What Is Shadow AI? Corporate Liability Risks and Who Is Accountable

What is shadow AI, and who is legally accountable for the risks it creates?

Shadow AI is the set of AI tools, scripts, and agents employees build or use without approval, and the liability it creates lands on executives who make AI claims, on the company as an entity, and on wherever the data ingestion happened. AI corporate liability risk is not theoretical anymore. Executives have been criminally charged for lying about AI capabilities. Courts are letting new legal theories against AI-scraped data survive motions to dismiss. State and international regulators have built penalty regimes that reach hundreds of millions of dollars. The part most companies still haven't connected: the shadow AI tools employees build without approval can trigger all three kinds of exposure at once, and nobody has claimed ownership of that risk.

Here's how that happened, and where the liability actually lands.

What does shadow AI actually mean?

Shadow IT was employees signing up for software the company never approved: a rogue Dropbox account, an unsanctioned Slack workspace. Shadow AI is different. It's employees using unapproved AI tools, chatbots, agents, and automations, often through personal accounts, to do real work with company data.

The difference that matters legally is what these tools do with what they touch. A shared file link is passive. An AI tool ingests, retains, and sometimes trains on whatever gets pasted into it. Citizen-development trends were already collapsing the line between employees using software and employees building it. AI accelerates that: an ordinary employee can now spin up a scraper, an agent, or a data pipeline with the same effective reach as an engineered system, minus any of the oversight.

The canonical case is still Samsung. In a single 20-day window in 2023, semiconductor engineers pasted confidential source code and meeting notes into ChatGPT three separate times. Samsung's response was to ban generative AI tools company-wide. That wasn't a hacking incident. It was employees trying to work faster, using a tool that quietly became a data-retention liability the moment they hit enter.

How big is the problem, really?

The gap between what executives believe and what employees actually do is the whole story here.

A global Okta survey found that 90% of executives are confident in their organization's visibility into AI tool usage, and 95% believe employees are using AI responsibly. Meanwhile:

• 52% of employees admit to using AI tools without approval, often through personal accounts.
• 58% of executives reported an AI-related security incident or near-miss in the past year.
• Of employees using unapproved tools, 54% share internal messages or emails, 45% share HR-related information, and 39% share confidential company documents.
• Only 34% of organizations apply the same security controls to AI agents that they apply to human employees.

A separate Gartner survey of 302 cybersecurity leaders found that 69% of organizations suspect or have evidence that employees are using prohibited public generative AI tools, yet only about 37% have a formal AI governance policy. Gartner projects that more than 40% of enterprises will experience a security or compliance incident tied to shadow AI by 2030.

This isn't a small population of reckless employees. It's a majority behavior pattern running underneath a governance structure built for a much smaller problem.

Can an executive actually go to jail for AI?

Yes. It has already happened. In April 2025, the SEC and the DOJ's Southern District of New York office filed parallel civil and criminal actions against Albert Saniger, founder and former CEO of the shopping app Nate Inc. Saniger had told investors his app completed online purchases autonomously using AI, machine learning, and neural networks, with automation rates between 93% and 97%. The actual figure, according to prosecutors, was effectively zero. Contract workers in the Philippines and elsewhere were processing the transactions by hand.

Saniger raised more than $42 million from investors on the strength of those claims. He is charged with securities fraud and wire fraud, each carrying a maximum sentence of 20 years in prison. That's personal criminal exposure, not a fine paid by the company's insurer.

This wasn't a one-off. The SEC's AI-washing campaign began in March 2024 with civil actions against investment advisers Delphia and Global Predictions, followed by charges against Rimar Capital and its owner Itai Liptz for false statements about the firm's AI use. The DOJ joining as a parallel criminal enforcer in the Saniger case, under an administration that changed hands in between, signals this is a durable bipartisan priority, not a political moment.

The common thread across these cases is misrepresentation, not the AI itself. Nobody is being prosecuted for using AI. They're being prosecuted for lying about what it does. That distinction matters for the shadow AI conversation, because a company that doesn't know what its employees have automated can't accurately represent its own AI capabilities to investors, regulators, or customers, even unintentionally.

How is the regulatory net tightening?

On top of the fraud angle, a separate compliance layer is forming around AI use itself, and it's moving fast enough to be unstable.

Colorado passed SB 24-205, the first comprehensive U.S. state law regulating high-risk AI systems, which imposed a duty of reasonable care on developers and deployers to prevent algorithmic discrimination, enforceable by the state attorney general. Then, in May 2026, Colorado repealed and replaced that framework with a narrower transparency-focused law, SB 26-189, dropping the risk management, impact assessment, and duty-of-care requirements entirely. The lesson isn't that regulation is retreating. It's that the rules are genuinely unsettled, and a company that built compliance around one version of the law can find itself out of date within two years.

The EU AI Act is the more stable reference point, and it's severe. Violations of prohibited AI practices carry fines of up to €35 million or 7% of global annual turnover, whichever is higher, with lower tiers of up to €15 million/3% and €7.5 million/1% for other violation categories. Any multinational with EU exposure now has to treat undocumented, employee-built AI workflows as a line item on its regulatory risk register, not just an IT nuisance.

Why does data input matter as much as output in the scraping wars?

The other front is about where AI systems get their data, and it's reshaping what counts as lawful access.

Reddit sued Anthropic in June 2025 in California state court, not over copyright but over breach of contract and unjust enrichment, arguing Anthropic scraped user content in violation of Reddit's terms. In October 2025, Reddit sued Perplexity and three data-scraping firms, Oxylabs, AWMProxy, and SerpApi, in New York federal court, this time under the DMCA's anti-circumvention provision. In July 2026, a Manhattan federal judge denied most of Perplexity's and SerpApi's motions to dismiss, letting the DMCA claims proceed.

Historically, scraping disputes ran through the Computer Fraud and Abuse Act, contract claims, and trespass-to-chattels theories that generally favored anyone scraping publicly available data. AI's scale and commercial value are pushing courts toward new theories, anti-circumvention and breach of contract among them, that could narrow what counts as lawful data access going forward.

This matters for shadow AI specifically because employees building their own scraping or data-aggregation tools are operating in exactly this contested space, usually with no legal review at all. A workflow an engineer considers a harmless internal script may sit on the same legal footing courts are actively re-litigating in cases with tens of millions of dollars at stake. Our piece on the tools employees are already running to build this kind of software walks through what these shadow workflows look like in practice: The Shadow AI Tech Stack: 5 Tools Employees Are Using to Build Their Own Apps.

Where does the liability actually land inside a company?

Put the three threads together and the exposure map looks like this:

• AI-washing exposure lands on the executives who make public or investor-facing claims about AI capability. If a company describes itself as "AI-powered" without knowing how much of that capability actually runs on unmanaged employee workflows versus manual work, that claim is now the kind of thing the SEC and DOJ have shown they will prosecute.
• Regulatory exposure under laws like the EU AI Act or state statutes lands on the company as an entity, through fines tied to global turnover, but board members and compliance officers increasingly carry personal accountability for oversight failures.
• Data and scraping exposure lands wherever the ingestion happened. A single employee pasting confidential documents into an unapproved tool, or building a scraper that touches a competitor's or partner's protected data, can create contract or DMCA liability that traces back to the company that employed them.

Shadow AI sits underneath all three categories at once, and it does so precisely because nobody owns it. IT doesn't see it because it was never provisioned. Legal doesn't review it because it was never a purchase. Security doesn't monitor it because it doesn't show up as sanctioned infrastructure. It's real software, doing real work, carrying real legal weight, and it belongs to no one on the org chart.

How do you actually govern shadow AI?

Banning shadow AI outright doesn't work. Employees route around bans, and roughly 80% of employees already use unsanctioned shadow IT applications regardless of policy, with usage growing rapidly year over year. Ignoring it works even less well, given the incident rates above. The realistic middle path treats employee-built AI workflows the way you'd treat any other piece of infrastructure the company depends on:

1. Discovery first. You can't govern what you can't see. Inventory the AI tools, scripts, and agents already running against company data before writing any policy.
2. Written policy with teeth. Only 37% of organizations currently have one, despite most suspecting unauthorized use is already happening. A policy nobody has written is a policy nobody is following.
3. Approved, capable tooling. Employees build shadow tools because sanctioned options are slower or weaker. Give them something legitimate that's actually competitive.
4. Equal controls for agents and humans. Right now only about a third of organizations hold AI agents to the same access and monitoring standard as human employees. That gap is where most of the exposure above actually originates.
5. Incident reporting without punishment. If employees fear discipline for admitting they used an unapproved tool, incidents stay hidden until they surface as lawsuits or leaks.

The code itself deserves the same scrutiny. AI-assisted tools produce working software fast, but not safe software by default. The governance approach for that layer is covered in Securing the Code Your AI Wrote: 5 Rules for the Copilot Era.

What is the bottom line for builders and buyers of internal AI tools?

The throughline across the Saniger prosecution, the Colorado about-face, the EU's fine schedule, and the Reddit litigation is the same: regulators and courts are done treating AI claims and AI data practices as marketing language. They're treating them as facts subject to verification, and increasingly, as facts subject to criminal liability when they turn out to be false.

Shadow AI sits at the exact intersection of both problems, misrepresentation risk and data-provenance risk, and it does so invisibly, because it was never built with a paper trail. The fix isn't prohibition and it isn't blind permission. It's ownership: knowing what your employees have built, what data it touches, and who's accountable when a regulator or a plaintiff's lawyer comes asking.

The companies that get ahead of this aren't the ones with the strictest ban. They're the ones who can answer, specifically and honestly, what their AI actually does, because that answer is now the difference between a compliance memo and a federal indictment.