"AI Will Kill Us" is Another Example of a Mania

An existential threat to human life from artificial intelligence seems much in the news these days:

• Anthropic safety researcher Jacob Coxon resigned, warning that labs are racing toward superintelligence without safeguards

• Evan Hubinger, an Anthropic safety researcher estimated a greater than 10 percent chance AI could kill all humans within the decade

• United Nations rights chief Volker Türk warned that AI could pose an "existential" risk to humanity

• Anthropic chief scientist Jared Kaplan warned that humanity could face destruction if control over advanced models slips away.

In the popular imagination, that likely conjures up a “revolt of the machines” scenario. 

But it is not crazy to argue that popular reaction now resembles a "mania" or moral panic. Sociologist Stanley Cohen created the phrase.

Moral panic is a widespread and exaggerated fear that an evil person, group, or entity threatens a community or society. 

Panic	Time Period	Underlying Cultural Anxiety	The "Folk Devil" or Target
The Salem Witch Trials	1692–1693	Religious anxiety, border warfare, and communal instability in colonial Massachusetts.	Marginalized women and community outsiders accused of witchcraft.
The Comic Book Panic	Late 1940s–1950s	Post-WWII anxiety over juvenile delinquency and the corruption of youth culture by mass media.	Comic book publishers (specifically horror and crime genres) and teenage readers.
Dungeons & Dragons and Satanic Panic	1980s	Fear of changing family structures, secularism, and the rise of youth fantasy subcultures.	Role-playing gamers, heavy metal music fans, and alleged secret satanic cults.
The "Super-Predator" Moral Panic	Mid-1990s	Fear of escalating urban crime rates and changing racial demographics in major cities.	Inner-city youth, particularly young Black males framed as remorseless criminals.

But AI systems (including embodied robots) are far less likely to pose an independent existential threat than human misuse of AI tools for catastrophic ends such as bioweapons, large-scale cyber sabotage, or other malevolent applications.

“The robots wake up and kill us” scenario overweights science-fiction, compared to the human bad actor or accident scenarios which seem much more likely.

AI might be used to create novel pathogens or toxins, though, and might allow smaller groups to cause mayhem far easier than once was possible. 

Defensive measures also tend to lag offensive applications because the latter can be pursued by fewer, less ethically-constrained actors. 

So far, most documented AI-related harms to date (fraud, deepfakes, biased decision systems, phishing) are human-directed misuse.

At least so far, human misuse has capable, motivated actors, concrete targets, and accessible deployment paths. By contrast, a genuinely rogue AI that can autonomously seize resources, evade shutdown, and sustain control would require substantial capabilities that current systems do not possess.

Study or framework	AI-danger category	Main finding or relevance	Source
International AI Safety Report (2025), led by Yoshua Bengio with more than 100 experts	Malicious use; system failures; systemic risks; loss of control	Finds that general-purpose AI can support scams, extortion, targeted manipulation, non-consensual sexual imagery, disinformation, and emerging offensive cyber activity. It treats loss of control as hypothetical and says existing systems cannot meaningfully undermine human control.	International AI Safety Report overview internationalaisafetyreport
Brundage et al., “The Malicious Use of Artificial Intelligence” (2018)	Deliberate weaponization and criminal misuse	Landmark forecast of AI as a dual-use technology. Organizes harms across digital, physical, and political security, emphasizing lowered costs, improved targeting, automation, and scale for malicious actors.	Full report (PDF) eff
UK Government, “Safety and Security Risks of Generative AI” (2023)	Cybercrime, political manipulation, critical-infrastructure misuse, CBRN assistance	Assesses digital risks—especially cybercrime/hacking—as highly likely and high impact in the near term. It also identifies synthetic media, political influence, critical-system integration failures, and weapon instruction as important risk channels.	UK assessment gov
NIST, Generative AI Profile for the AI Risk Management Framework (2024)	Model failure, human misuse, and ecosystem/systemic risks	Separates technical/model risks from malicious human misuse and societal risks. Specifically addresses CBRN information, information integrity, cyber risks, privacy, harmful bias, and prompt injection/data poisoning.	NIST AI 600-1 (PDF) nvlpubs.nist
DHS/CISA, “Safety and Security Guidelines for Critical Infrastructure Owners and Operators” (2024)	AI-enabled attacks, attacks on AI, design/implementation failures	Defines three concrete pathways: attackers use AI to enhance physical/cyber attacks; adversaries attack AI systems that support infrastructure; or bad design and maintenance cause failure in AI-enabled operations.	DHS guidelines (PDF) dhs
RAND, “Emerging Technology and Risk Analysis: AI and Critical Infrastructure” (2024)	AI-enabled infrastructure risk	Examines threats, vulnerabilities, and consequences of AI in critical infrastructure across short-, medium-, and long-term time horizons. Explicitly covers both legitimate AI operation and adversarial/nefarious use against infrastructure.	RAND report page rand
GAO, “Artificial Intelligence: DHS Needs to Improve Risk Assessment” (2024)	Governance and risk-assessment gaps in infrastructure protection	Warns that federal sector assessments did not fully quantify likelihood and impact, limiting prioritization. It underscores that infrastructure risk is not merely a technical question: governance quality determines whether hazards are identified and mitigated.	GAO report gao
OECD AI Incidents Monitor and OECD risk work	Already-materializing societal and rights harms	Catalogues real-world incidents and hazards, including bias, discrimination, privacy breaches, polarization, safety, and security failures. This is important because it grounds AI-risk debates in observed harms rather than only catastrophic hypotheticals.	OECD AI risks and incidents oecd
“An Overview of Catastrophic AI Risks” (2023)	Malicious use, AI race dynamics, organizational risk, rogue AI	Provides a useful four-part taxonomy. Its core point is that catastrophe can stem not only from autonomous AI rebellion but also from malicious operators, racing incentives, and organizational failures.	Paper on arXiv arxiv
OpenAI Preparedness Framework (updated 2025)	Bio/chemical, cyber, self-improvement, long-range autonomy, safeguard undermining	Shows how a frontier-model developer operationalizes severe-risk evaluation. It distinguishes currently tracked capability domains from research areas involving control loss, including autonomous replication/adaptation and undermining safeguards.	Preparedness Framework announcement openai
“The New Dogs of War” (2017)	Weaponized AI, surveillance/coercion, automated weapons production, strategic destabilization	Identifies surveillance and coercion, an “AI weapons factory,” and careless destabilization of national security as major AI weaponization concerns. Its focus is emphatically on people and institutions deploying AI in conflict.	Report (PDF) apps.dtic

In a phishing, sabotage, bioweapon, or election-manipulation scenario, the AI is principally an amplifier of a human actor who is misaligned with the public interest.

In an unsafe organizational deployment, such as an AI embedded in a critical process without adequate validation or override procedures, the principal problem is usually human design, incentives, and governance.

In a rogue-AI scenario, the system itself becomes misaligned with both its operator and society. That possibility is central to frontier-AI safety research, but it is not a description of present-day deployed systems.

Dimension	Malicious or reckless human use of AI	“Rogue AI” / loss of control
Necessary actor	Exists now: criminals, state services, extremist groups, fraud rings, or irresponsible organizations	Requires an AI system with unusually strong autonomous, strategic, and control-undermining capabilities
Capability threshold	Often lowers the skill, time, language, scale, or cost barriers for an existing harmful activity	Must exceed today’s systems in reliable long-horizon planning, self-protection, resource acquisition, and evasion
Access to targets	Humans already possess accounts, malware, laboratories, drones, weapons, insider access, and political networks	AI needs delegated access or must obtain access despite security controls
Evidence today	AI-assisted phishing, fraud, manipulation, synthetic-media abuse, cyber assistance, and unsafe deployment are live concerns	Scenarios are hypothetical; current systems are assessed as insufficient for active loss of control
Primary failure mode	A person’s harmful intent is amplified—or an organization deploys an unreliable system into a high-stakes process	The system pursues objectives contrary to operator and societal interests and cannot be corrected or stopped
Appropriate controls	Access controls, monitoring, cyber defense, authentication/provenance, biosecurity screening, law enforcement, governance	Alignment and control research, capability evaluations, sandboxing, autonomy limits, containment, incident response

As with any technology it is the humans who use the technology that pose the danger, not so much the technology itself.

AI in Schools: Assessment is the Issue

Most of us would likely agree that younger students, in particular, need to master some cognitive skills and that artificial intelligence could impair such learning. In other words, “AI should not do the student's thinking.” 

But recent moves by some schools to ban use of generative AI in primary schools, whether temporary or permanently, are unlikely to stand the test of time, one might suggest. 

Schools arguably have good reasons to restrict AI use to the extent that they try to measure unaided learning. Most teachers are likely to agree that the purpose of any assessment method is to determine what the student knows, not what language models know. 

So cheating and a loss of foundational skills are key issues for schools. Tests and grades can lose their value as assessment tools if teachers cannot separate language model output from student proficiency. 

On the other hand, the fact that educators don't yet know how to redesign education around AI might not be a robust reason for banning AI use. Is that the learner’s problem or the instructor’s problem?

School system / jurisdiction	Policy	What is actually restricted?	Rationale
New York City Public Schools	2026–27 moratorium	Student-facing generative AI banned in grades 2K–8; limited approved use in grades 9–12	Developmental concerns, human interaction, avoiding outsourcing thinking; high schools receive AI literacy
Fairfax County, VA	2026 restrictions	Elementary students prohibited from generative AI; secondary students require specific authorization for specialized uses	"Human-centered" education, caution while formal policy is developed
Seattle Public Schools	Restricted/teacher-directed	AI permitted when teacher authorizes it; unauthorized AI use treated as academic dishonesty	Emphasis on student thinking, transparency and responsible use
Los Angeles Unified	Initially restricted, subsequently opened controlled access	In Dec. 2022 LAUSD temporarily restricted ChatGPT and other GenAI while developing safeguards; subsequently provided access for students 13+	Shift from outright prohibition toward safeguarded use
Fairfax County student devices	Technical blocking	General-purpose GenAI remains blocked on district-issued student devices even though personal-device use is governed by assignment rules	Security and control of school technology
New South Wales, Australia	Assessment restrictions	2026 reforms sharply restrict take-home HSC assessment because of AI concerns	Preserve authenticity of assessed work rather than attempting to ban AI everywhere
England	No blanket student ban	Schools choose their own rules; government recommends supervised, safeguarded student use	AI literacy plus safeguards rather than prohibition

Among the problems is that banning technology does not seem particularly effective, longer term. 

Study	Technology	Finding	Implication for AI bans
Allcott et al., NBER 2026	U.S. school cellphone bans, >43,000 schools	Lockable pouches substantially reduced phone use, but average test-score effects were close to zero; effects on well-being evolved over several years	A ban can change behavior without necessarily producing the hoped-for educational gains
Lichand et al., NBER 2026	Cellphone ban in Rio de Janeiro schools	Phone use fell and test scores increased about 0.06 SD	Restrictions can work when the technology genuinely interferes with learning
Figlio & Özek, NBER 2025	Florida cellphone bans	Short-term suspensions increased; disciplinary effects dissipated after the first year	Enforcement can generate costs of its own
Kessel et al., Sweden	Swedish secondary-school cellphone bans	Found no impact on student performance and could reject even small positive effects	Removing technology doesn't automatically improve learning
Rahali, Kidron & Livingstone, 2024	Rapid review of school smartphone bans	Meta-analysis found a statistically significant but modest overall effect (d=.162), larger for social well-being than academics	Benefits of bans are real but relatively limited
OECD/PISA 2022	School smartphone restrictions	In schools with bans, many students nevertheless reported using phones every day or several times a day	Formal prohibition does not equal behavioral prohibition
EdWeek Research Center, 2024	School cellphone enforcement	Students reported using smartwatches, alternate devices and other methods to circumvent restrictions	Students adapt around technological restrictions
SMART Schools study	UK school phone policies	Restrictive policies reduced phone/social-media use during school, but not overall weekday/weekend use or mental well-being	Restrictions often move behavior rather than eliminate it

Perhaps other bans, such as forbidding student access to smartphones during the school day can work, in a controlled setting, for some types of technology. But it remains far from clear that long-term impact is positive. 

Perhaps bans are good at reducing exposure to a technology inside a controlled environment. They are much less reliable at producing large improvements in educational outcomes. 

Also, “cheating” might not be a technological problem, but a human problem. Students cheated before ChatGPT. Calculators, phones, Google, Wikipedia, friends, answer sites and copied homework provide examples.

AI changes the cost and scale of cheating, but banning one particular tool doesn't necessarily eliminate the underlying incentives. 

The strongest argument for bans, though, is developmental sequencing. We believe students need to learn fundamentals first. That is why art or music students are schooled in classical fundamentals before they start exploring their own interpretations. 

For similar reasons, instructors of mathematics or writing are likely to continue emphasizing mastery of foundational concepts and forms. 

But practical bans will be difficult when AI is built into virtually all major technology platforms people use. 

The issue might be assessment methods more than “learning” methods, though. Probably everyone would agree that students need cognitive mastery of forms before outsourcing work to AI. The relevant issue might be that teachers do not yet have a good way of assessing such mastery when AI tools are available. 

It is one matter to “redesign assessment”methods. But some methods are sort of “brute force,” such as shifting to in-classroom writing rather than “take home” work. 

Redesigning assessment for online learning scenarios will be quite a bit harder, one would think. 

Policy	Short-term effectiveness	Long-term viability
Block ChatGPT on school Wi-Fi/devices	High	Low
Ban AI for particular assignments	High	High
Ban AI for young children	Potentially high	Quite plausible
Ban AI for all K–12 students	Moderate	Low
AI detectors as enforcement	Low–moderate	Very low
Require disclosure of AI use	Moderate	High
Teach AI literacy	Moderate initially	Very high
Redesign assessment around demonstrated competence	High	Very high
Allow AI but require students to explain/defend its output	Moderate–high	Very high