The Paste Is the Disclosure: The Legal Risk of Unapproved AI Tools

An analyst has a model due at four o'clock. The company's approved AI assistant is slow and won't open the spreadsheet, so they paste the client's revenue figures into the chatbot they pay for personally, get a clean summary in eleven seconds, and make the deadline. Nothing breaks. No alert fires. Nobody will ever mention it.

Depending on what was in those cells, the analyst's employer may have just weakened its claim to a trade secret, put a privileged document at risk, made a disclosure that HIPAA presumes to be a breach, or started a 72-hour notification clock under the GDPR. None of those consequences requires a hacker, a software bug or bad intent. They follow from an act the law has understood for a very long time: telling someone who has no duty to keep quiet.

That is the argument of this piece. The legal risk of shadow AI is not new law struggling to catch up with new technology. It is old law, written around the question "who did you tell?", meeting a tool that makes telling effortless. The paste is the disclosure. Most of what comes after is cleanup.

How ordinary this has become

The instinct is to treat this as an edge case, so start with scale. It is closer to the median case.

When Microsoft and LinkedIn surveyed 31,000 people for the 2024 Work Trend Index, 75 percent of knowledge workers said they used AI at work, and 78 percent of those users were bringing their own tools rather than waiting for their employer to choose one. Two years later the habit is showing up in breach data. IBM's 2026 Cost of a Data Breach study put the global average cost of a breach at $4.99 million, a record, and a summary of the report by the law firm Baker Donelson notes that incidents involving shadow AI by employees more than doubled in a year, to 43 percent from 20 percent, while 68 percent of organizations lacked AI governance policies.

Read those numbers together and the picture is less "rogue employees" than a workforce routing around friction. People reach for whatever works. They do not experience pasting a document into a chatbot as disclosing it; they experience it as asking for help. The gap between what the act feels like and what it is in law is where the liability lives.

Four bodies of law, one question

Trade secret law, legal privilege, HIPAA and the GDPR come from different eras and different legislatures, and they disagree about almost everything. On one point they converge. Each turns, at some decisive moment, on whether information went to someone outside the circle of people obliged to protect it.

Trade secrets exist only while you guard them

A trade secret is an unusual kind of property, because it is defined by the owner's behavior. Under the federal Defend Trade Secrets Act, information qualifies only if "the owner thereof has taken reasonable measures to keep such information secret" (18 U.S.C. § 1839(3)). The Supreme Court stated the consequence plainly in Ruckelshaus v. Monsanto: "If an individual discloses his trade secret to others who are under no obligation to protect the confidentiality of the information, or otherwise publicly discloses the secret, his property right is extinguished."

A single paste into a consumer chatbot probably does not destroy a secret by itself. Courts weigh the whole program of protection, and an AI provider is not a newspaper. The sharper risk arrives later, in litigation. Suppose you sue a former employee for walking off with your pricing model, and the other side's first exhibit is a record of your own staff pasting that model into personal AI accounts whose terms let the provider use what users submit to improve its models. The question stops being "did they take it?" and becomes "did you ever really treat it as secret?"

That is one way to read what Samsung did in 2023. After staff uploaded sensitive code to ChatGPT, the company restricted generative AI on its own computers. The move was reported as a security story. It also protected Samsung's ability to say, in any future dispute, that it guarded its code. Our longer piece on source code in AI coding assistants covers the engineering side.

Privilege depends on keeping the circle closed

Attorney-client privilege protects confidential communications, and confidentiality is the whole mechanism. Bring an outsider into the conversation and the protection can leave with them. Courts have started applying that logic to AI. In February 2026, in United States v. Heppner, a federal judge in Manhattan held that a defendant's exchanges with a consumer AI tool were not privileged, in part because the provider's privacy policy defeated any expectation of confidentiality. Other courts have protected AI-assisted litigation materials as work product, so the law is genuinely unsettled; we walk through both lines of cases in our analysis of AI and the work product doctrine.

For a company, the exposure is wider than its lawyers' habits. A manager who pastes outside counsel's advice into a chatbot to make it readable for the team has done something structurally close to forwarding it to a stranger. The privilege belongs to the company, and almost any employee holding privileged material can put it at risk. Lawyers carry professional duties on top of this, which we cover in Can Lawyers Use ChatGPT?

HIPAA presumes the worst

Health privacy law takes the most aggressive position of the four. Under 45 C.F.R. § 164.402, a use or disclosure of protected health information that the Privacy Rule does not permit "is presumed to be a breach unless" the organization demonstrates a low probability that the information has been compromised. That demonstration rests on a risk assessment of at least four factors, including who received the data and "whether the protected health information was actually acquired or viewed."

Look at where the burden sits. A nurse who pastes a discharge summary into a personal chatbot, with no business associate agreement covering it, has made a disclosure the rule does not allow. To avoid treating it as a reportable breach, the hospital has to prove a negative about a system it does not control and has never seen inside. Often it cannot even say exactly what was pasted. We explain why consumer chatbots sit outside HIPAA in Is ChatGPT HIPAA Compliant?

The GDPR counts disclosure as a breach

European law reaches the same place by definition. Article 4(12) of the GDPR defines a personal data breach to include the "unauthorised disclosure of, or access to, personal data" in an organization's care. No attacker is required. An employee sending customer records to an AI provider the company has no data processing agreement with fits those words comfortably.

Once a controller becomes aware of a breach, Article 33 gives it 72 hours, where feasible, to notify the supervisory authority, unless the breach is unlikely to result in a risk to people's rights and freedoms. The notice has to describe the categories and approximate number of people and records involved. That produces a strange incentive: the clock starts at awareness, so not looking can feel safer than looking. It isn't. An organization that keeps itself blind will struggle to show the "appropriate technical and organisational measures" that Article 32 requires. The wider picture for European teams is in our GDPR and AI guide.

Why banning the tools rarely works

The reflex response is prohibition, and it has a short shelf life. A ban on company devices moves the activity to personal phones, where the company sees even less. The Work Trend Index figures describe people who adopted AI without asking in the first place, and a memo does not change their reason for doing it: the work is heavy and the tool makes it lighter.

A ban also points attention at the wrong thing. It is a rule about software, while the exposure is about data. A clinician summarizing a published study in an unapproved tool creates almost no risk. The same clinician pasting a patient note into an approved tool that was set up carelessly can create plenty. A sensible AI acceptable use policy is organized around categories of information for exactly this reason. Be honest with staff about visibility as well. Employees tend to assume their AI use is private, and on company systems it often is not, as we lay out in Can My Employer See My ChatGPT Prompts?

The question underneath all four

Put the doctrines side by side and they ask the same practical question at the moment of crisis: what exactly left, and who has it now?

HIPAA's risk assessment wants the nature and extent of the data and whether anyone viewed it. A GDPR notice wants categories and approximate numbers. A trade secret defendant will ask what measures you had in place and what your records show. Most organizations cannot answer any of these about AI use, because nothing in their systems records what was pasted into a browser tab. They find out the way Samsung did: after the fact, from the people involved.

Underneath the doctrine is a quieter change, and it explains why the problem keeps growing. For most of history, confidentiality was protected less by rules than by inconvenience. Leaking a document meant copying it, carrying it out, mailing it. Each step was a chance to reconsider, and each left a trace. Generative AI removed the steps. The distance between "I have this information" and "a third party has this information" is now a single keystroke, taken in the middle of trying to do good work. Policies written for a world of friction assume a pause that no longer exists.

So the control has to move to where the pause used to be: the moment before text leaves the device. Training still matters, and so do contracts with the providers you approve. But the only safeguard that works at the speed of a paste is one that sits in the path of the paste.

Frequently asked questions

Is it illegal for an employee to paste company data into an unapproved AI tool?

Usually not in itself. The legal consequences travel through other channels: the employer's confidentiality duties to clients, statutes like HIPAA and the GDPR that treat unauthorized disclosure as a breach, and trade secret law's demand for reasonable secrecy measures. The employee may also be breaching an employment agreement or company policy, which is a separate question from whether any statute was violated.

Does deleting an AI chat undo a disclosure?

No. The disclosure happened when the data was sent. Deleting a conversation may shorten how long a provider keeps it, but retention depends on the provider's terms, the account type and any legal holds, and deleted material can persist in backups or logs for a period. Our guide to whether ChatGPT saves your prompts covers the details for one major provider.

Does an approved enterprise AI tool remove these legal risks?

It removes many of them, because an enterprise agreement usually makes the provider a contracted processor bound to confidentiality, with training on your data excluded. It does not remove all of them. HIPAA still requires a business associate agreement, the GDPR still requires a lawful basis and a data processing agreement, and an approved tool can still be handed data that has no business being there.

Closing the distance

The organizations that handle this well will not be the ones with the longest AI policy. They will be the ones that accept people are going to paste, and plan for it: approve good tools quickly so there is less reason to go around them, write policy around data rather than brand names, and put a check at the moment of sending instead of relying on someone's judgment at four o'clock on a deadline.

That last check is why we built Locke. It runs on your device and catches names, account numbers, health details and secrets before a prompt leaves for the AI tools it screens, replacing them with placeholders so the model still gets a usable request while the sensitive values stay with you, and the unmasked text is never sent to Sonomos either. It will not write your policy or sign your vendor contracts, but it puts back the pause the paste took away. Download Locke and try it on the work you are already doing.