UNREGULATED AI: A CLEAR AND PRESENT DANGER

Artificial intelligence isn’t one invention with one neat set of consequences. The label covers everything from movie recommendations and medical tools to autonomous weapons and software agents that can browse websites, send messages, run code, and spend money.

We shouldn’t treat those systems alike. A chatbot inventing a citation is frustrating. A clinical tool inventing a medical fact could kill someone. A biased entertainment recommendation might narrow a person’s choices; a biased government risk score could cost them their freedom.

So the answer isn’t panic. It isn’t blind trust in the companies building these systems, either. The sensible approach is to match safeguards to the severity, scale, and reversibility of the potential harm.

AI Can Be Wrong Without Looking Wrong

Generative AI has a dangerous habit: it can present false information in polished, confident language. Researchers often call these errors “hallucinations.” A study presented at the 2024 Annual Meeting of the Association for Computational Linguistics examined hallucinations across models and tasks, reinforcing the point that factual unreliability is not a rare cosmetic flaw. It is a persistent technical problem that requires serious evaluation (Li et al., 2024).

Older software usually fails in obvious ways. A page crashes. A calculation returns an error—a database request times out. Generative AI can fail by giving an answer that sounds perfectly reasonable. Unless the user already knows the subject, there may be no clear sign that anything has gone wrong.

The consequences depend on the setting. A false claim in a brainstorming tool may waste a few minutes. The same behavior in medicine, law, finance, education, or public administration could lead to harmful treatment, wrongful accusations, financial losses, or denied public benefits.

That is why a single accuracy score tells us so little. A review by Kenthapadi, Sameki, and Taly argues that hallucinations differ in type and severity, so evaluations must account for the context and consequences of an error—not merely count how often a model gets something wrong (2024).

The starting rule should be simple: no AI system should make a consequential decision merely because its output sounds convincing. High-risk systems need reliable source data, independent validation, uncertainty reporting, qualified human review, and a genuine appeals process.

Privacy Risks Continue After Training

AI companies often build systems using enormous collections of text, images, audio, and behavioral data. Some of that material contains personal or sensitive information. Once the data has been absorbed into a model—or connected through an external retrieval system—it may be difficult to locate, correct, or remove.

The threat goes beyond models memorizing their training data. Retrieval-augmented generation systems pull information from external databases as they respond to users. Ni et al. identify privacy threats across this entire retrieval pipeline, meaning confidential information can leak from the surrounding data system as well as from the model itself (2026).

Medical applications make the stakes especially clear. In developing a large-language-model system to extract information from clinical reports, Posta et al. identified data security, privacy law, and governance as major deployment barriers (2026). Health records are unusually sensitive, but the same concerns apply to employment files, school records, legal documents, biometric data, and private communications.

Organizations should collect less data, retain it for shorter periods, and tightly control access. Sensitive information should be encrypted and kept out of general-purpose model training unless people have given meaningful consent. Privacy audits should test whether individual records can be recovered from a model or retrieval system.

People also need enforceable rights. They should be able to find out whether their data was used, correct inaccurate records, and request deletion where removal is technically and legally possible. A privacy policy buried behind a link is not meaningful control.

Automated Decisions Can Scale Discrimination

AI systems learn from historical data, and history contains plenty of unequal treatment. Feed those patterns into hiring, lending, policing, insurance, healthcare, or welfare systems, and discrimination can become faster, cheaper, and much harder to see.

Removing protected characteristics does not necessarily solve the problem. Postal codes, browsing histories, employment gaps, shopping behavior, and social connections can all serve as proxies for race, gender, disability, income, or other sensitive traits. A system can therefore produce discriminatory outcomes without explicitly using any protected category.

Scale makes the danger worse. A prejudiced individual might harm dozens of people. A widely deployed model can apply the same distorted rule to millions. And because automated decisions are wrapped in mathematics and technical language, institutions may treat them as more objective than they really are.

Before deploying a high-risk system, companies and public agencies should be required to conduct an impact assessment. That assessment should examine performance across relevant demographic groups and measure real outcomes—not just average technical accuracy. Regulators should have access to the results, and the public should see them whenever possible.

If an organization cannot explain the basis of a consequential decision or provide a meaningful appeal, it should not automate that decision.

Surveillance deserves even stricter treatment. Manan and Zakir argue that AI-assisted surveillance requires strong controls, transparency, and accountability because intrusive or biased systems can threaten privacy and other human rights (2025). Patil et al. similarly recommend mandatory human-rights impact audits for high-resolution satellite surveillance (2025).

Governments should prohibit indiscriminate biometric surveillance. Narrower uses should require legal authorization and proof that they are necessary and proportionate. “The technology was available” is not a legitimate reason to monitor everyone.

AI Makes Abuse Cheaper and Easier

AI does not need to invent a new crime to be dangerous. Making old crimes faster, cheaper, and easier to personalize is enough.

Language models can generate persuasive phishing messages at scale. Voice-cloning systems can impersonate family members, executives, or public officials. Automated tools may help attackers research targets, adapt malicious software, or identify vulnerable systems.

Brundage et al. warned that AI could expand existing security threats, introduce new ones, and change the character of attacks by increasing their scale and lowering the expertise required to carry them out (2018). That warning has only become more relevant as capable models have become easier to access.

AI can help defenders too. It can detect fraud, analyze malware, and speed up incident response. The problem is timing: offensive capabilities may spread before governments, companies, and ordinary users have adequate defenses.

Providers should evaluate powerful models for cyber misuse before release. When a system crosses a meaningful capability threshold, stronger controls may be necessary. These could include identity verification for dangerous functions, rate limits, monitored interfaces, secure activity logs, and restrictions on model weights when unrestricted release would create a serious, foreseeable risk.

Security researchers also need legal protection when they report vulnerabilities in good faith. Providers, in turn, should face firm deadlines for fixing serious defects.

Autonomous AI Raises the Stakes

The risk changes when AI stops merely offering advice and starts taking action. Do you remember the movie Patriot Games? The reality can be much worse.

Software agents can browse the web, send messages, execute code, purchase goods, and operate other digital tools. Embodied systems can act in the physical environment. Once a model can do things rather than describe them, a mistake can become an event.

An agent might misunderstand its goal, call an unsafe tool, reveal confidential information, or take a sequence of individually plausible steps that produces a disastrous result. Wu et al. describe how security risks evolve across language models, software agents, and embodied systems, combining perception errors, unsafe actions, and wider systemic effects (2026).

Vijayvargiya et al. also found that models struggle with systemic risks requiring more than a narrow, one-step safety check (2026). This exposes a weakness in many current evaluations: they test whether a model will produce an obviously harmful response, not whether it could cause harm over time while pursuing a broader goal.

Agents should therefore have limited permissions by default. Sensitive actions should require confirmation. Systems need spending caps, restricted tool access, tamper-resistant logs, continuous monitoring, and dependable kill or disconnect switches that are regularly tested under realistic emergency conditions.

Testing must reflect how these systems will actually be used. Short laboratory prompts are not enough. If an agent can affect critical infrastructure, financial markets, weapons, or large numbers of people, independent evaluation should be mandatory before deployment.

Lethal Decisions Must Remain Human Decisions

Autonomous weapons create the clearest need for a hard boundary.

A system that selects and attacks a target without meaningful human control raises legal, ethical, and security problems that ordinary product rules cannot address. This is not a question of whether the software offers a useful recommendation. It is a question of whether a machine should receive the final authority to take a human life.

Amoroso and Tamburrini connect meaningful human control over autonomous weapons to accountability, human dignity, and compliance with legal and ethical obligations (2020). Longpre, Storm, and Shah argue that international coordination is necessary because autonomous weapons may become increasingly accessible and cannot be governed effectively through unilateral action alone (2022).

Governments should negotiate binding international restrictions on weapons that select and engage human targets without meaningful human judgment. At a minimum, a human operator must understand the system’s limits, evaluate the context, approve any use of lethal force, and remain legally accountable for the decision.

Machines should not decide who lives and who dies.

Economic Harm Is Also a Safety Issue

Job displacement is often treated as separate from AI safety. It shouldn’t be.

Losing a job can threaten a person’s health, housing, family stability, and trust in institutions. A sudden wave of displacement can damage entire communities. Even if automation raises total productivity, the benefits may flow mainly to the owners of capital unless policy deliberately spreads them.

The evidence does not support the simplistic claim that AI will eliminate all work. Technology can replace tasks while creating new jobs and increasing demand elsewhere. The real questions are who bears the disruption, how quickly it happens, and who receives the gains.

Research by Solos and Leonard considers both labor displacement and rising inequality, suggesting that aggregate economic growth alone is a poor measure of whether AI’s economic effects are socially beneficial (2022).

Workers should receive advance notice when employers plan major AI-driven restructuring. Governments should strengthen unemployment insurance, wage insurance, portable benefits, retraining, and job-placement services. Employers should be encouraged to use AI to support workers rather than treating head-count reduction as the only measure of success.

Competition policy matters too. If a handful of companies control the leading models, computing infrastructure, and data, they can capture a disproportionate share of the gains—and gain enormous influence over the rules meant to govern them.

What Responsible AI Regulation Should Do

Good regulation should focus on actual risk. A photo filter and an autonomous weapons platform do not need the same rules.

Low-risk products may require little beyond ordinary consumer protection. Systems used in employment, credit, healthcare, education, insurance, policing, critical infrastructure, or government benefits need stricter requirements. These should include documented risk assessments, security testing, sound data-governance practices, incident reporting, independent audits, human oversight, and clear appeals procedures.

Developers of the most capable general-purpose models should have additional duties. Before deployment, they should evaluate dangerous capabilities, privacy leakage, deceptive behavior, foreseeable misuse, and failures that may appear when a model is connected to tools.

Evaluations should also examine behavior over time. It is not enough to ask what a model can do when directly prompted. Regulators and researchers need to know what it is inclined to do while pursuing a goal across an extended interaction. PropensityBench, for example, was developed to investigate latent safety tendencies through agentic evaluation (Sehwag et al., 2026).

Independent evaluators need secure access to models and relevant technical documentation. Regulators need trained staff, computing resources, and the authority to investigate serious incidents. Whistleblowers need protection. Companies should be required to report major failures promptly, just as firms in other regulated industries must report breaches, defects, and accidents.

Liability is equally important. If a company deploys an inadequately tested system and people are harmed, responsibility should not vanish into a maze of vendors, models, and subcontractors. The law should specify who is accountable for training, integration, deployment, and operation.

Companies that follow rigorous standards may deserve limited legal protection. Companies that hide risks, mislead regulators, or ignore foreseeable harms should face meaningful penalties.

Some uses should be banned. These include indiscriminate biometric surveillance, manipulative systems designed to exploit vulnerable people, certain forms of automated social scoring, and lethal systems operating without meaningful human control.

The Public Needs Power, Not Reassurance

The deepest weakness in AI governance is simple: the people affected by automated decisions often know the least and control the least.

Someone may never learn that an algorithm screened their résumé, flagged their insurance claim, ranked their child, denied them a loan, or placed them under suspicion. Even when they know AI was involved, they may have no practical way to challenge the result.

Basic procedural rights would help correct that imbalance. People should be told when AI materially influences a consequential decision. They should receive an understandable explanation, be able to correct inaccurate data, and have access to a qualified human reviewer with the power to reverse the result.

Communities should also have a say before public agencies introduce surveillance or automated decision systems—not after the damage is done.

Public-interest researchers need access to study deployed systems. Without independent scrutiny, society must rely on the claims of the same organizations selling the technology. Trade secrecy can protect legitimate intellectual property, but it should not become a blanket excuse for hiding discrimination, unsafe behavior, or violations of law.

AI can do real good. It can help doctors analyze medical data, make information more accessible, automate repetitive work, and give people useful new tools. Taking its risks seriously does not diminish those benefits. It helps preserve them.

The goal is not to make AI perfectly safe. No powerful technology is risk-free. The goal is to stop companies and governments from shifting the risks onto the public while keeping the rewards for themselves.

That requires enforceable standards, independent testing, strong privacy rights, meaningful human control, clear liability, and international rules for the most dangerous systems. It also requires the courage to say that some technologies are not ready—and that some uses should never be allowed.