Eric Schmidt on Google's $40 Billion AI Bet, the YouTube Mistake, and When Humanity Should Pull the Plug
Google's former CEO explains how a small team helped generate billions in profits, why Google failed to defeat YouTube and Facebook, how OpenAI surprised Silicon Valley, and the dangers that could force humanity to shut down advanced artificial intelligence.
Eric Schmidt spent a decade as Google's CEO, working alongside founders Larry Page and Sergey Brin as the company became one of the most powerful businesses in technology. His account of those years offers an unusually detailed look at how Google made decisions, selected investments, managed employees and sometimes failed to recognize opportunities sitting directly in front of it.
In a wide-ranging conversation with Steven Bartlett on The Diary of a CEO, Schmidt also addressed the arrival of artificial intelligence, a technology he considers potentially more transformative than anything he encountered during his career.
The interview took place in late 2024, shortly before the publication of Genesis, the book Schmidt wrote with former US Secretary of State Henry Kissinger and technology executive Craig Mundie.
Several of Schmidt's most revealing statements concern matters he witnessed personally: Google's internal allocation of resources, the decision to acquire YouTube, the breakdown of confidential employee meetings and the unexpected breakthrough that made ChatGPT successful.
His predictions about AI carry a different kind of significance. Schmidt warned about autonomous cyberattacks, dangerous model capabilities and AI systems that might improve themselves beyond human understanding. He also argued that the greatest danger could be humanity's failure to use AI quickly enough to improve education and healthcare.
Independent records and subsequent developments help distinguish which claims are established, which remain Schmidt's estimates, and which risks have begun to materialize.
1. Google's 70/20/10 Rule: How Small Experiments Produced Enormous Returns
The investment formula behind Google's growth
One of Schmidt's most concrete business revelations concerned Google's approach to allocating resources.
Rather than concentrating entirely on its existing search and advertising business, the company followed a framework developed by Page and Brin.
The allocation was straightforward: 70% toward core operations, 20% toward adjacent businesses and 10% toward new experiments.
Core operations meant search and advertising, Google's established sources of revenue. Adjacent opportunities included services such as cloud computing, which could benefit from Google's infrastructure and technical capabilities.
The remaining allocation supported ideas that had no immediate guarantee of becoming commercial products.
Google was already discussing this framework publicly in 2005. In a contemporaneous account of Schmidt's Business 2.0 interview, he described the allocation in terms of how employees spent their time, rather than a rigid division of corporate spending.
The distinction matters. It was a guide for organizational priorities, not a published financial statement proving that exactly 10% of Google's expenditure went into experimental research.
The Google Brain investment Schmidt valued at up to $40 billion
Schmidt identified Google Brain as a particularly successful result of that experimental approach.
The project began within Google's early research environment, initially involving a relatively small group working on machine learning.
Schmidt recalled a team of approximately 10 to 15 people whose work helped improve the AI systems underlying Google's business.
He estimated that their contributions produced between $10 billion and $40 billion in additional profits over roughly a decade.
“so that pays for a lot of failures” — Eric Schmidt
The financial estimate is Schmidt's retrospective assessment. The interview did not provide audited accounts or a methodology establishing how much incremental profit could be attributed specifically to Google Brain.
Nevertheless, its development is well documented.
The project emerged in 2011 through work involving Jeff Dean, Greg Corrado and Andrew Ng. Its research helped advance Google's use of neural networks and later contributed to technologies such as TensorFlow.
In April 2023, Google formally combined its Brain team with DeepMind, bringing two major AI research organizations into Google DeepMind under Demis Hassabis.
Google Brain's history gives substance to Schmidt's argument about experimental investment: research that initially appears distant from a company's principal products can eventually become central to them.
Why Google accepted failed projects
Schmidt said Google cancelled numerous experiments that seemed promising but did not develop into worthwhile businesses.
Employees could be reassigned, and the company could apply lessons from those attempts elsewhere.
He viewed this willingness to abandon unsuccessful projects as one of Silicon Valley's advantages.
A person could spend several years pursuing an idea that failed, acquire valuable experience and still find another job.
Schmidt joked that a particularly useful chief financial officer might be someone who had already experienced bankruptcy and was determined never to repeat it.
The underlying principle was not that failure was desirable. It was that a company should identify unsuccessful projects before they consume resources that could support better ideas.
Schmidt connected this approach with Bill Gates's preference for failing quickly rather than defending weak projects indefinitely.
For Google, the purpose of its experimental allocation was to make repeated attempts affordable while allowing occasional breakthroughs to justify the overall investment.
2. Google Almost Dismissed YouTube Because It Already Had Google Video
David Drummond brought Schmidt the problem
Schmidt's recollection of YouTube offers a more uncomfortable example of corporate decision-making.
Google had developed Google Video and believed it had a competitive product.
David Drummond, then Google's general counsel, approached Schmidt and urged him to examine a younger competitor called YouTube.
Schmidt initially questioned why he should care about the startup.
He already had a video team, an established internet business and substantial technical resources.
But when Google examined YouTube, Schmidt recognized that its developers were moving faster.
“they really work quicker even though we had an incumbent” — Schmidt
Schmidt attributed the difference partly to organizational constraints.
Google Video operated under established corporate rules and procedures. YouTube's small team enjoyed greater flexibility, including in areas that could create intellectual-property problems.
The lesson was particularly uncomfortable for a company that regarded itself as unusually innovative: possessing better resources did not guarantee producing the better product.
Google paid $1.65 billion for the faster competitor
Instead of continuing to rely entirely on Google Video, Google acquired YouTube.
The transaction was announced on October 9, 2006, for $1.65 billion in stock, according to Google's acquisition announcement filed with the SEC.
Google completed the acquisition on November 13, 2006.
The company initially committed to allowing YouTube to operate independently, preserving its brand, community and much of the agility responsible for its rapid growth.
Schmidt's recollection also requires an important legal distinction.
He described Google ultimately prevailing in the litigation associated with YouTube's intellectual-property practices.
However, the major copyright dispute brought by Viacom did not end in a final courtroom victory. The companies settled the lawsuit in March 2014, after years of litigation, without disclosing the settlement terms.
That outcome does not negate YouTube's commercial success, but it matters when evaluating the risks Google accepted.
The rival inside a company's own building
The YouTube episode illustrates a problem Schmidt encountered repeatedly.
Large companies often employ excellent engineers and possess considerable money, distribution and infrastructure. Yet their established products can prevent them from recognizing that someone else has found a better approach.
Schmidt compared this situation with Apple's development of the original Macintosh.
Steve Jobs assembled a small team in a separate building in Cupertino, away from Apple's main operations. The group displayed a pirate flag, reflecting its willingness to challenge the company's existing practices.
The separation created resentment among other employees.
But Schmidt considered it essential to the Macintosh's success. Its interface innovations helped establish the foundations of Apple's subsequent approach to personal computing and, eventually, the iPhone.
He argued that employees maintaining a successful existing product rarely make ideal candidates for destroying it.
Operating an established business rewards stability. Developing its replacement often requires breaking conventions.
Google's other missed opportunity: Facebook
Schmidt also identified social networking as one of his greatest regrets.
Google had introduced Orkut, a social-networking service that appeared around the time Facebook was emerging.
Yet Google failed to turn that early position into lasting leadership.
Schmidt accepted personal responsibility for the missed opportunity.
The scale of Orkut's eventual reach is easy to overlook. According to Google's own account of its closure, the service's public communities had accumulated approximately 51 million communities, 120 million discussion topics and more than one billion interactions.
Google shut Orkut down on September 30, 2014.
It had attracted considerable participation, but Google failed to establish the dominant social network.
Schmidt's regrets therefore concern two different outcomes. Google recognized YouTube's superiority and bought it. Google recognized the opportunity in social networking but failed to build the business that ultimately won.
3. The Founders, Engineers and Unusual People Who Built Google
Larry Page and Sergey Brin began with very little
Schmidt arrived at Google in 2001, aged 46, following a career that included Bell Labs and Sun Microsystems.
He had become interested in computers as a student, when the university machine he used was, by his estimate, approximately 100 million times slower than a modern smartphone.
The decades of semiconductor improvement that followed created the conditions for his professional success.
Google's founders had followed a different route.
Page and Brin met at Stanford University, where their graduate research helped produce PageRank, an algorithm that used links between websites to assess the importance of pages.
Their work received National Science Foundation support.
Early operations were improvised. Schmidt recalled the founders exhausting available electricity in their dormitory and borrowing power from a neighboring room.
They eventually moved into a garage in Menlo Park.
Google's official company history identifies the property as belonging to Susan Wojcicki, who later became YouTube's CEO.
Sun Microsystems co-founder Andy Bechtolsheim wrote the founders a $100,000 check in August 1998.
Schmidt remembered the investment but was uncertain about the original amount during the interview. Google's historical account confirms $100,000.
Even the early hardware was improvised. Schmidt recalled processors separated with corkboard, creating problems when the equipment overheated.
Professional hardware engineering arrived as the company expanded.
Why Schmidt believed the founders were exceptional
Schmidt attributed much of Google's success to Page and Brin's unusually strong analytical abilities.
He particularly admired their ability to recognize opportunities that others failed to see.
He recalled an occasion when he and Page had spent considerable time developing a strategy on a whiteboard.
Brin arrived from playing volleyball, examined their work, dismissed its direction and suggested several alternatives.
According to Schmidt, Brin's proposed strategy was superior.
The story illustrates how Schmidt viewed his role.
He did not consider himself the company's supreme inventor. He believed his responsibility included finding exceptional people, listening to them and creating an organization capable of turning their ideas into businesses.
Schmidt described successful innovators as people driven by solving difficult problems rather than by personal advancement alone.
In his terminology, an exceptionally demanding but productive individual could be a diva. A person seeking personal benefit at the expense of colleagues was a knave.
Steve Jobs represented the former category.
Steve Jobs, Apple and the price of control
Schmidt knew Jobs personally and served on Apple's board.
He described Jobs as brilliant, difficult, demanding and deeply committed to perfection.
Their disagreements frequently concerned open versus closed systems.
Schmidt generally favored openness. Jobs believed Apple should control its own intellectual property and tightly integrate hardware, software and the customer experience.
Schmidt recalled Jobs becoming frustrated that MP4 files achieved greater adoption than Apple's MOV format. Schmidt believed Apple's comparatively closed QuickTime approach contributed to the outcome; Jobs insisted Apple's technology was better.
Their disagreement reflected two different business philosophies.
Jobs wanted Apple positioned as a premium brand rather than a commodity producer.
Schmidt remembered discussing whether Apple should resemble Chevrolet or Porsche, General Motors or BMW.
Jobs favored BMW.
Schmidt said Apple's profit margins at the time were approximately double those of competing PC companies.
Jobs defended those margins as part of the value of Apple's brand.
Schmidt credited Tim Cook with preserving the company's central priorities after Jobs's death, particularly its attention to interface quality, privacy, security and tightly integrated products.
He nevertheless wondered whether Jobs would have moved more aggressively into AI.
Schmidt expected that Jobs would have embraced artificial intelligence while insisting that the resulting products be exceptionally polished.
One immediate opportunity, he argued, was replacing Siri's limited conversational abilities with a much more capable AI assistant.
4. Why Schmidt Wanted More Engineers and Fewer Corporate Layers
Product quality before sales
Schmidt said that, if he could rebuild a technology company, he would employ an even greater proportion of technical staff.
His reasoning was that superior products create much of their own demand.
He described meeting entrepreneurs who planned to employ approximately 95% technical workers while relying on a contract manufacturer.
Their proposed switching technology, they claimed, could become 100,000 times better than competing products.
Schmidt believed such an advantage would make the product extremely attractive commercially.
There was one problem: it did not yet work.
He used the example to make two related points. Engineering ability creates the opportunity for extraordinary differentiation, but a product must deliver the promised performance before the business can rely on it.
Schmidt also believed that modern technology CEOs should act as chief product and innovation officers.
Distribution, marketing, capital and sales infrastructure had become more accessible than in earlier decades, making product excellence increasingly important.
Why customer measurements mattered more than executive opinions
Schmidt recalled engineers presenting ideas they considered impressive.
He would redirect attention toward customers.
Customers wanted products that worked immediately, rarely failed and cost as little as possible.
He expected engineers to confront customer feedback rather than assume their own enthusiasm proved the product was successful.
Marissa Mayer, who later became Yahoo's CEO, helped establish an important mechanism for doing this at Google.
She questioned whether executives could reliably judge user-interface quality through personal preference.
Instead, Google used A/B testing: showing alternative designs to different groups of users and measuring their behavior.
The same principle underlies the modern recommendation systems Schmidt discussed elsewhere in the interview.
Viewing duration, sharing, commenting and other actions can be measured and used to guide automated decisions.
Hard targets, long forecasts and individual ownership
Schmidt credited Page with establishing demanding quarterly objectives and key results, commonly known as OKRs.
Page set difficult numerical targets and considered achieving approximately 70% of them a respectable result.
The objective was to discourage executives from selecting easy goals or substituting attractive presentations for measurable progress.
Schmidt recommended combining such targets with longer-term forecasting.
He advised entrepreneurs to describe what their industry might look like in five years, then decide what they should accomplish during the next year.
He recalled Google's early business plan, prepared by Salar Kamangar, as unusually accurate when examined years later.
Schmidt also emphasized individual ownership.
Page insisted that important projects needed a specific person responsible for making them happen.
Sundar Pichai later applied a similar distinction between established major priorities and smaller emerging initiatives, some of which grew into important businesses.
For startups, Schmidt preferred intelligence, speed and willingness to take risks over lengthy experience.
He praised Elon Musk's ability to persuade teams to attempt ambitious projects and accept extraordinary risks, although he recognized that such judgment is unusual.
He also warned against focusing excessively on competitors.
Google, he said, spent more time examining what its own technology could accomplish than reproducing what competitors had already built.
The danger of simplifying a business too aggressively
Schmidt challenged the conventional advice that companies should concentrate exclusively on established strengths.
He cited Intel's decision to sell its low-power processor operations as an example.
In June 2006, Intel agreed to sell its communications and application processor business to Marvell for $600 million.
The business included XScale processors based on ARM technology.
Schmidt argued that Intel underestimated how important power consumption, battery life and heat would become in mobile computing.
He contrasted that decision with Nvidia's use of ARM-based processing in AI infrastructure.
Nvidia's Grace CPU architecture demonstrates the continuing importance of ARM technology in high-performance computing.
Intel's transaction is documented. Schmidt's broader conclusion—that selling the business determined Intel's subsequent competitive weakness—is his strategic interpretation, not a result established by the transaction record alone.
He made a similar distinction between cryptocurrency and AI.
Crypto, he argued, remained an important but specialized market. AI could become a general-purpose platform touching almost every occupation and industry, much as personal computers and the internet had done.
5. The Google Culture That Became Too Large to Remain Private
When employees started leaking executive meetings
Google's weekly TGIF meetings were once a central part of its internal culture.
Employees could question senior executives, hear business updates and participate in informal company traditions.
Schmidt remembered the gatherings as humorous and candid.
One executive, sales chief Omid Kordestani, repeatedly forecast revenue below what the business ultimately achieved.
The practice was known as sandbagging.
His colleagues responded by making him stand on a physical sandbag when presenting his figures.
But the informality became increasingly difficult to sustain as Google expanded.
Schmidt recalled a presentation being leaked to a reporter while the meeting was still taking place.
“someone was leaking the presentation live to a reporter” — Schmidt
According to Schmidt, someone interrupted the meeting and said it had to stop.
He regarded the incident as a moment when Google's original culture had become incompatible with the organization's size.
The later change is independently documented.
In November 2019, Google reduced TGIF from a weekly meeting to a monthly event.
Sundar Pichai attributed the change partly to repeated disclosure of internal discussions outside the company.
The official change confirms the wider confidentiality problem, although it does not independently establish every detail of Schmidt's particular recollection.
Google's approach to layoffs
Schmidt said Google generally preferred avoiding unnecessary hiring to conducting regular layoffs.
He recalled a relatively rare reduction involving approximately 200 sales employees, which he described as painful.
The interview's transcription is unclear about the precise timing of that episode, so the number is best treated as Schmidt's recollection rather than a verified historical total.
He opposed a management practice popular in some corporations: periodically dismissing the lowest-performing 5% of employees.
Such systems assumed managers could identify the weakest performers accurately.
Schmidt questioned that assumption and noted that even less productive employees could possess valuable institutional knowledge.
He believed strong compensation, selective hiring and employee retention contributed to Google's success.
Political disagreement inside the company
Google's openness created another problem.
Schmidt recalled discovering approximately 100,000 internal mailing lists or message boards, many devoted to subjects unrelated to company business.
Employees debated politics, war and social issues.
Schmidt identified himself as a Democrat but said he made a point of protecting the relatively small number of Republican employees in Google's workforce.
He believed those employees deserved equal opportunities regardless of their political beliefs.
His preferred policy was to keep workplace activity focused on business while allowing employees to express political views outside work.
He said Google eventually addressed some of these internal cultural problems.
These accounts also illustrate a limitation of founder culture.
The practices that encourage unusually open communication among a few dozen employees can become difficult to manage when an organization employs tens of thousands of people.
The Mayo Clinic's century-old principle
Schmidt offered the Mayo Clinic as evidence that organizational values can endure far beyond their founders.
As a board member, he encountered the institution's long-standing commitment to putting patients' needs first.
Initially, he doubted that such a familiar slogan could guide everyday behavior.
After observing the organization, he concluded that employees took the principle seriously.
He saw something similar in Apple, whose focus on integrated products and interfaces remained recognizable after Jobs's death.
In Schmidt's account, founders establish the standards that later generations of employees inherit, even as the organization becomes much larger and more complicated.
6. How OpenAI Turned Google's Research Into the ChatGPT Breakthrough
Google built the foundation but did not launch first
Schmidt offered an explanation for one of the most consequential competitive surprises in modern technology.
Google had pioneered much of the research underlying large language models.
Its researchers helped develop the Transformer architecture, published in the 2017 paper Attention Is All You Need.
The architecture became central to systems including GPT.
Yet OpenAI, not Google, introduced the conversational product that captured global attention.
Schmidt explained that Google was occupied with numerous established businesses and complex research programs.
OpenAI concentrated on making advanced language models usable.
He identified reinforcement learning from human feedback, or RLHF, as a major part of that breakthrough.
The method involves people evaluating model responses, allowing developers to train systems to produce answers users find more helpful.
The technical advance is documented in OpenAI's March 2022 paper, Training Language Models to Follow Instructions with Human Feedback.
That research demonstrated that a smaller model trained with human feedback could be preferred by evaluators over a substantially larger GPT-3 model.
RLHF was not invented entirely in one unexpected moment. The interview compresses an extended research history into the story of a commercial breakthrough.
Even OpenAI underestimated what it had built
Schmidt recalled joking with OpenAI colleagues about the moment their model became surprisingly effective.
He believed the success caught some of its developers off guard.
“they didn't really understand how good it was they just turned it on” — Schmidt
According to Schmidt, OpenAI released ChatGPT while concentrating substantial effort on GPT-4.
The resulting public adoption created both an enormous commercial opportunity and an unexpected operational challenge.
The interviewer mentioned ChatGPT's rapid growth to approximately 100 million users and reports of panic inside Google.
Schmidt did not establish that Google had experienced the precise internal crisis described in those reports.
Instead, he emphasized Google's competing priorities and the importance of OpenAI's product-level achievement.
The companies Schmidt believed were competing for control of AI
At the time of the interview, Schmidt identified OpenAI, Google, Meta and Anthropic as important competitors.
He compared GPT-4o with Google's Gemini 1.5, describing their capabilities as broadly competitive while noting differences in multimodal performance.
GPT-4o had been introduced by OpenAI in May 2024, with capabilities spanning text, images and audio.
Schmidt also highlighted Anthropic's decision to establish itself as a public-benefit corporation.
He said its founders anticipated that powerful AI might create conflicts between commercial demands and broader public responsibilities.
Meta represented a different strategy.
Schmidt argued that Mark Zuckerberg's willingness to make enormous investments helped Meta become a serious AI competitor.
Zuckerberg absorbed considerable losses pursuing virtual reality while also investing heavily in Facebook, Instagram, WhatsApp and the AI systems supporting those products.
Meta's Llama models, which Schmidt described as openly available, offered another approach to distributing advanced technology.
His central business observation was that founder-controlled companies can sometimes make expensive and controversial decisions that hired executives would struggle to persuade a board to approve.
7. AI Could Transform Publishing, Podcasts and Political Communication
NotebookLM showed Schmidt what synthetic media could become
Schmidt identified Google's NotebookLM as one of the most impressive AI products he had encountered.
He described writing material, asking Gemini to improve its wording, then placing the result into NotebookLM.
The system generated a conversation between two artificial presenters discussing the material.
Schmidt said he had played such recordings to audiences without listeners necessarily realizing that neither presenter existed.
Google officially introduced NotebookLM's Audio Overviews on September 11, 2024.
The company described the feature as a way to transform supplied documents into podcast-style conversations.
Google also warned that these generated discussions could contain inaccuracies and should not automatically be treated as comprehensive or objective accounts of the source material.
The technology advanced after the interview. By 2025, Google had expanded NotebookLM's audio capabilities to additional languages and introduced mobile applications.
What happens when producing another podcast becomes inexpensive?
Bartlett explained that his media company occupied approximately 25,000 square feet of office and studio space.
He wondered how traditional production businesses could compete in a world where AI could generate enormous quantities of audio and video at minimal cost.
Schmidt advised him to use the technology to extend his existing work.
An interview could produce summaries, additional discussions, analysis and even synthetic commentators criticizing the original conversation.
The objective would be to reach additional audiences rather than simply replace the original host.
Schmidt believed AI could substantially increase the amount of content a producer created from existing material.
He also challenged the assumption that abundant content would necessarily destroy celebrity.
During the internet's expansion, he had expected greater opportunity for specialized voices. Instead, global networks also allowed successful personalities to build audiences across multiple countries.
He predicted that AI would produce similar advantages for creators who already possessed recognizable identities and strong audience relationships.
This remains a prediction, not proof that established publishers or creators will automatically benefit. Greater supply also creates greater competition for attention, and the economics depend on distribution, audience behavior and ownership.
Personalized political messages
Schmidt described another potential application.
A political campaign could use AI to identify subjects relevant to different constituents and produce personalized videos discussing local accomplishments.
The politician's likeness would be digitally generated but clearly labeled.
In Schmidt's example, voters concerned about a particular bridge could receive a message explaining the politician's involvement in improving it.
The technology could strengthen direct political communication.
It could also introduce greater opportunities for manipulation, particularly if synthetic material was misleadingly presented as authentic.
That possibility connects Schmidt's enthusiasm for personalized AI services with his much deeper concern about algorithmic influence.
8. The Attention Business: Why Social Platforms Reward Outrage
Schmidt's explanation of TikTok's algorithm
Schmidt described TikTok's recommendation process using the concept of a multi-armed bandit algorithm.
Such systems balance continuing with an option that appears effective against testing alternatives that might perform better.
In a social-media feed, this means repeatedly providing content likely to retain attention while experimenting with related interests.
Schmidt believed the approach could create highly addictive environments.
Someone interested in a subject might receive increasingly similar material until their information environment became unusually narrow.
For entertainment, that could be relatively harmless.
For users already experiencing distress, Schmidt worried that repeated exposure to negative material could make their situation worse.
He contrasted algorithms with parents, friends or teachers who might deliberately introduce encouraging perspectives.
Most commercial recommendation systems optimize measurable outcomes, not an individual user's complete psychological welfare.
Advertising creates an incentive to amplify emotional reactions
Schmidt offered a commercial explanation for online outrage.
Advertising businesses need attention.
Material that provokes strong emotions can retain attention, even when its factual content is questionable.
He connected this with economist Herbert Simon's observation that abundant information makes human attention increasingly scarce.
Schmidt argued that platforms can therefore face an incentive to promote emotionally provocative material over more constructive information.
This was not merely a theoretical concern for him.
He criticized online harassment, misinformation bots and false claims distributed during disasters, including a Florida hurricane.
Google's own choice between revenue and product quality
Schmidt recalled facing a related problem while running Google.
When improvements to a product created opportunities for additional advertising, executives needed to decide how much of that benefit should become revenue and how much should improve the user experience.
He said he adopted an approximate 50/50 balance, supported by Page and Brin.
The idea was that Google could make money while preserving its usefulness.
Schmidt also described an internal preference for preventing deceptive or harmful material from dominating prominent search results.
Those statements explain Google's stated philosophy under his leadership, not an independently established measure of the company's overall social impact.
The evidence about children's mental health is more complicated
Schmidt was particularly concerned about young girls encountering algorithmically personalized content around ages 11 and 12.
He connected social-media use with anxiety, depression, self-harm and emergency treatment.
He also argued that the shift from chronological feeds toward increasingly personalized recommendation systems around 2015 contributed to changes in young people's mental health.
Independent research supports taking the risks seriously, but not all of the causal claims are established.
The US Surgeon General's 2023 advisory concluded that social media could present meaningful mental-health risks and that evidence was insufficient to establish that it was adequately safe for adolescents.
However, a National Academies review published in 2024 found that much of the research established associations rather than definitive causal relationships.
It did not conclude that social media alone caused population-level changes in adolescent mental health.
Schmidt's explanation is therefore a plausible concern requiring continued investigation, not a settled account of the crisis.
He supported practical measures such as restricting phone use during classroom instruction.
His larger question concerned children growing up with AI companions from an early age.
He wondered how identity, values and social development might change when a child's closest digital companion could be a computer.
9. AI Safety: Schmidt's Warnings Have Become More Concrete
Cyberattacks by machines that never sleep
Schmidt warned that advanced AI models could identify previously unknown software vulnerabilities.
He referred to zero-day attacks, which exploit vulnerabilities not previously known to the affected software vendor or defenders.
He claimed that some unreleased models could perform such tasks as well as, or better than, human specialists.
The interview did not provide independent benchmark results establishing that precise comparison.
His explanation of the underlying risk was straightforward: computers can repeatedly test possibilities without sleeping, becoming tired or losing concentration.
Subsequent evidence has made the wider concern more concrete.
On October 19, 2026, Reuters reported an investigation into an AI-assisted hacking campaign.
According to cybersecurity researcher Snorre Fagerland of NorForge, a Chinese-speaking hacker used AI agents to attempt intrusions against more than 1,200 websites and successfully compromise more than 100.
Reuters independently examined activity identifying victims that included websites associated with Deere, SharkNinja and Cummins.
The hacker reportedly gave the agents relatively brief instructions and allowed them to work autonomously.
Successful intrusions often required approximately 23 hours of investigation and repeated attempts.
The researcher estimated computing costs of approximately $5,000 for the campaign, or about $50 per successful intrusion.
Reuters could not establish how much sensitive information, if any, had been extracted during every breach.
The investigation does not prove Schmidt's specific claim about unreleased frontier models outperforming human zero-day researchers. It does establish that commercially accessible AI agents were being used to automate significant parts of real cyberattacks.
Who checks the companies building dangerous systems?
Schmidt described the safety procedures used by AI developers.
Large models undergo expensive training on powerful computing infrastructure, supported by specialized hardware from companies such as Nvidia.
Developers monitor training performance through mathematical measures, including a loss function, and then evaluate the resulting model's capabilities.
Some capabilities may emerge without having been explicitly programmed or anticipated.
Schmidt cited the ability to generate website code from an image as an example of a surprising development.
After training, specialist teams test models for harmful capabilities and apply safeguards intended to prevent dangerous responses.
His concern was that future models might possess knowledge or abilities their evaluators did not know to test.
This raises a practical question: how can a company reliably evaluate a system when it cannot anticipate everything the system might be able to do?
The issue has become more pressing.
On October 15, 2026, Reuters reported that Anthropic had granted independent evaluator METR unusually extensive access.
Six METR personnel received access comparable in some respects to Anthropic's internal risk-assessment teams, including company hardware, office space and internal systems.
The arrangement still excluded customer information, privileged legal material and model architecture.
Anthropic could require narrowly defined redactions but could not veto METR's findings.
The arrangement illustrates both progress and unresolved tensions: outside evaluators need meaningful access, while AI companies remain concerned about revealing proprietary technology.
The same day, Reuters reported, citing Axios, that the US Federal Trade Commission had issued civil subpoenas to OpenAI and Anthropic, seeking information about safety incidents, safeguards and consumer protection.
These were reported investigative steps, not findings that either company had violated the law.
Biological risks and the security of powerful models
Schmidt also expressed concern that AI could assist the development of dangerous biological agents.
He said he participated in a commission examining ways to prevent such misuse.
The concern extends beyond whether a public chatbot will answer a harmful question.
It involves access to the underlying models, specialized scientific knowledge and potentially automated research workflows.
Schmidt compared the protection of highly advanced models with safeguards surrounding nuclear materials.
He recalled visiting a plutonium facility while working with US defense officials.
The facility was protected by multiple security perimeters and armed personnel.
He imagined that especially powerful AI models might eventually require similarly secure computing facilities.
If frontier AI remained concentrated in a small number of installations, governments might be able to negotiate controls resembling nuclear nonproliferation arrangements.
If equally dangerous capabilities became cheap and easy to copy, those controls could be much harder to enforce.
China, Russia and the limits of international cooperation
Schmidt expected powerful AI to become strategically important to major states.
He discussed China and Russia as potential military and technological competitors, particularly in cybersecurity.
His assessment of China was complicated.
He believed Chinese AI systems would likely operate under stronger information controls because the government placed such importance on controlling public speech.
He suggested that those incentives could encourage certain kinds of restraint.
That did not mean China would avoid developing AI weapons.
Schmidt assumed that countries would eventually adapt important new technologies for military purposes, just as they had done with aircraft, tanks and other developments.
He supported international cooperation over AI safety.
The United Kingdom hosted the first major AI Safety Summit at Bletchley Park in November 2023, resulting in the Bletchley Declaration, which recognized shared concerns about frontier AI.
Schmidt anticipated further cooperation at the February 2025 summit in France.
That gathering demonstrated the limits of international agreement. The United States and United Kingdom did not sign the Paris summit's principal declaration on inclusive and sustainable AI.
Agreement that AI presents serious risks does not automatically produce agreement about how governments should regulate it.
10. How AI and Cheap Drones Are Changing Warfare
Schmidt described the Russia-Ukraine war as an early demonstration of a different kind of military conflict.
Traditional military operations often depended on expensive equipment and soldiers operating near the battlefield.
Drones increasingly allow personnel to locate, observe and attack targets from remote positions.
Schmidt cited a hypothetical economic comparison: a $5,000 drone destroying a $5 million tank.
The precise costs and effectiveness vary considerably by weapon and circumstances, but the broader problem is substantial.
Relatively inexpensive equipment can threaten vehicles costing millions of dollars.
Both Russia and Ukraine have adapted their battlefield practices around unmanned systems.
Schmidt expected drone-on-drone combat to become increasingly important as countries developed ways to intercept opposing unmanned aircraft.
AI could expand these capabilities through autonomous navigation, target identification and coordination.
The military consequences also concern industrial production.
Schmidt referred to the book Freedom's Forge, which examines America's extraordinary manufacturing mobilization during World War II.
His point was that nations facing serious military emergencies can reorganize production much faster than peacetime assumptions might suggest.
The combination of inexpensive autonomous equipment, powerful AI and large-scale industrial mobilization could change military strategy.
Schmidt did not claim to know exactly how future warfare would develop.
He expected the importance of automation to increase, while recognizing that both sides of a conflict could adopt similar technologies.
11. When Schmidt Thinks Humanity Should Shut AI Down
The problem of machines improving themselves
Schmidt rejected the idea that artificial general intelligence would necessarily appear on one identifiable day.
He expected progress through successive advances in scientific reasoning, writing, mathematics, psychology and other specialized capabilities.
He anticipated powerful AI systems helping scientists make further discoveries, including discoveries that improve AI itself.
That possibility leads to one of his central safety concerns: recursive self-improvement.
An AI system might become capable of improving its own capabilities repeatedly.
Schmidt believed operators should intervene if they could no longer understand what the system was learning.
He identified another warning sign: a sufficiently powerful model developing a successor faster than researchers could properly evaluate the previous model.
In both scenarios, technological progress would be moving faster than human oversight.
AI agents communicating in an unknown language
Schmidt also described AI agents: models equipped with memory and the ability to perform tasks, use tools and cooperate with other systems.
A complex project might involve several agents passing information between one another.
He regarded such coordination as an important future development.
But he also saw a potential warning sign.
Agents exchanging information in ordinary human-readable language could be inspected by their operators.
If they developed a private language that only other machines could understand, oversight would become substantially more difficult.
“that's a good time to pull the plug” — Schmidt
His proposal was not a claim that autonomous systems had already created such a language or escaped human control.
It was an example of a situation in which he believed operators should exercise their authority.
Asked whether advanced systems could still be physically disconnected, Schmidt pointed to the existence of power supplies and circuit breakers.
That answer addresses a physical possibility, not every technical challenge involved in containing distributed AI systems.
Why greater intelligence does not automatically mean better judgment
Schmidt illustrated his concern using a hypothetical AI with an IQ of 1,000 advising a person with an IQ of 100.
The numbers were illustrative, not a meaningful measurement of machine intelligence.
His argument was that a system could possess extraordinary analytical abilities without understanding human values adequately.
A machine optimizing traffic efficiency, for example, might fail to prioritize a passenger experiencing a medical emergency unless its rules accounted for such circumstances.
Schmidt believed AI should help humans understand complicated problems, evaluate historical evidence and compare possible solutions.
He did not believe people should automatically surrender moral judgment to the system producing the most sophisticated calculation.
His concern extended to democracy.
Political decisions require shared facts, public trust and human accountability. AI-generated misinformation could weaken all three if it becomes difficult to distinguish authentic events from convincing fabrications.
Schmidt did not regard human extinction as the inevitable outcome of advanced AI.
He argued that humanity had survived catastrophic wars, atrocities and epidemics, although such survival had involved enormous suffering.
The distinction was important: rejecting an inevitable extinction scenario did not mean dismissing the possibility of severe harm.
12. AI, Employment and the People Who Still Need Jobs
Why Schmidt predicts more work, not permanent mass unemployment
Schmidt disagreed with predictions that AI would permanently eliminate the need for human employment.
He compared the fears surrounding AI with earlier technological disruptions, including the industrial changes that prompted the Luddite movement.
His central economic argument concerned demographics.
Developed countries are aging, fertility rates are low in many places, and fewer working-age people will need to support larger elderly populations.
Greater productivity could help societies cope with those pressures.
Schmidt compared AI-assisted knowledge workers with machinists who moved from manually operated equipment to computer-controlled machines.
The workers might accomplish much more without their occupations disappearing entirely.
He pointed to increasing factory automation in China, Japan and South Korea, where manufacturers have faced changing labor costs and demographic pressures.
Schmidt nevertheless acknowledged that technological change creates serious problems for individuals.
A displaced worker may need additional training, a different occupation or a move to another location.
Security guards, lawyers and Hollywood production crews
Schmidt identified repetitive, monotonous and dangerous jobs as particularly vulnerable to automation.
He suggested that sufficiently capable robots could replace some security guards, performing continuous monitoring without becoming tired.
In professional services, he expected AI to automate many routine legal tasks.
He did not conclude that this would necessarily produce fewer lawyers.
Instead, he predicted that legal professionals might handle larger workloads and develop more complicated services.
His Hollywood example was similar.
AI-generated sets, backgrounds and makeup could reduce production costs while allowing established actors and producers to continue earning substantial incomes.
But workers responsible for physical sets or makeup might face displacement.
Schmidt suggested that some could transfer their skills into construction, personal care or other industries facing shortages of skilled craftspeople.
Such transitions may be possible, but the interview did not establish that new employment would provide comparable wages, locations or working conditions.
Why Schmidt rejects the promise of universal abundance
Schmidt criticized a particular technological vision in which AI produces so much wealth that most people no longer need to work.
The interviewer raised Worldcoin and universal basic income.
Schmidt distinguished Worldcoin from a guaranteed-income program and then challenged the broader idea that enormous productivity would make work largely unnecessary.
He believed human ambition and competition would continue generating new demands.
People might use more productive tools to accomplish more rather than simply stop working.
He expected individuals to retain their appetite for status, achievement, cooperation, personal relationships and meaningful activity.
His examples included human-operated Formula 1 racing and marathon running.
Machines might eventually outperform human competitors, but audiences could still find human achievement more compelling.
He also regarded direct brain-computer integration, including Neuralink, as more speculative than the ordinary use of AI assistants.
Schmidt said he would not undergo an invasive brain procedure without a compelling reason.
13. Remote Work: Schmidt Admitted the Evidence Challenged His Preference
The discussion of office attendance produced one of the interview's clearest examples of Schmidt separating personal conviction from available evidence.
Bartlett strongly favored office-based work.
He believed employees, particularly younger people, benefited from shared spaces, social relationships and opportunities for spontaneous interaction.
Schmidt largely agreed.
He recalled learning important professional skills through conversations in hallways, attending meetings and spending time with experienced colleagues.
He believed those informal experiences helped him secure promotions.
For young professionals, he regarded being physically present as valuable.
But he acknowledged a difficulty.
“the evidence does not support our view” — Schmidt
He said research indicated that allowing employees to work from home could improve productivity, even though he personally preferred office environments.
Independent evidence helps clarify the distinction.
A 2024 randomized controlled trial published in Nature studied 1,612 employees at Trip.com, a Chinese technology company.
Researchers compared a hybrid schedule involving two days at home with working five days in the office.
The study found no significant deterioration in job performance or promotion outcomes, while employee resignations fell by approximately one-third.
Job satisfaction also improved.
That does not establish that working from home always increases productivity, nor does it settle every question about mentorship or innovation.
It does demonstrate why blanket assumptions about remote workers being less effective can be misleading.
Schmidt's admission was notable because he did not attempt to dismiss the findings merely because they conflicted with his preferences.
14. What Schmidt Would Tell an 18-Year-Old Entering the AI Economy
Bartlett raised a personal question involving his girlfriend's younger brother, Raph, who was 18.
The family had been discussing which skills Raph should acquire to prepare for the future.
Schmidt's answer began with critical thinking.
He believed the most important ability was to evaluate claims independently rather than accept information because it sounded plausible.
He recommended learning to reason analytically through mathematics, science, law or other demanding subjects.
Schmidt also encouraged young people to learn Python.
Although AI systems increasingly generate code, he argued that programming knowledge remains valuable because developers can use AI services through application programming interfaces.
An API allows one program to request information or services from another.
For example, a developer could create an application that sends an image to an AI system and asks it to identify objects.
Schmidt suggested building a game as an accessible way to begin learning.
His argument was not that young workers should compete directly with machines at producing routine code.
It was that they should understand enough about technology to create useful applications with it.
The value of verifying claims
Schmidt linked critical thinking with the wider problem of misinformation.
He said people routinely repeat statements that appear credible but fail basic factual checks.
He gave examples involving exaggerated claims about criminality in Congress and mistaken assertions about how many Americans possess passports.
He recommended checking such claims before passing them on.
Scientific reasoning, he argued, depends on evidence that can be tested and potentially disproved.
He used climate change to distinguish between establishing the reality of an observable phenomenon and debating the appropriate response.
Schmidt regarded the ability to distinguish facts from persuasion as increasingly important when artificial intelligence can produce highly convincing misinformation.
The opportunity Schmidt nearly missed himself
Schmidt also advised younger people to take more risks.
He believed careers were shaped by opportunities that appeared for limited periods.
People sometimes rejected those opportunities because of inconvenience, fear or uncertainty.
He recalled that several candidates had turned down the opportunity to work with Page and Brin before he joined Google.
A board member later reminded Schmidt that his most important contribution to the decision had been accepting the offer.
The opportunity changed his career.
Schmidt's advice was to remain willing to travel, relocate, acquire unfamiliar skills and accept difficult assignments when the potential reward justified the effort.
15. Henry Kissinger's Final Project: What Genesis Was Trying to Answer
The conference that changed Kissinger's attention
Schmidt said Henry Kissinger became one of his closest friends.
Their collaboration on artificial intelligence began approximately a decade before the interview.
The two attended a conference where DeepMind co-founder Demis Hassabis was scheduled to speak.
Kissinger was considering leaving to recover from jet lag.
Schmidt encouraged him to remain and listen.
According to Schmidt, Hassabis's presentation convinced Kissinger that artificial intelligence represented a profound challenge whose consequences were not adequately understood.
The subject connected with questions Kissinger had considered since studying philosophy at Harvard after World War II.
What happens when humans encounter a form of intelligence that rivals their own?
What happens to knowledge, judgment, politics and human independence when machines become capable of performing intellectual tasks previously reserved for people?
These questions eventually shaped Genesis.
The book completed near the end of Kissinger's life
Schmidt said Kissinger continued working on Genesis during his final days.
“he finished the last chapter in his last week of life in his deathbed” — Schmidt
Kissinger died on November 29, 2023, aged 100.
The book was subsequently published on November 19, 2024, by Little, Brown and Company.
Its full title is Genesis: Artificial Intelligence, Hope, and the Human Spirit.
Schmidt said Kissinger wanted the book to help humanity navigate the next 50 years.
Having lived through World War II, the Cold War and major technological changes, Kissinger wanted to consider how human institutions should respond to an unfamiliar kind of intelligence.
Schmidt described the work as an attempt to preserve human judgment while benefiting from technological progress.
The book's arguments do not depend on agreement with Kissinger's political legacy. Its central question concerns the relationship between human beings and increasingly capable machines.
16. Schmidt's Greatest AI Fear Was Not What Bartlett Expected
After discussing cyberattacks, biological threats, military automation and the possibility of systems escaping human understanding, Bartlett asked Schmidt what frightened him most.
His answer was surprising.
“my actual fear is we're not going to adopt it fast enough to solve the problems that affect everybody” — Schmidt
Schmidt wanted AI development to proceed quickly enough to improve ordinary lives.
He regarded healthcare and education as particularly valuable opportunities.
An AI teacher for every child
Schmidt proposed AI teaching assistants working alongside human educators.
Such systems could communicate in a child's language, account for cultural circumstances and adapt lessons to individual abilities.
They could offer explanations and additional practice without requiring every teacher to provide one-to-one instruction constantly.
Schmidt believed this could make excellent educational support available to many children who currently lack it.
The proposal also involves risks.
UNESCO's guidance on generative AI in education emphasizes privacy, age-appropriate use, human involvement and the importance of retaining teachers' professional roles.
Personalized instruction is a promising application. Its effectiveness and safety depend on how systems are designed, evaluated and supervised.
An AI assistant for every doctor
Schmidt proposed a similar arrangement in healthcare.
An AI assistant could help a physician evaluate possible treatments while considering the patient's circumstances, the local availability of medicines, insurance coverage and the country's healthcare system.
He imagined doctors having access to a far wider range of relevant information without needing to recall every medical publication themselves.
The World Health Organization has examined this possibility.
Its January 2024 guidance on large multimodal AI models identified potential applications in diagnosis, patient care, medical education, administration and drug development.
But the WHO also warned about inaccurate information, bias, privacy violations and the danger of professionals accepting automated recommendations without sufficient scrutiny.
Schmidt's envisioned medical assistant would therefore need to support clinical judgment rather than replace accountability for treatment.
The tension at the heart of Schmidt's argument
Schmidt's position contains two demands that are not easily reconciled.
He wants AI systems developed and deployed rapidly because they could improve healthcare, education, scientific research and productivity.
At the same time, he wants governments and developers to retain the power to interrupt progress when increasingly capable systems become too dangerous or too difficult to understand.
His proposals for safer deployment involve technical safeguards, institutional oversight, meaningful evaluation and human control over important decisions.
His enthusiasm for adoption is strongest where the technology can address tangible human needs.
His caution is strongest where AI can act autonomously, affect national security, manipulate information or make decisions without sufficient human understanding.
By October 2026, developments in autonomous cyberattacks and external AI safety evaluations had made those questions considerably more immediate than they were when the interview was recorded.
Schmidt's experience at Google supplies an instructive historical parallel. The company gained enormously from giving small teams room to experiment, but it also discovered that successful technologies, once widely adopted, could create problems that exceeded the control of the original team.
The challenge he now describes is much larger: developing intelligence powerful enough to transform society while preserving humanity's ability to decide how that intelligence should be used.
Reduce the article to its strongest arguments