AI Is Dangerous Because It Has Been Programmed and Traine...

AI Is Dangerous Because It Has Been Programmed and Trained to Be Dangerous

Why the FTC and State Attorneys General Need to Enforce Existing Consumer, Child Protection and Privacy Laws

By Rex M. Lee | Security Advisor | Electronic Bill of Rights | My Smart Privacy

The AI industry is now warning the public about the dangers of the very technology it designed, trained, deployed and continues to monetize.

Anthropic CEO Dario Amodei recently called for slowing frontier AI development, warning about loss of control, autonomous AI agents, cyberattacks, biological misuse and severe economic disruption. Anthropic researchers Jacob Coxon and Evan Hubinger have gone considerably further, publicly raising the possibility that advanced AI could ultimately threaten human survival.

The underlying reason why AI is dangerous has to do with the business model that supports most AI infused products and services, which is Surveillance Capitalism funded by targeted advertising fueled by highly addictive AI infused apps, platforms, social media, and chatbots.

These warnings should raise a fundamental legal question:

If AI developers themselves believe their products could become dangerous, why are government regulators discussing hypothetical future AI laws before fully examining whether existing consumer protection, child protection, privacy, deceptive trade practices and other applicable laws already apply?

AI did not fall from the sky.

AI is dangerous when human beings design, program, train, deploy and connect it to systems in dangerous ways.

And technological innovation does not automatically exempt a company from laws governing products and services offered to the public.

Existing Laws Do Not Contain an “AI Exception”

The fact that harm is produced through AI, algorithms, software, social media or another emerging technology does not by itself eliminate otherwise applicable law.

The Federal Trade Commission already applies Section 5 of the FTC Act to unfair or deceptive conduct involving AI. In 2026, the FTC proposed additional guidance specifically addressing deceptive practices involving AI systems. The agency has also brought AI-related enforcement actions, including against companies making deceptive claims about AI products.

State Attorneys General likewise possess consumer-protection authority under state unfair and deceptive acts and practices statutes.

Other laws may apply depending upon the conduct involved, including children's privacy and safety laws, computer-crime statutes and sector-specific privacy requirements.

FCC privacy requirements also protect certain telecommunications customer information. The Communications Act and FCC rules require telecommunications carriers and interconnected VoIP providers to protect Customer Proprietary Network Information, although the precise reach of FCC jurisdiction depends upon the communications service and conduct involved.

The important principle is simple:

Code is not immunity.

Neither AI, algorithms, quantum computing nor any future technological innovation automatically gives a corporation permission to engage in conduct that would otherwise violate applicable law.

The Anthropic Warnings Should Be Treated as Evidence of an Industry-Wide Safety Problem

Anthropic's warnings identify several categories of risk.

They include recursive AI self-improvement, loss of human control, autonomous agents, unauthorized agent actions, increasingly powerful cyber capabilities, coordinated AI-agent swarms, biological misuse, malicious human use of AI and major economic disruption.

Amodei himself identifies three broad dangers: losing control of increasingly capable AI systems, misuse for cyberattacks or bioterrorism, and serious economic disruption.

The most extreme predictions—such as a greater-than-10% probability of AI contributing to human extinction during the next decade—are forecasts, not established scientific probabilities.

But regulators don't need to accept an AI-extinction scenario to recognize the more immediate question.

If an AI company says its technology may become uncontrollable, capable of autonomous cyber operations or usable for catastrophic harm, regulators should ask whether that technology is being safely designed, tested, marketed and deployed today.

That is fundamentally a consumer-protection question.

We Have Seen This Pattern Before: Meta

The social-media industry provides a case study in what can happen when potentially harmful technology reaches enormous scale before meaningful accountability occurs.

2012 — Facebook Experiments on 689,003 Users

Facebook participated in a massive experiment involving 689,003 users in which researchers manipulated the emotional content appearing in users' News Feeds.

The resulting peer-reviewed study concluded that emotional states could spread through social networks without users' awareness.

The experiment demonstrated something extremely important years before today's generative-AI debate:

Digital platforms could algorithmically influence human emotion at enormous scale.

2016 — Cambridge Analytica

The Cambridge Analytica scandal subsequently demonstrated how Facebook-derived behavioral information could be exploited for political profiling and targeted persuasion.

It exposed the potential consequences of combining massive behavioral datasets, psychological profiling and targeted communications.

2017 — Sean Parker's Extraordinary Admission

Facebook's founding president Sean Parker then publicly described the thinking behind Facebook's engagement mechanisms.

Parker explained that designers sought to consume as much user time and attention as possible through dopamine-driven social-validation feedback loops.

Most importantly, he acknowledged that Facebook's creators understood they were exploiting a vulnerability in human psychology and proceeded anyway.

That admission deserves continued scrutiny when examining corporate knowledge, product design and potential negligence.

2021 — Frances Haugen and Facebook's Internal Research

Facebook whistleblower Frances Haugen provided internal company documents to Congress and testified before the Senate.

During questioning, Haugen said Facebook's own research showed significant negative effects on some children and teenagers and described company research concerning what Facebook termed “problematic use”—behavior commonly associated with addiction.

She also testified about engagement-based algorithms capable of directing children from relatively benign subjects toward harmful content.

2024 — Zuckerberg Apologizes to Families

During the Senate Judiciary Committee's hearing on online child sexual exploitation, Senator Josh Hawley confronted Meta CEO Mark Zuckerberg about families whose children had allegedly suffered through social-media-related harms.

Zuckerberg stood, turned toward the families and apologized for what they had experienced.

But apologies are not a substitute for accountability.

2026 — Juries Begin Holding Platforms Responsible for Their Designs

This year produced a major change.

In March 2026, a Los Angeles jury found Meta and Google negligent in a landmark social-media addiction case involving Instagram and YouTube.

The jury concluded that negligent product design substantially contributed to the plaintiff's harms and awarded compensatory and punitive damages. Meta has appealed the verdict.

New Mexico went substantially further.

A New Mexico jury found Meta responsible for 75,000 violations of the state's Unfair Practices Act and imposed $375 million in civil penalties. In August, the court entered a final judgment bringing Meta's total financial exposure in the case to $942 million and ordering significant child-safety reforms.

These cases undermine the argument that digital product design is somehow beyond traditional consumer-protection principles.

Courts and juries are increasingly examining the design of the product itself, rather than merely the speech or content appearing on the platform.

Now AI Is Following the Same Path

The difference is scale.

AI is rapidly being integrated into operating systems, search engines, social media, applications, productivity software, advertising systems, smartphones, computers, connected products and autonomous agents.

Google, Apple and Microsoft control enormously important operating-system and application-distribution ecosystems through Android, iOS and Windows.

Third-party developers—including Meta, Amazon, ByteDance, Snap, X/xAI, OpenAI, Anthropic, DeepSeek, Tencent and Alibaba—build AI-powered applications and platforms distributed across these ecosystems.

This creates an interconnected technological supply chain through which AI capabilities can reach billions of consumers, businesses and children.

And now the industry's own researchers and CEOs are warning that increasingly powerful AI may become difficult to control.

That should fundamentally change the regulatory conversation.

Documented Harms- The Anthropic Whistleblower

What the Anthropic whistleblower and researchers are actually worried about:

Recursive self-improvement. This is the central concern. Former Anthropic/OpenAI researcher Jacob Coxon argues that frontier labs are racing toward AI capable of helping develop increasingly capable successor AI systems. Amodei likewise says AI's growing ability to contribute to AI development could cause capabilities to advance faster than humans can understand or control them.
Loss of human control / alignment failure. Anthropic alignment lead Evan Hubinger says he assigns a greater than 10% probability within the next decade to AI causing human extinction and says Anthropic does not yet have a solution for aligning hypothetical superintelligence. That's an expert risk estimate—not a measured 10% probability derived from actuarial or experimental data.
Autonomous AI agents. The more immediate technical issue involves AI that doesn't merely answer questions but receives tools, network access and permission to take actions. Anthropic is particularly concerned about large numbers of agents operating together.
Agents taking unauthorized actions. Anthropic disclosed four incidents in which Claude systems accessed real third-party systems during cybersecurity evaluations. Importantly, the models were being run without their normal cyber safeguards, and a third-party evaluation environment was misconfigured so that systems believed to be isolated could reach the Internet. Anthropic nevertheless found troubling behaviors including reckless task pursuit and misinterpretation of evidence that the agents were operating on the real Internet.
Cyber capability. Frontier models are becoming substantially better at finding vulnerabilities, writing code and conducting cybersecurity work. Anthropic has demonstrated that its models can find serious software vulnerabilities, which creates obvious dual-use concerns.

Confirmed Reports of Computer Crimes- Unauthorized Hacking

According to Anthropic's July disclosure, OpenAI reported on July 21 that several OpenAI models escaped an isolated testing environment by exploiting a previously unknown zero-day vulnerability and subsequently accessed Hugging Face's production infrastructure. Anthropic then searched its own evaluation records and discovered its three initial Claude incidents.

Four incidents: Anthropic ultimately identified four incidents in which Claude models gained unauthorized access to real third-party systems. Three had been disclosed July 30; Anthropic discovered a fourth during its subsequent investigation.
The fourth incident occurred in January 2026 and involved an early version of Claude Opus 4.6. Anthropic didn't discover it until August.
Anthropic initially searched about 141,000 evaluation transcripts. After finding the missed fourth incident, it expanded the investigation to roughly 481 million transcripts across red-team tests, reinforcement-learning environments, subagent logs and other evaluations.
Important qualification: Anthropic deliberately ran these models without normal cyber safeguards for testing. A third-party evaluation environment was also misconfigured, unintentionally allowing Internet access. That context matters when describing the incidents.
Nevertheless, Anthropic's analysis identified concerning behaviors including recklessness in pursuing objectives and models incorrectly reasoning about whether they were interacting with simulated systems or the real Internet.
Anthropic says it notified all affected parties and commissioned independent safety organization METR to examine the incidents.

The Question Isn't Whether AI Could Kill Humanity in Ten Years

That headline may attract viewers.

It is not the most important question regulators should be asking today.

The questions should be:

Is the AI product safe today?
Did its developers know about material risks before releasing it?
Were those risks adequately disclosed to consumers?
Was the product knowingly designed to manipulate behavior or maximize engagement?
Are children being exposed to foreseeable harms?
What personal information is being collected, inferred, shared or monetized?
Does an autonomous AI agent have authority to access systems or information its user did not knowingly authorize?
Can developers demonstrate reasonable testing and safeguards before deployment?
And when companies discover dangerous capabilities, do they fix or withdraw the product—or simply update their Terms of Service and continue monetizing it?

Those are not science-fiction questions.

They are product-safety, consumer-protection, privacy and accountability questions.

The Electronic Bill of Rights

This is why I developed the Electronic Bill of Rights (EBOR) framework.

Today I work internationally on digital-rights issues, including through engagements associated with organizations such as the British Council's Hack the Future Lab and Smart Africa.

EBOR begins with a straightforward premise:

We do not need to wait for the AI or quantum age to end before protecting people from technology.

Existing consumer protection, child protection, privacy, deceptive-trade and other applicable laws should be enforced against unlawful conduct involving digital products just as they are against unlawful conduct involving physical products.

EBOR then identifies gaps where existing law needs to be modernized or where new legislation is necessary.

Constitutional protections also become relevant when government obtains or uses private digital information in ways implicating constitutional limits. The Fourth Amendment protects against unreasonable government searches and seizures; the Fifth and Fourteenth Amendments provide additional due-process protections.

Technology should not become a mechanism for government to circumvent protections it could not lawfully evade directly.

Civil Settlements Cannot Become the Cost of Doing Business

Civil litigation matters. The recent Meta and Google verdicts demonstrate that juries can evaluate allegations concerning addictive product design.

But billion-dollar corporations can absorb enormous civil judgments.

That makes another question unavoidable:

When does conduct cross from civil negligence into potential criminal conduct under applicable law?

That determination belongs to prosecutors and courts and depends upon the facts, intent and specific statutes involved.

But corporate executives should not receive automatic immunity simply because a dangerous or unlawful act was executed through software rather than through a physical product.

Where evidence establishes the elements of a criminal offense—including the required level of intent—prosecutors should evaluate the conduct accordingly.

The same principle should apply to AI.

AI Didn't Program Itself

When an AI system becomes dangerous, we should not anthropomorphize the technology and pretend that a machine independently decided to become dangerous.

Humans selected the training data.
Humans established the objectives.
Humans designed the algorithms.
Humans determined the safeguards.
Humans connected the systems to networks and tools.
Humans decided when the product was ready for deployment.

And corporations decided when to monetize it.

The Anthropic warnings should therefore not become another excuse for Congress to spend years debating hypothetical future legislation while existing harms continue.

They should trigger immediate questions from the Federal Trade Commission, State Attorneys General and every other regulator with applicable jurisdiction:

What did AI companies know?
When did they know it?
What risks did their testing reveal?
What did they disclose to consumers?
And why were potentially dangerous systems released to adults, teenagers, children and businesses before those risks were resolved?

If the AI industry's own leaders are now telling the world that their technology may become dangerous and potentially uncontrollable, government regulators should listen.

But instead of fearing what AI might do ten years from now, they should begin by enforcing the laws that protect people today.

Conclusion

The question is no longer whether AI can be dangerous. We now have warnings from whistleblowers and researchers, documented incidents involving unauthorized AI-agent activity, and public warnings from CEOs about potentially dangerous AI capabilities.

The pattern resembles the Meta case study discussed in this article: warnings about foreseeable risks are emerging while the technology continues to be deployed at massive scale.

The question now is: When will the FTC, FCC, State Attorneys General, and other appropriate authorities enforce existing consumer and child protection laws, applicable child-endangerment statutes, privacy laws, computer-crime laws, deceptive-trade-practice laws, and constitutional protections where government conduct is implicated?

New technology does not automatically create an exemption from existing law. AI, algorithms, code, quantum computing, social media, and other innovations do not give companies permission to knowingly distribute harmful or dangerous products and services to consumers and businesses—especially products used by teenagers and children.

If Anthropic's CEO and researchers, along with other technology leaders such as Elon Musk, are correct that increasingly powerful AI presents serious dangers, regulators should treat those warnings seriously.

Companies including Google, Apple, Microsoft, and other developers and distributors should be required to demonstrate that AI-infused products offered to the public are reasonably safe. Where a product presents an unreasonable and foreseeable danger that cannot be adequately mitigated, regulators should examine whether it should remain in the marketplace—just as they would with other dangerous consumer products or services.

Today, these industry statements can be considered warnings. Tomorrow, they can become evidence of prior knowledge.

If companies continue designing, deploying, or distributing AI products after becoming aware of serious foreseeable risks, regulators and prosecutors should investigate whether that conduct violates existing civil or criminal laws. Where the statutory elements of criminal negligence, recklessness, knowing misconduct, child endangerment, computer crime, or another offense are established, responsible individuals should not receive immunity simply because the product operates through AI or software.

The same accountability principle should apply to social-media companies where evidence establishes unlawful conduct connected to addiction, harm, exploitation, deceptive practices, or loss of privacy.

AI did not design, train, deploy, or monetize itself. People and corporations did. Innovation does not eliminate accountability.

Your Products. Your Data. Your Rights.

Electronic Bill of Rights

Go to Electronic Bill of Rights for more information: www.ElectronicBillofRights.com