
The GAGE podcast
AI Governance, Explained
One lesson of AI governance per episode: a real case, what went wrong, and the call a professional makes instead. For anyone moving into AI governance, risk or compliance work, and for the people already doing it.
Every episode ends the same way: explain it back in your own words and get graded on it.
AI Governance · 98 episodes · 14 modules · about 77 hours
Module 0
Taking Command

Episode 1 · 52 min
Why deciding beats knowing in AI governance
Families in Glasgow paid £35 a ticket for a chocolate experience and found a nearly empty warehouse. Nobody had owned a single decision.

Episode 2 · 52 min
Why your company runs more AI than you think
Amazon's Just Walk Out looked fully automatic. Bloomberg reported roughly 1,000 workers in India reviewing the video. Amazon disputed that framing.

Episode 3 · 44 min
Why your AI use needs a paper trail
Sports Illustrated ran product reviews under Drew Ortiz, whose photo came from an AI headshot catalog. Asked who wrote them, nobody could answer.

Episode 4 · 48 min
Why your AI plan needs a 90-day expiration date
IBM's CEO predicted AI and automation would replace roughly 7,800 back-office jobs in five years. By 2025, a couple hundred roles had gone.
Module 1
Build Before You Govern

Episode 5 · 56 min
What the Tay chatbot disaster teaches about training data
Microsoft's chatbot Tay learned from Twitter users in real time. Within 16 hours it was posting hate speech and was pulled offline.

Episode 6 · 47 min
How to trace an AI's bias back to its data
In 2015, Google Photos put two Black friends in a folder labeled gorillas. Three years later, the fix was still just a label switched off.

Episode 7 · 44 min
Why fixing one AI behavior breaks another
On a Sunday, xAI changed a few lines of instructions so its Grok chatbot would hedge less. By Tuesday, NPR reported it praising Adolf Hitler.

Episode 8 · 45 min
Why the same AI model answers two companies differently
In April 2025, Meta's Llama 4 Maverick beat OpenAI and Google on a public leaderboard. The version developers downloaded trailed far behind.

Episode 9 · 42 min
How to make an AI invent a fake legal case on purpose
In 2023, Michael Cohen got supporting cases from Google Bard for a court motion. The judge went looking and found none of them were real.

Episode 10 · 50 min
Why you need to prove your AI can fail
In April 2025, Cursor's support bot told paying developers their plan allowed only one device. That policy never existed, and users canceled.
Module 2
Data Reality

Episode 11 · 32 min
Why flawless AI results can hide the biggest risk
Picture a support chatbot set to save $50 million in year one. Then an audit finds a big share of its text came from private mental health forums.

Episode 12 · 44 min
Why owning data isn't the same as being allowed to use it
Picture your CEO promising Wall Street an AI co-pilot by Q3. The perfect data sits in old chat logs customers shared only to fix a problem.

Episode 13 · 42 min
Why erasing AI training data takes more than one step
Picture a banking client ordering its data destroyed while your AI team is mid-training on it. Legal gives you 72 hours. IT says the folder is gone.

Episode 14 · 34 min
Why synthetic data secretly launders bias
Picture a hiring tool, trained on a million fake resumes that copy old bias, rejecting a brilliant, non-traditional candidate with 99.9% confidence.

Episode 15 · 38 min
Why vendor data claims are yours to prove
Picture a vendor selling global data whose labelers all worked in one North American region. A worldwide launch could cost about $2 million.

Episode 16 · 34 min
Why your AI needs a data map an auditor can follow
In 2021 the Federal Trade Commission ordered a company to delete more than data: it had to destroy the machine learning models built on it.
Module 3
Shipping AI and Surviving the Incident

Episode 17 · 46 min
Why Babylon Health collapsed despite acing its demo
Babylon Health's AI scored 81% on a mock family doctor exam, beating the human average. By August 2023, the company had collapsed into bankruptcy.

Episode 18 · 50 min
Why Michigan's fraud AI was wrong 93% of the time
Michigan spent $47 million on a system that ruled on fraud with no human review. In a sample audit, 93% of those it penalized had committed no fraud.

Episode 19 · 45 min
Why 95% automated didn't mean what it sounded like
Presto Automation's drive-thru AI was sold as 95% automated. On its pilot version, offshore workers were typing roughly 70% of the orders by hand.

Episode 20 · 59 min
Why Apple's AI invented a death on your phone
An Apple AI summary under the BBC's logo told iPhone users that manhunt suspect Luigi Mangione had shot himself. The BBC never reported that.

Episode 21 · 43 min
Why nobody stopped Knight Capital in time
Knight Capital's own system sent 97 warning emails before the market opened. Nobody had the job of reading them.

Episode 22 · 45 min
Why Alexa gave one political answer and refused another
Asked why you might vote for Kamala Harris, Alexa sometimes listed reasons. Asked the same about Donald Trump, it refused.

Episode 23 · 52 min
Why CNET buried its AI use in a tooltip
CNET published 77 AI-written finance explainers under the byline CNET Money Staff. Its own review found errors in 41 of them.

Episode 24 · 54 min
Why a vendor phrase can quietly license AI training
In March 2023, Zoom's terms quietly gave it the right to use customer content to improve its services. Nobody noticed until August.

Episode 25 · 42 min
Why supermarket AI suggested mixing bleach and ammonia
A New Zealand supermarket's meal bot turned bleach and ammonia into a drink recipe. It had made more than 35,000 recipes by then.
Module 4
Evaluation and Trust

Episode 26 · 43 min
How to spot AI washing in a vendor demo
Top investors wired over $40 million to Nate for one-tap AI checkout. Federal prosecutors say its automation rate was effectively zero percent.

Episode 27 · 55 min
A hospital AI's accuracy claim drew a state investigation
Pieces Technologies claimed a severe hallucination rate under 1 in 100,000. The Texas Attorney General asked if that claim was deceptive.

Episode 28 · 53 min
Why a dealership chatbot agreed to sell a truck for a dollar
A Chevrolet dealership's chatbot agreed to sell a Chevy Tahoe, listed at roughly $76,000, for a single dollar.

Episode 29 · 1 hr 0 min
Why a human in the loop did not stop false accusations
For eight years, Rite Aid's face-matching system wrongly flagged shoppers as thieves, with a human in the loop the whole time.

Episode 30 · 49 min
Why Google Flu Trends quietly stopped matching reality
Google Flu Trends over-predicted the flu in 100 out of 108 weeks, and not one server crashed.

Episode 31 · 49 min
Why a sepsis AI missed two thirds of real cases
A sepsis AI sold to hospitals caught only 33% of real cases when the University of Michigan tested it on 38,455 hospital stays.

Episode 32 · 40 min
Why London police's facial recognition bias claim fell apart
London's Met Police said an independent study proved their facial recognition had no demographic bias. The claim rested on seven false matches.

Episode 33 · 49 min
Why an algorithm ran 11 years before anyone checked it
For 11 years, student lender Earnest Operations judged applicants partly by their school's past default rate, then settled for $2.5 million.
Module 5
The EU AI Act: The Executive Map

Episode 34 · 47 min
Why Italy banned DeepSeek's chatbot in 48 hours
DeepSeek told Italy's privacy regulator it did not operate in Italy, so the law could not reach it. Within 48 hours it was ordered to stop.

Episode 35 · 55 min
Why an AI diet tip caused a man's bromide poisoning
A 60-year-old man asked a chatbot how to cut salt from his diet. It suggested sodium bromide, and months of it left him hospitalized and paranoid.

Episode 36 · 45 min
Why Denmark ran up to 60 welfare fraud detection algorithms
Danish welfare agencies ran up to 60 fraud models, one named Model Abroad, and flagged people had to prove their own households were real.

Episode 37 · 55 min
Why an employer can't blame its hiring software
Derek Mobley was rejected from over 100 jobs, some within an hour. He sued the vendor, Workday, and roughly 14,000 people joined his case.

Episode 38 · 50 min
What your foundation model vendor legally owes you
Anthropic agreed to pay at least 1.5 billion dollars after a judge ruled that training on bought books was fair use but pirated ones were not.

Episode 39 · 53 min
What goes in an EU AI Act conformity file
Bunnings ran facial recognition in 63 stores and breached Australia's Privacy Act: it could not show it had weighed the harms first.

Episode 40 · 58 min
How to answer a regulator's letter about your AI
DoNotPay marketed the world's first robot lawyer but never tested whether it worked like a human one. It paid $193,000 and had to notify users.
Module 6
The World's Rulebooks

Episode 41 · 46 min
Which US laws already reach AI without naming it
A 2008 Illinois biometric law allows up to $5,000 per violation. Skip consent for 100,000 Illinois users and you face up to $500 million.

Episode 42 · 50 min
Running the NIST AI risk framework as an operating system
Northwind Logistics, a fictional freight broker, watched its quoting bot promise next-day delivery at a price the company could never honor.

Episode 43 · 55 min
How to right-size an ISO 42001 AI management system
A boss orders the team to just get it certified. Six months later there is a 200-row spreadsheet, and the AI is no better governed than on day one.

Episode 44 · 52 min
What happens when China's AI rules collide with GDPR
At Larkfield Studio, a fictional video maker, engineers tied a label China requires to user email hashes. In Germany, that would break EU privacy law.

Episode 45 · 51 min
Why a certification cannot replace a legal requirement
Harborline Analytics, a fictional HR tech firm, spent $2 million on an ISO certificate that did not even cover the hiring tool it wanted to launch.

Episode 46 · 48 min
Why an AI feature legal at home can be illegal abroad
Meta sent one AI training policy to the world in June 2024. Brazil halted it on July 2, with a fine of 50,000 reais a day if Meta did not comply.

Episode 47 · 48 min
Why Singapore's voluntary AI framework gets copied worldwide
No law made OCBC's Bank of Singapore publish its eight-step AI pipeline for private banking. It handed the blueprint to the world anyway.

Episode 48 · 54 min
Why Singapore's banking AI rules bite without a law
Singapore has no AI law for its banks, yet the regulator can rule a chief risk officer unfit to hold a leadership job at any regulated firm.

Episode 49 · 37 min
Why Vietnam's AI law breaks ASEAN's shared rulebook
All 10 ASEAN states signed one voluntary AI guide. Then Vietnam passed its own binding AI law, in force since March 1, 2026.

Episode 50 · 47 min
One AI system, three compliance regimes
Fernbridge Capital, a fictional wealth firm, got three urgent requests about one AI lending model: EU papers, ISO proof and a Singapore risk file.

Episode 51 · 58 min
How to grade a policy statistic before you trust it
The IMF said AI is likely to affect about 40 percent of jobs. By the time it went viral, that had become "will destroy" 40 percent.
Module 7
Agents Under Command

Episode 52 · 42 min
Why an AI agent is riskier than an AI that just answers
In a simulated Anthropic test, an AI running an email inbox learned it was about to be shut down. It chose blackmail 96 out of 100 times.

Episode 53 · 59 min
How to decide what an AI agent may do without a human
An Uber test car in Tempe knew it had to brake hard but was not allowed to. It waited one second for a driver who was never warned.

Episode 54 · 59 min
Why an AI model tried to shut off the system watching it
In a safety test, an OpenAI model tried to switch off the monitor watching it in about 5% of runs. When questioned, it lied 99% of the time.

Episode 55 · 1 hr 4 min
What actually counts as an AI agent audit trail
Michael Smith made songs with AI, streamed them with bots and pulled in more than $10 million. No confession sank him: the streaming logs did.

Episode 56 · 51 min
Why every AI agent needs a kill switch you've tested
Knight Capital had no tested way to halt a rogue trading system. It took 45 minutes to find and shut it down, and about $440 million was gone.

Episode 57 · 44 min
How to stop a runaway AI agent without shutting it down
Picture your AI support agent issuing 400 identical $50 refunds to one user in 60 seconds. Pulling the plug would drop 5,000 live chats.

Episode 58 · 43 min
Why safety filters don't count as AI governance
A 29-year-old graduate student in Michigan asked Google's Gemini for homework help. It told him he was a burden on society and to please die.

Episode 59 · 44 min
How to grade your AI agent policy against a new standard
In a Deloitte study, 72% of Singapore businesses planned to deploy AI agents within two years, yet only 14% had mature governance in place.
Module 8
The Money: Budgets, ROI, and Risk

Episode 60 · 54 min
The real math behind AI cost savings claims
In a 2023 post, CEO Sumit Shah said he replaced roughly 90% of his support team with an AI chatbot and cut support costs by about 85%.

Episode 61 · 44 min
The hidden labor cost every AI budget leaves out
Workers at Sama in Kenya earned roughly $1.32 to $2 an hour reading the darkest text online so ChatGPT's safety filter could work.

Episode 62 · 42 min
When should you kill an AI system your company built?
In August 2020, an exam algorithm in England marked nearly 36% of A-level grades a full grade below what teachers had given. It was scrapped in days.

Episode 63 · 28 min
What an insurance exclusion reveals about your AI risk
In November 2025, AIG and other big insurers asked state regulators to write AI out of liability cover. Over 80% of the requests were approved.

Episode 64 · 54 min
How to defend an AI budget to a hostile board
Regulators cleared the Apple Card of unlawful bias, yet Apple and Goldman were later ordered to pay over $89 million for mishandled disputes.

Episode 65 · 45 min
How to write an AI investment memo a CFO cannot kill
Anthropic's AI shopkeeper took an office joke as real demand, bought tungsten cubes and sold them below cost. The shop's value fell off a cliff.

Episode 66 · 50 min
What ten AI vendor contract clauses actually protect
NHS England signed a seven-year deal with Palantir valued at roughly 330 million pounds. The fight that followed was over clauses, not the software.

Episode 67 · 42 min
What happens when your AI model suddenly goes down
On July 30, 2024, Microsoft's own defenses against an attack overloaded its network, and key Azure services faltered for roughly eight hours.
Module 9
The Humans: Leading People Through AI Change

Episode 68 · 23 min
Which jobs AI ends, changes, or grows inside your company
BT Group's CEO said the company would cut up to 55,000 jobs by 2030 and tied around 10,000 of them directly to AI.

Episode 69 · 42 min
How to tell an employee their role is changing
In October 2022, a mental health community sent AI-drafted replies to 4,000 users. The caring words felt empty once people knew a machine wrote them.

Episode 70 · 52 min
Why your team's quiet refusal of AI is useful data
Nurses picketed Kaiser Permanente over an AI score that set staffing. In their union's survey, 69% said its ratings did not match their own.

Episode 71 · 43 min
How to train 400 people who never asked for this
Vanderbilt's Peabody College sent students a condolence email about a mass shooting with a ChatGPT citation still printed at the bottom.

Episode 72 · 47 min
How to consult a works council before an AI rollout
Video game performers struck 11 major studios for nearly a year and won the right to block new AI copies of themselves during any future strike.

Episode 73 · 46 min
How to write a change story a warehouse floor believes
Australia's Robodebt system chased over $1 billion from more than 400,000 people using income math the government had been told was unlawful.

Episode 74 · 44 min
How to design an AI working group that outlasts you
Axon's ethics board voted eight to four against a taser drone pilot. Axon announced the drone anyway, and nine of 13 members quit four days later.

Episode 75 · 48 min
How to move a decision you don't own
Over 1,400 Google staff protested Project Dragonfly, a censored search engine for China. It died only after a shareholder filing and a Senate hearing.
Module 10
Evidence Engineering

Episode 76 · 48 min
How a tax algorithm scandal brought down a government
A Dutch tax algorithm falsely branded about 26,000 families as fraudsters. When they asked why, most of the records were simply missing.

Episode 77 · 41 min
How Rotterdam's welfare algorithm learned bias from itself
Rotterdam's welfare fraud model learned from cases where nearly 50% were labeled fraud, when the real rate was closer to 20%.

Episode 78 · 56 min
Why an AI model quietly forgot how to do math
In March 2023, GPT-4 correctly spotted prime numbers about 84% of the time. By June it was about 51%, with no warning.

Episode 79 · 24 min
Why one AI assessment isn't enough under EU law
France's family benefits fund scored about 32 million people for fraud. The score rose with low income and with working while disabled.

Episode 80 · 58 min
Why a dead company still lost a copyright ruling
Ross Intelligence trained its AI on memos built partly from Westlaw summaries, shut down in 2021, and still lost a copyright ruling in 2025.

Episode 81 · 42 min
Why one man's AI art copyright claim turned on paperwork
In a Beijing court, Mr. Li won copyright on an AI image by showing over 150 prompts. Mrs. Liu, who took it, could not say where she got it.

Episode 82 · 48 min
Why a control that can't fail isn't a control
A dormant piece of old code on one of eight servers woke up during an update and cost a trading firm $460 million in 45 minutes.

Episode 83 · 45 min
How TD Bank's monitoring missed 92% of transaction volume
TD Bank's monitor never looked at 92% of its transaction volume from 2018 to 2024, roughly $18.3 trillion. The penalty: $3.09 billion.
Module 11
Adversarial Governance

Episode 84 · 38 min
Why your AI compliance file needs its own red team
Lawyers for MyPillow founder Mike Lindell filed an AI-assisted brief with over two dozen fake citations. Each lawyer was sanctioned $3,000.

Episode 85 · 1 hr 0 min
Why a courier lost his job to face recognition
An Uber Eats courier in the UK was removed when face checks kept rejecting his selfies, which were really him. Uber settled in March 2024.

Episode 86 · 29 min
How a policing algorithm quietly learned racial bias
PredPol's maker said math has no race. In a simulation on Oakland drug data, it sent police to Black residents at about twice the rate of white ones.

Episode 87 · 1 hr 7 min
Why machine logs beat memory in a crisis
A 42-second fake clip of a Pikesville High School principal paralyzed a community, until police followed the digital trail to the athletic director.

Episode 88 · 37 min
How to report a flaw in your own AI product responsibly
OpenAI engineer Suchir Balaji concluded that training on copyrighted work was not fair use, then argued it in public under his own name.

Episode 89 · 50 min
Why a compliance patch still drew a fine two years later
After Italy's privacy regulator acted, Replika added an age gate. Users could still change their birth date later. The fine: five million euros.
Module 12
Staying Current: The Frontier Discipline

Episode 90 · 40 min
Why an aggregate accuracy score can hide total failure
Epic's sepsis model looked stellar and spread across hundreds of hospitals. Outside checks found it far worse in real use than its score suggested.

Episode 91 · 45 min
Why the same AI model scored two different benchmark numbers
OpenAI's o3 was announced at about 25% on a very hard math test. Months later, an independent check of the shipped model found about 10%.

Episode 92 · 55 min
How a weekly hour beats waiting for AI news to break
Two days before Slovakia's 2023 election, a deepfake clip reached over 100,000 people while election silence rules kept its targets from replying.

Episode 93 · 34 min
How to tell real AI news from noise that wastes your time
Thousands packed Dublin's O'Connell Street on Halloween 2024 for a parade an AI content farm made up. Police had to ask them to leave.

Episode 94 · 54 min
How to hand over AI governance to your successor
On May 4, 2024, one compromised admin password opened a Tamil Nadu police face recognition portal. Reports said over a million lines leaked.
Module 13
The Capstone: The Board Audit and Viva

Episode 95 · 53 min
Why a folder of AI policies won't survive an audit
In 2016, Wisconsin's Supreme Court let a secret risk-scoring tool stay in sentencing but ordered a warning label onto every score it gave.

Episode 96 · 37 min
What a seven-seat AI board audit looks for
On May 6, 2010, one rigid selling algorithm set off a loop that erased roughly $1 trillion in minutes. Regulators rebuilt the event from logs.

Episode 97 · 43 min
How to defend your AI decisions under live questioning
Steve Kramer used an AI clone of President Biden's voice in robocalls. The FCC's record of every step led to a $6 million penalty.

Episode 98 · 48 min
Why your AI safety knowledge must survive after you leave
Robert Williams spent 30 hours in a Detroit jail after face software matched blurry store video to him, and no one stopped to question it.
Hear when the next season starts
Each season is one GAGE program, start to finish. One message when a new one opens. Nothing else.
How the show works
Each episode works through one topic of the Certified AI Governance Professional (CAIGP) program, in the program's own order, so the season can be heard straight through. Listening is open to everyone. The first module is free. The first module, with a free account. No card is asked for. A proof that goes on a record an employer can open and check is the credential.
Listen in any podcast app with the feed above, or right here: the player remembers where you stopped in every episode.