TThe Diary of a CEO
← All episodes
26 March 2026

AI Whistleblower: We Are Being Gaslit By The AI Companies! They’re Hiding The Truth About AI

3Frameworks
15Insights

Frameworks in this episode

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 2

Myth Buster20:30

The brain analogy behind scaling is still a hypothesis

Hao says some influential AI researchers treat the brain as a statistical engine and infer that larger statistical models could approach or exceed human intelligence. She stresses that this is a scientific hypothesis associated with particular researchers, not an established consensus across neuroscience and psychology. In her view, the distinction matters because global investment and infrastructure decisions follow from it.

  • Hao attributes the statistical-brain hypothesis to researchers including Ilya Sutskever and Geoffrey Hinton
  • The hypothesis supports a belief that greater model scale can yield greater intelligence
  • Hao says scientists outside AI debate this account of the brain
  • She argues that the hypothesis drives data collection and data-center expansion
  • The segment does not establish which theory of human intelligence is correct

It's a hypothesis that they have. It's not one that has been proven by science.

Karen Hao · 21:00

these companies they are driving their future actions based on this hypothesis

Karen Hao · 23:30
#scaling#intelligence#neuroscience#ai-research
Myth Buster1:08:00

Why scaling does not improve every AI capability equally

Hao rejects the idea that a model becomes uniformly more capable simply because it is scaled. She describes a jagged frontier: firms select capabilities, gather relevant data, and hire people to train those tasks, often prioritizing commercially valuable domains. Bartlett argues that useful outputs may matter more than resemblance to human intelligence, leaving a genuine disagreement about how much the mechanism matters.

  • Models can be strong at some tasks and weak at others
  • Hao says companies deliberately choose which capabilities to improve
  • Training requires capability-specific data and human work
  • She names finance, law, medicine, health care, and commerce as commercial priorities
  • Bartlett emphasizes practical outputs such as driving and surgery
  • The speakers do not resolve whether future systems will become broadly intelligent

they actually can only do some things for some people

Karen Hao · 1:09:00

They pick what capabilities they want to advance.

Karen Hao · 1:11:00
#ai-capabilities#scaling#training-data#intelligence

Hot Take· 1

Hot Take1:29:00

Bartlett's three role pools he currently sees as hard to replace

Bartlett describes how AI-agent experiments are changing his hiring decisions. He currently prioritizes deep domain experts who can direct agents, highly curious people who are proficient with agents, and people with strong in-person relationship skills. He later speculates that even these pools could face pressure if capability improves, so this is his present operating view rather than a durable labor forecast.

  • Deep experts can formulate and judge specialist work delegated to agents
  • Agent-proficient, highly curious workers can create business leverage
  • In-person relationship and community skills remain valuable in Bartlett's companies
  • The view comes from Bartlett's own companies and investment portfolio
  • He does not claim the three groups are permanently protected from automation

really deep expertise is very very valuable

Steven Bartlett · 1:30:00

people with extremely great IRL people skills

Steven Bartlett · 1:31:00
#hiring#ai-agents#expertise#human-skills

Explainer· 7

Explainer07:00

How Hao says AGI changes meaning with the audience

Hao argues that artificial general intelligence lacks an agreed scientific destination and that OpenAI has described it differently to lawmakers, consumers, Microsoft, and the public. She presents those shifting definitions as a way to mobilize regulation, adoption, or capital. The segment reflects Hao's interpretation of company rhetoric rather than a settled account of intent.

  • Hao says there is no scientific consensus defining human intelligence
  • She argues that an undefined destination lets companies redefine AGI
  • She cites promises about social problems, digital assistance, revenue, and economically valuable work
  • Different formulations may serve different audiences in her analysis

There are no goalposts for this field

Karen Hao · 08:00

This is like not a coherent vision of one technology.

Karen Hao · 09:30
#agi#definitions#openai#rhetoric
Explainer15:00

Why Hao says views of Sam Altman split so sharply

Hao says interviewees tended to see Altman either as an exceptional technology leader or as manipulative and dishonest. Her explanation is alignment: people who share his preferred future value his persuasion and ability to mobilize resources, while opponents feel those abilities are being used against their own goals. This is a pattern Hao inferred from her interviews, not a psychological diagnosis.

  • Hao reports unusually polarized opinions among her sources
  • Supporters value Altman's storytelling, fundraising, and recruiting abilities
  • Critics may experience the same abilities as manipulation
  • Hao uses Dario Amodei's departure from OpenAI as an example of a vision conflict

No one has in between feelings about him.

Karen Hao · 15:30

it really comes down to what that person's vision of the future is

Karen Hao · 15:30
#sam-altman#persuasion#leadership#polarization
Explainer33:30

How access can pressure technology journalism

Hao and Bartlett describe access to prominent executives as a powerful incentive that companies can grant, delay, or withdraw. Hao says losing OpenAI access early in her career initially felt damaging but ultimately forced her to report through other sources. Bartlett adds his own claim that an unnamed AI figure's team has repeatedly dangled a possible appearance, which he interprets as an attempt to influence coverage.

  • Hao says access is a major incentive for technology journalists
  • She reports that OpenAI declined participation after criticizing an earlier profile
  • She says she still completed more than 300 interviews for her book
  • Bartlett presents his unnamed booking experience as his own interpretation
  • Both argue that withholding access can discourage critical voices

a really big carrot that they can give to technology journalists is access

Karen Hao · 38:30

They don't need to open the front door for me.

Karen Hao · 41:30
#journalism#access#media#openai
Explainer1:17:00

Why self-driving adoption is more than a capability problem

Hao argues that statistical driving systems can perform well in places for which they are trained yet fail to generalize uniformly across locations. She also identifies trust, legal responsibility, and infrastructure as barriers to widespread adoption. Bartlett challenges her with Tesla's safety and sales claims, but the episode does not provide comparable data sufficient to establish a universal safety advantage.

  • Training data is labeled for vehicles, pedestrians, lights, and lane markings
  • Hao says probabilistic systems cannot be made entirely error-free
  • She accepts that a trained system may outperform local human drivers in some places
  • She disputes that this result automatically transfers to every location
  • Public trust and responsibility after a fatal crash remain adoption questions
  • Bartlett's safety comparison is presented as a host claim, not established in the episode

It depends on whether the Tesla was trained to specifically navigate the place that you're driving.

Karen Hao · 1:19:00

part of it is also a legal problem

Karen Hao · 1:20:00
#self-driving#safety#autonomy#regulation
Explainer1:21:30

AI job loss comes from technology and executive choices

Hao rejects a binary choice between AI replacing every job and AI having no employment effect. She argues that losses come both from improving capabilities and from executives accepting cheaper, merely adequate systems or using AI rhetoric while downsizing. She also warns that new work may be lower quality and that removing entry and middle roles can break progression into senior expertise.

  • Some work is being automated as models gain selected capabilities
  • Executives can cut roles before a model fully matches human performance
  • Company overhiring or investor messaging can also shape how layoffs are described
  • Displaced professionals may move into lower-status data-annotation work
  • Entry-level and middle roles are rungs through which future experts develop
  • The episode offers examples and interpretations, not a complete causal estimate of labor-market change

There are definitely jobs that are being automated away because of the capabilities of their models.

Karen Hao · 1:24:00

it breaks the career ladder

Karen Hao · 1:28:00
#jobs#automation#career-ladder#executive-decisions
Explainer1:41:00

The hidden annotation labor behind AI systems

Hao explains data annotation as human work that supplies examples and labels used to train AI systems. Citing a New York Magazine report, she describes highly educated displaced workers waiting for short, unpredictable projects and racing to complete tasks before they disappear. The working-condition examples are Hao's summary of that reporting, not independently investigated within the episode.

  • Human examples helped turn language models into conversational systems
  • Annotation is one input into reinforcement-learning processes
  • Third-party firms compete to provide work quickly and cheaply
  • Hao says workers may have little control over when tasks appear or end
  • She argues that fragmented work can devalue expertise and reduce personal agency
  • The segment distinguishes product convenience from the labor conditions supporting it

data annotation is the process of teaching these chat bots or or any AI system

Karen Hao · 1:42:00

they are incentivized to pit workers against each other

Karen Hao · 1:44:00
#data-annotation#hidden-labor#working-conditions#ai-training
Explainer1:49:00

Hao's case that AI data-center costs fall locally

Hao argues that large AI facilities can increase electricity costs, strain grid reliability, compete for fresh water, and add local air pollution when powered by nearby fossil-fuel generation. She cites projects in Abilene, Louisiana, and Memphis, including her claim that xAI used 35 methane-gas turbines for its Colossus facility. The episode includes corrected size comparisons and no source documents, so the figures and health effects remain Hao's reported claims rather than independently verified facts here.

  • Hao corrects an initially displayed comparison before giving updated facility sizes
  • She says OpenAI's planned Abilene facility would use more than 20% of New York City's power
  • She says a Louisiana Meta facility would be about one-fifth the size of Manhattan and use half New York City's average power demand
  • Data centers may require water both for power generation and cooling
  • Hao reports that xAI used methane-gas turbines in a Memphis community
  • Her statements about worsened asthma, respiratory illness, and lung-cancer burden are claims in the interview, not clinical findings established by it

Power utility increases, grid reliability decreases.

Karen Hao · 1:51:30

they literally smelled what seemed like a gas leak in all of their living rooms

Karen Hao · 1:52:30
#data-centers#energy#water#public-health

Story· 2

Story10:00

Hao's account of how Altman and Musk became OpenAI rivals

Hao says Altman's 2015 existential-risk language closely mirrored Elon Musk's public concerns while Altman was recruiting him to co-found OpenAI. She then recounts lawsuit documents and her reporting about the later contest over leadership of a proposed for-profit entity. Her account is partly interpretation and reporting around an ongoing legal dispute, not a judicial finding that Altman manipulated Musk.

  • Musk co-founded OpenAI after publicly warning about existential AI risk
  • Hao interprets Altman's early language as tailored to persuade Musk
  • She says internal emails showed Ilya Sutskever and Greg Brockman initially favored Musk as chief executive
  • Hao reports that Altman persuaded Brockman that giving Musk control could be dangerous
  • Musk left after the leadership preference shifted toward Altman

certainly from Musk's perspective, he does feel manipulated

Karen Hao · 12:00

If I'm not CEO, I'm out.

Karen Hao · 15:00
#sam-altman#elon-musk#openai#leadership
Story41:30

Hao's reconstruction of the board decision to fire Altman

Drawing on roughly six or seven sources, Hao reconstructs concerns raised by Ilya Sutskever and Mira Murati about chaos, divided teams, and inconsistent representations under Altman's leadership. She says independent directors concluded that conduct which might not justify dismissal at an ordinary startup met a different threshold because they believed OpenAI could build transformative technology. The surprise firing excluded key stakeholders, triggering the campaign that restored Altman days later.

  • Sutskever reportedly approached independent director Helen Toner with concerns
  • Hao says Murati and Sutskever assembled messages and other documentation
  • The concerns included team conflict, instability, and inconsistencies between presentation and action
  • Directors reportedly moved quickly because they feared Altman's persuasive response
  • Microsoft and other stakeholders were largely excluded before the decision
  • The backlash led to Altman's reinstatement

the problem will not be fixed unless Altman is removed

Karen Hao · 44:30

every single person that is affected by this decision is now extremely angry

Karen Hao · 50:30
#openai#board-governance#sam-altman#corporate-crisis

Q&A· 2

Q&A1:15:30

Hao argues for radiologists using AI, not being replaced by it

When asked whether AI will surpass surgeons and radiologists, Hao points to research she says supports combining a radiologist's judgment with an AI tool. She claims this combination can improve the accuracy and timing of some cancer diagnoses. No study is identified in the episode, so the medical claim should be treated as Hao's research summary rather than verified clinical guidance.

  • Past predictions that radiologists would soon be unnecessary missed their stated deadline
  • Hao frames AI as an input into expert judgment
  • She claims the combination improves some diagnostic outcomes
  • The transcript does not name a study, cancer type, patient group, or effect size
  • The discussion does not support replacing individual medical advice with an AI system

the best outcomes for people in a healthcare setting is for the radiologist to have the AI model in their hands

Karen Hao · 1:16:00
#radiology#healthcare-ai#diagnosis#human-ai
Q&A1:35:30

Klarna's CEO clarifies its AI and headcount story

In a live call, Klarna chief executive Sebastian Siemiatkowski says AI increased customer-service speed and quality in the company's experience while human interaction became more valuable for premium service. He reports that headcount fell from about 6,000 to under 3,000 over two to three years as revenue doubled, primarily through limited recruitment and natural attrition rather than layoffs. These are the chief executive's claims in the episode and are not independently audited there.

  • Siemiatkowski says the company introduced AI early in customer service
  • He claims customers valued faster and better interactions
  • Klarna also increased emphasis on human service for higher-value interactions
  • He reports roughly halving headcount while doubling revenue
  • He attributes the reduction mainly to a hiring slowdown and natural attrition
  • The episode does not provide financial records or workforce data to verify the figures

AI has allowed us to be do more with less people

Sebastian Siemiatkowski · 1:36:30

we have avoided layoffs and instead relied on natural attrition

Sebastian Siemiatkowski · 1:37:00
#klarna#customer-service#headcount#ai-efficiency

Takeaway· 1

Takeaway54:30

Why former OpenAI leaders built competing AI companies

Hao observes that several prominent people recruited into OpenAI later left after conflicts and founded competing organizations. She interprets the pattern as powerful leaders seeking control over their own vision of AI rather than accepting another executive's direction. The episode lists examples but does not establish that every departure had the same cause.

  • Elon Musk later founded xAI
  • Dario Amodei founded Anthropic
  • Ilya Sutskever founded Safe Superintelligence
  • Mira Murati founded Thinking Machines Lab
  • Hao says competing visions and a desire for control connect these departures

They want to have control over their own vision of this technology.

Karen Hao · 57:00
#ai-companies#founders#openai#competition