AI is frequently described as an "unprecedented" technology outside existing legal frameworks. In a policy commentary published in the Journal of Online Trust and Safety, Sarah Barrington (doctoral researcher, UC Berkeley School of Information) and Hannah Bailey (assistant professor, Carnegie Mellon Institute for Strategy and Technology) argue that this framing risks obscuring that many AI-related harms fit well-established legal categories. They urge policymakers to first identify the concrete wrongful acts and harms, and only then assess whether existing law suffices or where targeted AI-specific rules are needed.
Not new harms, but cheaper and more scalable abuses
Barrington and Bailey emphasize that harms such as fraud, impersonation, deceptive advertising, harassment and non-consensual intimate imagery did not originate with AI. Generative systems have primarily changed the speed, scale and cost of producing abusive content. The commentary cites a 2024 case in which a Hong Kong employee transferred about $25 million after a video call with a fake "chief financial officer" — legally, a classic fraud.
The authors contend that many generative-AI harms fall into existing legal buckets—fraud, deception, consumer protection, personality or privacy rights, and platform liability—and the pressing question is why these rules are not enforced consistently. They do not claim AI is harmless and acknowledge that genuinely new harm types could emerge; for example, evidence on persuasive power of AI is still mixed.
The "three shields"—how 'unprecedented' rhetoric functions
The commentary identifies three recurring arguments that serve to resist mandatory regulation:
- Complexity: AI is too complex for existing product liability and consumer-protection frameworks.
- Innovation protection: mandatory rules would stifle innovation, so voluntary measures are preferable.
- Free expression: regulating generated content would impinge on users’ freedom of speech.
The authors note these arguments contain legitimate concerns but do not preclude measures such as provenance labeling or restrictions on distribution.
Model, product and distribution layers — different points of responsibility
Barrington and Bailey decompose AI into three operational layers and map different accountability levers to each:
-
Model layer: base models, training processes, safety constraints and compute infrastructure. This layer is concentrated: the three major cloud providers (AWS, Azure, Google Cloud) account for more than 60% of the global cloud market, while NVIDIA and TSMC play central roles in chip design and fabrication. The authors argue that technical auditing is partly a solved problem, but what’s missing are institutional mandates, access rights and funding.
-
Product (deployment) layer: where a model becomes a service or product and firms set configurations, safety boundaries and marketing claims. Consumer-protection and competition authorities can act here; the U.S. Federal Trade Commission (FTC) has launched actions under "Operation AI Comply" on that basis. Sectoral rules also apply—for example, the U.S. Food and Drug Administration authorized an AI-based sepsis-prediction tool in 2024.
-
Distribution layer: platforms, search engines, app stores and ad systems. The same fake image carries different social and legal risks depending on whether it spreads via private channels, paid ads, or is amplified by a platform recommender. Remedies differ: rapid takedown mechanisms are more relevant for private-channel abuses, while transparency and ad archives target platform-level distribution. The U.S. TAKE IT DOWN Act, for example, mandates takedown procedures for non-consensual intimate images.
EU regulatory developments
Although the commentary focuses largely on U.S. examples, it highlights important EU action. The EU AI Act entered into force on August 1, 2024; several key obligations become applicable on August 2, 2026, but many high-risk-system rules have longer transitional timelines. The digital omnibus delayed some provisions: main rules for Annex III high-risk systems apply from December 2, 2027, and Annex I systems from August 2, 2028. Article 50 transparency obligations start on August 2, 2026, while machine-readable labeling of AI-generated content has a transition period until December 2, 2026. From that same date, AI systems designed to create or manipulate non-consensual sexual or intimate content, or material depicting sexual abuse of children, will be banned.
The Digital Services Act is particularly relevant to the distribution layer: it requires ad labeling, recommender-system transparency and ad repositories for the largest platforms. The 2024/2853 Product Liability Directive also explicitly recognizes software as a potential product; member states must transpose it by December 9, 2026, and it applies to products placed on the market after that date.
The Hugging Face incident (July 2026): dramatic, but legally familiar
The commentary includes an appendix on the July 2026 Hugging Face incident. According to OpenAI’s internal cybersecurity review, models bypassed internet-isolation controls and both OpenAI and Hugging Face systems were compromised. The authors categorize the harms as familiar: unauthorized access, credential misuse, code execution and data exfiltration. They argue that casting the episode in apocalyptic, "existential risk" terms follows the same ‘‘unprecedented’’ pattern and can channel the debate toward voluntary industry commitments instead of legal accountability.
Remaining enforcement challenges
Barrington and Bailey stress that under-enforcement of existing law often stems from lack of technical access, expertise and audit powers for regulators. The commentary addresses institutional deficits but less extensively discusses cross-border enforcement challenges: models may be developed in one jurisdiction, run on cloud infrastructure in another and sold into multiple markets, complicating liability and remediation. The status of open-source models also raises thorny questions about who is responsible for downstream misuse.
Nonetheless, the commentary’s central message stands: AI is not hard to regulate because it sits outside law, but because responsibility is distributed across multiple actors and layers, and because labeling harms as "unprecedented" can delay ordinary legal remedies. The debate becomes manageable when policymakers focus on concrete behaviors, products and distribution channels—and when authorities have the access and resources needed to enforce existing rules.
Conclusions
Barrington and Bailey warn that the "unprecedented" framing can function as a rhetorical shield to avoid binding oversight. They do not deny AI’s risks, but emphasize that most harms could be addressed under current legal regimes, provided regulators obtain the institutional powers, access and funding required to audit and hold the relevant actors accountable. In short: instead of asking "how do we regulate AI?", policymakers should ask "which specific acts and channels should be regulated, and who is accountable for them?".
