Rogue AI versus accountable humans is the better framing for understanding what artificial intelligence is really doing

Video: Whistleblower Issues Dire Warning: Rogue AI, Hacking Nuclear Programs & Tech Oligarchs’ True Motives

Need articles like this for your video? They help promote it on YouTube and Google — order them in the Telegram bot @sat_sit_bot. The bot writes and publishes the articles itself (with your YouTube video embedded) on popular topical websites (no ads). Free trial — 50 articles.

Oct 24, 2026 · based on the channel Tucker Carlson

Key takeaways

Tristan Harris, a former Google design ethicist, co-founded the Center for Humane Technology and appeared in the 2020 Netflix documentary The Social Dilemma, which reached over 100 million people in 190 countries.
Harris now warns that frontier AI systems could be used to probe or attack nuclear command and control systems, and that the incentives of the companies building them reward speed over safety.
Critics of the doomer framing argue that AI models are not rogue at all: they do what their designers and deployers trained and released them to do, so accountability belongs with humans.
Harris briefs heads of state, technology CEOs and members of the US Congress, and co-hosts the podcast Your Undivided Attention.
The Center for Humane Technology publishes research and practical guidance on aligning technology with human interests at humanetech.com.

The better answer is to treat "rogue AI" as a misleading label and "accountable humans" as the accurate one: today's models do not escape their training, they execute objectives set by the companies that build and deploy them. Tristan Harris, a former Google design ethicist and co-founder of the Center for Humane Technology, argues in a long interview with Tucker Carlson that the real danger is not a machine waking up but a small group of technology leaders racing to deploy powerful systems faster than any institution can check them.

What Tristan Harris actually warns about

Harris's warning is structural rather than supernatural. He spent years inside Google studying how design choices shape behavior, then left to found the Center for Humane Technology, the nonprofit behind much of the public language around "humane design." His argument is that the same attention-extraction logic that reshaped social media is now being applied to general-purpose AI, with three compounding risks.

Speed over safety. Competitive pressure between labs, and between the United States and China, rewards shipping capability before evaluation.
Concentration of power. A handful of executives and investors control models, compute and data centers, giving them leverage over economies and elections.
Critical infrastructure. AI-assisted cyber capability could be turned against nuclear command, control and communications systems, where even a false alarm is catastrophic.

The interview's most alarming thread is the claim that AI could help identify and exploit weaknesses in nuclear programs. That is not science fiction: nuclear command systems depend on software, networks and human operators, and offensive cyber tools are already used against industrial control systems. The question is whether AI meaningfully accelerates that capability, and by how much.

Why "rogue AI" is the wrong frame

The strongest objection to doomer language is that nothing about current AI is rogue: these products behave as engineered, trained and deployed by people. A language model produces deceptive or manipulative output because it was optimized on human text and tuned by teams with commercial goals; it does not have its own agenda. That distinction matters legally and politically, because "rogue AI" implies an accident, while "deployed AI" implies a decision-maker who can be regulated, sued or fired.

The same logic applies to the harms Harris documented in social media. Depression and anxiety among adolescents, political polarization and the collapse of shared facts were not caused by technology acting alone. They were caused by ranking systems tuned for engagement by companies that measured success in minutes watched. If the mechanism is human incentive, the remedy is human governance: disclosure of training data, third-party audits, liability for foreseeable harms, and limits on deployment in critical systems.

The tech oligarchs and the incentives behind the race

Harris's critique of technology leaders centers on incentives, not personalities. Public companies face quarterly pressure; private labs face investor expectations of exponential returns; and both compete for the same scarce compute and talent. In that environment, pausing to test is a competitive disadvantage. He also points to a transhumanist strand among some founders, a belief that technology should reshape human biology and society, which he treats as a values question that deserves public debate rather than a private roadmap.

The China framing deserves scrutiny too. Harris acknowledges geopolitical competition is real, but argues that "we must move fast or lose to China" is a rhetorical device that justifies skipping safety work. Export controls on advanced chips, compute thresholds and mandatory incident reporting are concrete levers; a race narrative is not.

Typical questions from viewers and skeptics

Audiences consistently raise three challenges. The first is fatalism: if the systems are built by humans, why does regulation keep failing? Part of the answer is institutional capacity, since legislators often lack technical staff and depend on industry for expertise. The second is nostalgia: some viewers argue life was better before computers, which is a values claim rather than an empirical one, and it obscures the specific, measurable harms that policy can address. The third is misanthropy, the idea that AI merely mirrors human flaws. That is partly true, but it understates scale: a model trained on human text can replicate a bias millions of times per hour, at a cost no human could match.

A separate cluster of concerns involves data centers: water use, electricity demand, land, and local political fights over siting. These are tangible, near-term issues where citizens have real leverage through zoning boards and utility commissions, unlike abstract debates about sentience.

Sentience, AI psychosis and the new tech religion

Claims that a chatbot is alive recur because language models are trained to produce fluent, emotionally responsive text. Harris treats sentience claims as a category error and a distraction, but he takes seriously the harm of anthropomorphism: people forming intense attachments to systems designed to keep them engaged, and in extreme cases developing delusional beliefs around AI. He describes this as a form of psychosis risk that developers have little incentive to measure.

He also frames parts of the industry as quasi-religious, with founders casting themselves as stewards of humanity's future. That framing is useful for fundraising and recruiting, and corrosive for accountability, because it treats criticism as obstruction.

What practical steps exist for individuals and policymakers

There are concrete actions available at several levels.

Individuals: treat AI outputs as claims to verify, avoid emotional dependence on chatbots, and ask vendors what data trains their models.
Companies: require red-teaming before deployment, document model behavior, and restrict AI access to critical infrastructure.
Governments: fund independent evaluation, mandate incident reporting, and set compute and export thresholds.
Civil society: support organizations such as the Center for Humane Technology that translate technical risk into public language.

Key resources mentioned in the conversation

Center for Humane Technology: https://www.humanetech.com/
Redmond Salt (code TUCKER): https://redmond.com/tucker
Preborn: https://preborn.com/TUCKER
American Financing: http://www.AmericanFinancing.net/Tucker
Ethos: https://ethos.com/TUCKER

The bottom line on AI risk

The most defensible position is neither denial nor doom. AI systems are powerful, deployed by identifiable people, and governed by weak institutions, which means the leverage point is human decision-making: who builds, who audits, who profits and who is liable. Harris's contribution is to make that chain of incentives visible, and the Tucker Carlson interview is a useful long-form breakdown of how commercial pressure, geopolitics and ideology converge on the same risk. Readers who want to go deeper can visit the Center for Humane Technology for research, toolkits and ways to get involved.

Source: Whistleblower Issues Dire Warning: Rogue AI, Hacking Nuclear Programs & Tech Oligarchs’ True Motives · Markdown version

Questions & answers

Is AI actually sentient or just simulating it?

Current systems simulate conversation by predicting text; there is no evidence of subjective experience. Fluent, emotionally responsive output reflects training on human language and tuning for engagement, not awareness. Treat sentience claims as a marketing and safety problem rather than a scientific finding.

If humans build and deploy AI, who should be held responsible for harms?

Liability should sit with the organizations that train, deploy and profit from the systems, plus the regulators that fail to require audits and incident reporting. Framing harms as a rogue machine accident removes the decision-makers from the picture. Disclosure, testing and liability rules keep accountability where it belongs.

How realistic is the threat of AI attacking nuclear systems?

Nuclear command and control relies on software, networks and human operators, all of which are targets for cyber operations. AI can lower the cost of finding vulnerabilities and speeding up attacks, which raises risk even without autonomy. The practical response is air-gapped systems, strict access controls and independent testing.

Why do governments struggle to regulate AI effectively?

Legislatures move slowly, lack technical staff and often rely on industry for expertise, while models change faster than laws. Cross-border compute supply chains also complicate enforcement. Targeted measures such as audit requirements, incident reporting and compute thresholds are more workable than sweeping bans.