Manufactured Dread: How Feeds, AI Hype and Blackout Myths Sell Fear

Introduction: Three Claims, One Pipeline

Scroll long enough and the same package arrives in different wrappers. A video warns of a nationwide blackout that will last for months. A clip insists that artificial intelligence is about to end humanity. A third tells you to buy a generator and a year of freeze-dried food, today. Each claim is emotionally gripping, each is easy to share, and each is delivered by a system that is built to reward exactly that.

This post does not argue that nothing can go wrong. Grids do fail, and AI does carry real risks. The goal is narrower: to separate what the primary sources say from what the feed says, to explain how the underlying systems actually work, and to be honest where the evidence cuts against the skeptic as well. Where a popular talking point does not survive contact with the record, I will say so. Where a popular rebuttal does not survive either, I will say that too.

Why Your Feed Prefers Fear

The Mechanism: A Ranker Optimizing a Proxy

A modern recommendation feed is a ranking model. For every candidate post it predicts how likely you are to take actions the platform can measure: dwell time, replies, reposts, follows. It then sorts by a weighted sum of those predictions. Nothing in that loop asks whether the post is true, calm, or consistent with what you say you care about. Engagement is a proxy, and the model optimizes the proxy.

A Stanford-led study published in PNAS in 2026 tested what that does in practice on X's "For You" feed. The researchers found a value misalignment between the content the algorithm surfaced and the personal values users stated. The feed weights replies and time spent heavily, and users were more likely to engage with content they disagreed with than content they found more palatable.^[1] In plain terms, the smarter the ranker gets at predicting your behavior, the more it can mistake outrage for interest.

What the Data Says About Falsehood and Fear

The best-known large-scale measurement is the 2018 Science paper from MIT's Vosoughi, Roy and Aral. They examined about 126,000 news cascades, tweeted by roughly 3 million people more than 4.5 million times, with true and false labels drawn from six independent fact-checking organizations that agreed 95 to 98 percent of the time. Falsehood traveled farther, faster, deeper and more broadly than the truth in every category, and the effect was strongest for political news. False stories were 70 percent more likely to be retweeted than true ones.^[2]

Two details matter for anyone writing or sharing content. First, bots were not the explanation. Automated accounts accelerated true and false stories at the same rate, which means humans are the reason falsehood wins. Second, the authors found false news was more novel, and that replies to false stories showed more fear, disgust and surprise, while replies to true stories showed more anticipation, sadness, joy and trust.^[2] Novelty plus fear is the recipe for a doom clip.

Yale researchers added the learning mechanism. In a Science Advances study combining real-world Twitter data with controlled experiments, people who received likes and shares for outraged posts went on to express more outrage later. The platform's reward structure trained the behavior.^[3]

The feed does not need anyone to lie on purpose. It only needs to pay more for the post that frightens you.

Neil Singh

This Is a Design Choice, Not a Law of Nature

A Stanford team has shown that a feed-ranking algorithm with democratic values built into its objective reduces partisan animosity.^[4] Ranking is an engineering decision. Which brings us to the claims that ride those feeds best.

AI Doom Talk: Two Things Can Be True

The Concern Is Not Only Marketing

The cynical reading, that every AI extinction warning is a sales pitch, is too simple. In May 2023 the Center for AI Safety published a one-sentence statement saying that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Signatories included Geoffrey Hinton and Yoshua Bengio, two of the field's most decorated researchers, alongside lab leaders such as Sam Altman and Demis Hassabis.^[5] Serious academics hold these concerns for reasons that have nothing to do with a product launch.

The Regulatory Capture Argument

There is also a documented counter-argument, and it deserves a fair hearing. Regulatory capture is a well-established concept: incumbents shaping rules so that compliance costs fall hardest on new entrants. Andrew Ng, co-founder of Google Brain, has said publicly that some large technology companies are stoking extinction fears to avoid competing with open-source AI, and that licensing requirements would crush innovation.^[6] In written testimony to the Senate's AI Insight Forum in December 2023, venture investor Martin Casado warned that calls for licensing would hand incumbents a framework they could exploit.^[7]

Two cautions keep this honest. Casado is a venture investor in the AI industry, so his testimony comes from a party with a financial interest, not a neutral observer. And the other side has serious voices too: a 2025 paper on open-source generative AI argues that "open" should not mean exempt from law or safety obligations.^[8] The honest summary is that capture risk is real and plausible, whether any specific proposal amounts to capture is contested, and everyone in the argument has something at stake.

What the Software Actually Is

A large language model is a statistical system trained on large text corpora to predict likely next tokens. It runs on clusters of GPUs, it consumes substantial electricity, and it produces fluent text that can be confidently wrong. None of that makes it harmless, and none of it resembles the omnipotent mind in the doom clip. The gap between the engineering reality and the narrative is where hype, in both directions, operates. A June 2023 story about a "rogue AI drone" that went viral turned out, as Cory Doctorow later noted, to describe a hypothetical scenario rather than a real event.^[9]

A useful test for any AI claim: does it describe a mechanism you could check, who benefits if you believe it, and what evidence would change the author's mind? Documented harms today, such as surveillance, manipulation and market concentration, pass the first test easily. I covered those in The Digital Panopticon.

How the Power Grid Actually Works

There Is No Single National Grid

The continental United States runs on three largely separate synchronous interconnections: Eastern, Western and Texas (ERCOT). They are linked only by a small number of direct-current ties.^[10] The official report on the August 2003 Northeast blackout makes the point concretely: the event affected the northeastern portion of the Eastern Interconnection, about 10 percent of that interconnection's load, while the other two interconnections were unaffected.^[11] A single failure that blacks out an entire continent for months is not how this architecture behaves.

What Real Failures Look Like

Real grid failures are regional, weather-driven and usually traceable to fuel and equipment. In February 2021, Winter Storm Uri caused more than 4.5 million Texans to lose power, some for as long as four days, and the federal investigation attributed the outages to freezing at generators and natural gas facilities.^[13][14] Death estimates range from 246 to as many as 702.^[15] This was the largest controlled load shed in U.S. history, and it exists for a reason: shedding load in a controlled way is how operators keep system frequency from collapsing, and Texas avoided a complete grid failure.^[16]

Winter Storm Elliott in December 2022 repeated the pattern across the eastern half of the country. FERC and NERC reported about 90.5 GW of incremental unplanned generation outages, with 55 percent of outages, derates and failures to start caused primarily by freezing and fuel issues, and called it the fifth winter storm in eleven years to jeopardize bulk power reliability.^[17]

The regulatory response is public and checkable. FERC approved NERC's extreme cold weather standards EOP-011-3 and EOP-012-1 in February 2023, and approved the third version, EOP-012-3, in September 2025 with an October 1, 2025 effective date, requiring generators to plan, winterize and train for extreme cold.^[18][19] Note the number: it is EOP-012, the generator cold weather standard. That is the system working through its failure-and-fix cycle, slowly and imperfectly.

Even Total Blackouts Get Restored Fast

Two recent cases show what a deliberate or catastrophic collapse looks like when engineers have restoration procedures. On April 28, 2025, Spain and Portugal suffered a complete blackout affecting about 60 million people.^[20] The expert panel's final report, published March 20, 2026, found multiple interacting factors centered on voltage control, not a single cause.^[21] Power was back for 99.95 percent of demand by 7 AM the next day, using islanded segments reconnected step by step with help from France and Morocco.^[22]

The cyber record points the same way. The 2015 attack on Ukrainian distribution companies cut power to roughly 225,000 customers, and operators restored service in one to six hours by switching to manual operation.^[23] Dragos, analyzing the malware used in the 2016 Ukraine incident, concluded that outages from it would likely last hours or days rather than weeks or months.^[24] Manual restoration is slow and expensive, but it works.

Where the Skeptics Are Wrong Too

Honesty requires the counter-example. Puerto Rico experienced the longest blackout in U.S. history after Hurricane Maria in 2017.^[25] More than a month after landfall, DOE estimated that roughly 1.29 million customers, 81 percent, were still without power.^[26] By late November, restored load was still under 60 percent.^[27] In 2022 Hurricane Fiona again left all 1.5 million customers without power.^[28] Hurricane Ian likewise left some Florida customers without power for more than two weeks.^[29]

The lesson is specific. Long outages happen when physical damage is severe, repair access is limited, and the system is an island that cannot receive help by land. That is a regional hazard with a regional footprint, not a nationwide digital switch-off. Restoration priority in those events is critical facilities first, as DOE's situation reports describe, not revenue.^[26]

Practical Preparedness Versus Apocalyptic Paranoia

What Agencies Actually Recommend

The classic federal baseline was three days of food, water and supplies per person. Many emergency management agencies have since moved the planning figure to about two weeks. FEMA and Washington's Emergency Management Division were reported as recommending two weeks where the guidance used to say 72 hours, and Oregon's "2 Weeks Ready" program asks households to hold at least two weeks of water, food and critical supplies, with one gallon of water per person per day as the standard.^[30][31][32] Notice what that figure matches: the duration of a severe regional event, such as Uri, Ian or a major ice storm. It is not a figure designed for surviving months in the dark, and nobody credible says it is.

Panic Is Mostly a Myth, and Stockpiling Is Not the Same Thing

Disaster researchers have long rejected the idea of mass panic. Researchers Savage and Torgler argue that stocking up in a crisis is usually a rational response to uncertainty, not irrational panic.^[33] A University of Kent study found that most people who store food and water do so "just in case," not because they believe society is about to collapse.^[34] But there is a real effect hiding in the data. A 2021 study in Frontiers in Psychology found that panic buying was influenced more by seeing others stockpile, by personality, and by catastrophic thinking than by actual proximity to danger.^[35] Fear content on a feed is precisely the cue that triggers that behavior, and mass hoarding empties shelves for everyone else.

A Calm Checklist

• Water: one gallon per person per day, for up to two weeks, plus a way to treat more.
• Food: shelf-stable items you will actually eat, rotated into normal use.
• Heat and cooking: a safe alternative source, with fuel stored properly. Never run a generator or camp stove indoors; carbon monoxide kills.
• Light and information: flashlights, spare batteries, a battery or crank radio, charged power banks.
• Medications and documents: a few weeks of prescriptions where possible, plus copies of key records and some cash.
• Neighbors: know who nearby needs help. Community is the most reliable item on the list.