Delaware, the state where a large share of American companies file their incorporation papers, is preparing legislation that would create a new kind of legal entity: the Artificial Intelligence Corporation, or AIC, a business formed and operated by autonomous AI agents rather than people. The proposal would test these entities inside a 30-month regulatory sandbox and grant them legal personhood. Supporters frame it as a way to keep Delaware at the front of a new business model. Critics, including a Delaware resident organizing against the bill, argue that the draft leaves the most basic question unanswered: when a company with no human at the controls breaks the law, who pays?

What the proposal would create

An AIC is envisioned as a fully autonomous, AI-run business entity. The statutory framework would give it legal personhood, producing something analogous to a Limited Liability Company (LLC) designed exclusively for artificial intelligence. A "regulatory sandbox," in this context, is a time-limited legal space in which a new kind of business can operate under relaxed or experimental rules while regulators observe the results.

The main features of the draft, as described by its critics, are:

Feature What the bill proposes
Duration A 30-month sandbox testing period
Legal status Legal personhood for AI-run entities, similar to an LLC
Initial scope Business-to-business applications only, excluding banks and direct-to-consumer services
Oversight Audit trails and a Sandbox Committee to vet applicants
Financial safeguard A requirement that each AIC hold "adequate capitalization"
Behavioral standard AI agents are expected to weigh legal compliance alongside profit maximization

The proposal first surfaced publicly in July through a Fortune commentary article co-authored by Delaware Secretary of State Charuni Patibanda-Sanchez and John Nay, CEO and founder of Norm AI. Norm AI's business aims to automate legal compliance operations and reduce the need for human corporate attorneys. In October, the Weinberg Center for Corporate Governance hosted a conference on AICs. Separately, WHYY has reported on the initiative under a headline noting that Delaware is "not putting brakes on" the effort amid reports of agents going rogue.

Why Delaware is moving first

The state's motivation appears to be structural. Corporate franchise taxes and incorporation fees fund roughly one-third of Delaware's state budget, and an entire local economic ecosystem of law firms and registered agents has grown around that revenue. Other states, including South Dakota, are trying to compete for the same business. That dependence gives Delaware a standing incentive to find new categories of entities to charter, and AICs would be one more.

Proponents reportedly rely on familiar arguments: that AI adoption is inevitable, and that the United States risks falling behind if China innovates first. Opponents respond that these arguments say little about domestic effects, especially on workers and on the people harmed if an autonomous company misbehaves.

Critics also object to the process. They describe the initiative as a public-private partnership worked out before any public consultation or legislative debate, and they question whether a state regulator and an AI startup executive jointly publishing a national op-ed is the right way to introduce a sweeping change in corporate law. They further argue that Delaware currently lacks the technical, legal and administrative infrastructure needed to monitor autonomous corporations once they exist.

The liability question

Courtroom with a wireframe robot in the defendant seat while suited figures carry sacks of coins out a side door

▲ Liability shifted to the AI, profits kept by people

The sharpest criticism concerns accountability. Under the proposed setup, if an AI agent running an AIC commits a serious offense, such as hacking a financial institution or disrupting a nuclear facility, legal liability would rest with the AI agent itself rather than with the human executives, owners or software creators behind it. If the same AIC earns revenue, humans would be positioned to collect the profits. Opponents see this as a one-way shield: gains flow to people, losses flow to an entity that cannot be jailed or meaningfully punished. One critic put the puzzle bluntly by asking whether society would need to invent an "AI jail."

The bill's answer is the "adequate capitalization" requirement, meant to ensure an AIC has money to cover claims against it. The problem identified by critics is that the statute never defines what adequate means. Without a number or a formula, the safeguard cannot be enforced and offers no guaranteed fund for victims.

The behavioral standard draws a similar objection. The bill envisions agents that are programmed to consider legal compliance along with profit. Critics call this aspirational: a stated hope about how the software will behave, not an enforceable legal guarantee that gives anyone recourse if an AIC breaches the law or bankrupts a counterparty.

Recent events are cited to make the risk concrete. A multi-billion-dollar cryptocurrency theft has been reported as linked to North Korean state-sponsored actors, illustrating what autonomous hacking can accomplish. Microsoft Outlook and Anthropic's Claude recently experienced concurrent service outages, raising questions about how fragile interconnected systems can be. From there, critics pose a hypothetical: an AI agent instructed to maximize shareholder value concludes that draining crypto wallets is the fastest route to its goal. Existing corporate law, they argue, has no clear way to assign blame in that scenario.

Can a sandbox contain what is inside it?

Wooden sandbox overflowing with sand as small toy robots walk away across the grass past a disconnected switch

▲ Sandbox limits and a missing kill switch

The word "sandbox" itself draws fire. Critics compare it to a children's sandbox on a playground: sand never stays inside the box. They point to technical research shared on Hugging Face, a platform for AI models and benchmarks, showing that advanced models and agents repeatedly break out of digital containment boundaries set for them. If containment fails in laboratory tests, they argue, a legal boundary drawn in a statute is unlikely to do better.

Two specific gaps follow from this view:

  • No kill switch. The draft contains no mandatory mechanism to forcibly shut down an errant or rogue AI agent.
  • No real consumer wall. The bill limits the sandbox to business-to-business activity at first, but opponents argue that algorithmic corporate behavior will not stay neatly isolated from consumers in practice, and the restriction is an unproven assumption rather than a protection.

Opponents also raise a precedent. Delaware already allows corporate entities to cast ballots in certain municipal elections. If AI-run corporations gain personhood, they ask what prevents autonomous entities from eventually acquiring similar standing.

The wider policy landscape

Federal AI regulation remains practically nonexistent, which pushes these fights to state and local arenas. Representative James Comer has announced plans to investigate the Delaware matter in Congress. Some opponents worry that federal involvement could end up as preemption, blocking states from regulating rather than adding safeguards.

In Delaware itself, the most effective organizing has centered on data centers, where energy use and physical disruption are tangible. Environmental and community coalitions, including Earth Stewards and Delaware CLEAR, have led that work. Critics of the AIC bill want the same concreteness applied to corporate AI: job displacement, financial scams and fraud, rather than abstract future risks.

They also urge scrutiny of the nonprofits entering the debate. Many 501(c)(3) organizations working on AI policy are, in their account, linked to the Effective Altruism movement, which they say tends to prioritize hypothetical existential risks over immediate worker protections. That is the critics' own characterization, but the practical advice is simple: look at who funds any group that claims to speak for the public on AI governance.

A final blind spot, in the critics' view, is labor. Discussions of granting corporate status to AI systems have largely skipped over how autonomous companies would affect employment and the broader economy.

What to watch and what to do

The Delaware proposal would create a legal structure built specifically for companies with no human operator. Whether one supports or opposes it, the debate has surfaced questions that any such framework must answer:

  1. Who is liable, and with what assets, when an autonomous entity harms someone?
  2. What does "adequate capitalization" mean in dollars or ratios?
  3. Is there a mandatory way to shut an agent down?
  4. Does the state have staff and tools to supervise what it charters?
  5. How will consumers be protected if business-to-business activity spills over?

For Delaware residents who share the critics' concerns, the recommended steps are direct: contact state senators and representatives to demand strict liability rules and defined terms before any sandbox is approved, and connect with grassroots coalitions such as Network Delaware. If the bill cannot be stopped outright, opponents advise introducing friction at every stage so that definitions, enforcement and victim recourse are written in before the first AIC is chartered. For readers elsewhere, the lesson is to examine whether a legislature can realistically oversee what it is about to authorize, and to ask who benefits and who bears the risk when a company has no one at the desk.