in that language, framing a light, human lead-in before the AI topic deepens.
- •Hilke’s long-term apartment in Greenpoint and neighborhood change
- •Polish-language interactions and Hilke’s one memorized phrase
- •Trevor’s skepticism about the politeness logic
- •Humor as a bridge into heavier tech-and-society themes
- 3:27 – 5:29
An investigative journalist who now investigates AI—and builds tools
Hilke introduces her work: investigating AI’s real-world impact, especially in employment, and thinking about AI’s effect on journalism and factual society. Trevor probes why AI became the dominant focus of her reporting.
- •Hilke’s shift from broad investigations to AI as a primary beat
- •Focus on societal implications: winners, losers, and power
- •AI’s impact on journalism and information integrity
- •Curiosity-driven approach: “What is going on here?”
- 5:29 – 7:27
The Lyft-ride origin story: the first 'robot interview' rabbit hole
Hilke recounts the moment that hooked her: a Lyft driver describing a job interview conducted by a robot. She describes discovering a hidden ecosystem of HR tech vendors and realizing how widespread algorithmic hiring already was.
- •2017 story: baggage handler interview via automated system
- •Discovery of HR tech conferences and vendor ecosystem
- •AI in hiring was widespread but largely invisible to the public
- •Framing: transformative tech requires scrutiny beyond engineering
- 7:27 – 19:15
Debunking emotion AI: your face and voice don’t predict job performance
Hilke describes early vendor claims about emotion recognition in hiring—facial analysis, voice intonation, and “emotion scanning.” She explains how verification with experts undermined those claims and why such signals are unreliable in high-stakes selection.
- •Vendors claimed facial/voice cues predict job success
- •Experts: emotion-from-face and intonation are weak/invalid for hiring
- •Job interviews create nervous behaviors misread as "happiness" or traits
- •Journalistic method: trust-but-verify exposes shaky scientific claims
- 19:15 – 21:22
How much hiring is AI-driven? The funnel, mass rejection, and invisible screening
The group breaks down how AI is used at the top of the hiring funnel to process huge volumes of applicants. Hilke explains why concrete statistics are hard to obtain, but describes common practices across large platforms and Fortune 500 employers.
- •No central registry: AI use is inferred from surveys and reporting
- •High-volume hiring drives automation (millions of applications)
- •AI is often used early to reject and narrow applicant pools
- •Efficiency incentives override questions about accuracy and fairness
- 21:22 – 23:29
One-way video interviews: talking to a screen that doesn’t talk back
Hilke explains one-way video/audio interviews: applicants record answers without a live interviewer, sometimes under tight time windows. She details how companies may use AI to rank these recordings, especially in high-volume roles.
- •Applicants receive a link and record answers solo
- •Often used for entry-level/high-volume/high-turnover roles
- •Some companies have humans review; others algorithmically rank
- •Candidates frequently assume a human will watch—often not true
- 23:29 – 27:06
Humans are biased at hiring—and AI may be too, without proof it improves outcomes
Trevor raises the paradox: humans are biased and inconsistent, but AI may not be demonstrably better. Hilke notes companies rarely validate whether AI-selected candidates truly perform better, while leaders acknowledge qualified candidates are rejected.
- •Hiring is hard to do fairly at scale; humans cut corners too
- •Hilke sought audits comparing AI vs traditional methods—companies refused
- •Survey: leadership admits tools reject qualified candidates (~90%)
- •Core issue: efficiency is measurable; hiring quality often isn’t validated
- 27:06 – 31:16
AI vs. AI hiring: applicants use AI to beat the employer’s AI
With the rise of LLMs, applicants can generate resumes and even automate applications at scale. The conversation highlights the arms race dynamic and questions what hiring processes mean if both sides are optimizing for the algorithm rather than the job.
- •LLMs make it easy to mass-generate resumes and cover letters
- •New tools can auto-apply on a candidate’s behalf
- •Arms race: applicants optimize to pass filters; employers optimize to filter
- •Bigger question: if AI hires AI-assisted applicants, what’s the point?
- 31:16 – 37:21
The Amazon keyword trap: when 'women' gets you downgraded
Hilke and Trevor walk through the Amazon-style failure mode: training on past “successful” resumes can encode existing gender disparity. The system learns patterns correlated with incumbents, punishing words like “women’s” even when irrelevant to qualifications.
- •Model trained on historical resumes replicates past disparities
- •Keyword correlations become proxies for gender (e.g., “women’s soccer”)
- •Pattern matching substitutes for job-relevant criteria
- •Unsupervised pattern discovery can yield discriminatory scoring rules
- 37:21 – 50:20
Weird assessments: balloon-popping games, space-bar tests, and personality claims
Hilke describes gamified assessments used by large multinational companies and how opaque they can feel to candidates. The hosts unpack how trivial actions are converted into speculative personality inferences, raising validity and fairness concerns.
- •Candidates are pressured to comply (e.g., 48-hour windows)
- •Games claim to infer traits like risk tolerance or challenge-seeking
- •Space-bar speed and timing can be recorded and analyzed
- •Personality is a weak predictor; people can grow and adapt over time
- 50:20 – 1:02:37
Bias at scale: zip codes, commutes, redlining, and plausible deniability
The discussion expands from keywords to broader statistical shortcuts like commute distance and zip codes. Hilke explains how seemingly neutral variables can reproduce segregation and discrimination, while AI provides cover for companies to deny intent.
- •Commute distance predicts quitting statistically—but isn’t job capability
- •Zip-code screening can mirror historical segregation/redlining
- •AI adds a veneer of objectivity that reduces scrutiny
- •Opacity creates plausible deniability and limits legal accountability
- 1:02:37 – 1:12:14
Beyond hiring: workplace surveillance, ‘productivity theater,’ and algorithmic firing risk
The conversation shifts to staying employed: AI-powered monitoring proliferated during remote work and continues across offices and warehouses. Hilke explains how metrics like emails, meetings, and bathroom breaks drive ‘productivity theater’ and can trigger punitive decisions.
- •Remote work accelerated surveillance and behavioral monitoring
- •Digital traces: emails, Zoom behavior, Slack activity, print volume
- •‘Productivity theater’ emerges: perform busyness to satisfy metrics
- •Warehouse-style productivity algorithms and bathroom-break monitoring risks
- 1:12:14 – 1:20:13
What can be done: privacy, worker co-determination, transparency, and real audits
Hilke outlines actionable levers for lawmakers, companies, and workers. She contrasts Germany’s worker councils and co-determination with US workplace monitoring norms, argues for transparency mandates, and highlights EU-style data access as a tool for accountability.
- •Policy: stronger privacy protections and disclosure of monitoring tools
- •Workplace governance: worker councils/co-decision on surveillance tech
- •Company practice: demand evidence, audits, and validity testing from vendors
- •Worker leverage: data-access rights (EU example) enable challenges/settlements
