Shadow AI in the Workplace: Recording Consent, Privilege Waiver, and Trade Secret Risks

The meeting starts. A small notification appears in the corner of the screen: “Otter.ai has joined the call.” Or maybe it does not appear at all. A sales manager added an AI notetaker to the team’s shared calendar weeks ago, and now it joins every meeting automatically. Most participants have never been asked whether they consent to being recorded. No one told them their words would be transcribed, processed by a third-party vendor, and potentially used to train a language model.

This is shadow AI in the workplace — the unauthorized or unsanctioned use of AI tools by individual employees or teams, outside of any policy framework, without legal review, and often without meaningful notice to the people affected. It is one of the most significant and underappreciated legal risks facing businesses in 2026.

The legal exposure comes from multiple directions simultaneously. State recording consent laws create civil and criminal liability for unauthorized transcription. Attorney-client privilege can be waived when confidential communications pass through third-party AI systems with permissive terms of service. Trade secret protection can be compromised when proprietary information is disclosed to a vendor without adequate contractual safeguards. And AI-generated records from these tools can become discoverable in litigation — creating spoliation risk if they are not preserved and evidence problems if they are.

This post addresses each of these risks in practical terms, and explains what employers must do to manage them.

What Shadow AI Actually Looks Like

“Shadow AI” is a term for AI tools that employees use without formal corporate authorization or policy oversight. It is the workplace AI equivalent of “shadow IT” — the unsanctioned software and cloud services that employees have always adopted faster than legal and IT can review them.

In 2026, shadow AI in the workplace most commonly takes these forms:

AI notetakers and meeting transcription tools — Otter.ai, Fireflies.ai, Fathom, tl;dv, and similar products join video calls, record audio, generate transcripts, and produce meeting summaries. Employees add them to recurring meetings, sometimes without disclosing this to other participants.

Consumer-tier generative AI tools — Employees use free or personal-account versions of ChatGPT, Claude, Gemini, and others to draft documents, summarize information, analyze data, and assist with work tasks. Consumer-tier tools typically have terms of service allowing the provider to use inputs for model training.

AI-powered browser extensions and add-ons — Tools that summarize emails, suggest replies, transcribe voice notes, or analyze web content, often processing communications and documents in the background without explicit user awareness.

AI writing and editing assistants — Grammarly, Notion AI, and similar tools that process document content, potentially including sensitive or privileged information.

What makes these tools “shadow” AI is not that they are necessarily bad tools — many are genuinely useful. What makes them risky is that they operate outside any legal, IT, or compliance review framework, meaning the company has no visibility into what data is being shared, with whom, on what terms.

Recording Consent Laws: The Immediate Liability

The most immediate and concrete legal risk from AI notetakers and transcription tools is violation of state recording consent laws.

Federal law — specifically the Electronic Communications Privacy Act and the federal Wiretap Act — prohibits intentional interception of oral or electronic communications without the consent of at least one party. This is a “one-party consent” standard at the federal level. If the person adding the AI tool to the meeting counts as a consenting party, federal law may not be violated.

But eleven states and the District of Columbia have enacted “all-party consent” or “two-party consent” recording laws, which require that every participant in a recorded conversation must consent before the recording occurs. These states include California, Connecticut, Florida, Illinois, Maryland, Massachusetts, Michigan, Montana, New Hampshire, Oregon, Pennsylvania, and Washington.

California’s law is the most aggressively enforced. California Penal Code section 632 prohibits recording a confidential communication without the consent of all parties. Section 637.2 provides a civil remedy of $5,000 per violation or three times actual damages, whichever is greater. “Per violation” is interpreted by courts to mean per recording, not per lawsuit — so a recurring meeting notetaker that attends fifty meetings without proper consent could theoretically generate $250,000 in civil exposure from a single participant.

The ongoing litigation against Otter.ai illustrates the scope of this risk. Beginning in August 2025, a class action was filed in California federal court alleging that Otter.ai’s AI Notetaker recorded and transcribed meeting participants without the consent of all parties, in violation of the California Invasion of Privacy Act and the federal Wiretap Act. A consolidated complaint was filed in December 2025. The core allegation is that Otter obtained consent from the meeting host who added the assistant, but not from the other participants whose voices it recorded. Fireflies.ai faces separate biometric privacy suits in Illinois.

What makes this litigation directly relevant to employers is the question of who bears liability when an employee’s use of shadow AI violates recording consent laws. If an employee adds an AI notetaker to business meetings without authorization, and participants in those meetings include people in all-party consent states, the employer faces potential vicarious liability for the employee’s actions, as well as direct liability to the extent the practice was widespread and the company should have known about it.

The critical question is whether you have a policy — and whether you have implemented it. A company that has a clear, communicated prohibition on unauthorized AI recording tools, and enforces it, is in a far better position than a company that has no policy and discovers after litigation that its employees were running AI transcription on every customer call.

Privilege Waiver: When Legal Communications Pass Through AI

Attorney-client privilege protects confidential communications between a client and their lawyer for the purpose of obtaining legal advice. Work product doctrine protects documents and materials prepared in anticipation of litigation. Both protections depend on confidentiality being maintained. When a communication or document is disclosed to a third party outside the privileged relationship, the privilege is typically waived.

This creates a specific problem when employees use consumer-tier AI tools to process legally privileged communications.

Consider the scenario: a general counsel uses a free-tier AI assistant to summarize a legal memo about pending litigation. Or a manager uses ChatGPT to help draft a response to a lawyer’s email. Or an AI notetaker transcribes a meeting in which in-house counsel provides legal advice about an employment dispute. In each case, privileged communications are being shared with a third-party AI provider whose terms of service may allow the provider to use that data for model training, share it with employees, or store it indefinitely.

The February 2026 ruling in United States v. Heppner, decided by Judge Jed Rakoff in the Southern District of New York, illustrates the problem with precision. In that case, a defendant used Anthropic’s Claude AI tool to prepare documents outlining his legal defense strategy and potential arguments. Judge Rakoff ruled that these exchanges were neither protected by attorney-client privilege nor work product doctrine. The AI is not an attorney, so communications with it are not inherently privileged. And Anthropic’s terms of service — which allowed user data to be used for model training and disclosed to third parties — defeated the confidentiality requirement for privilege.

The distinction between consumer-tier and enterprise-tier AI tools matters here. ChatGPT Enterprise and similar enterprise-grade products offer zero data retention options, contractual confidentiality commitments, and terms that prohibit using customer data for model training. Consumer and free-tier products typically do not. An employee using a personal ChatGPT account for work tasks is sharing information on consumer-tier terms — terms that courts have found defeat privilege claims.

Employers face a compounding problem: most employees do not distinguish between enterprise and consumer AI tools. A policy that authorizes use of “AI assistants” without specifying which specific tools, at which service tier, with which data handling terms, is not a policy that protects privilege.

Trade Secret Risks: Disclosure Without Adequate Safeguards

Federal trade secret law under the Defend Trade Secrets Act and state trade secret laws under the Uniform Trade Secrets Act require a trade secret holder to take reasonable measures to maintain secrecy as a condition of protection. If a company discloses trade secret information to a third party without adequate confidentiality protections, it may lose trade secret status for that information — permanently.

When employees use shadow AI tools to process proprietary business information — competitive analyses, product roadmaps, customer lists, pricing models, source code, manufacturing processes — they may be transmitting trade secrets to AI vendors without any contractual confidentiality protection in place. The company has no nondisclosure agreement with the AI vendor at the employee’s tier. The vendor’s terms of service typically disclaim confidentiality obligations for free-tier users and reserve the right to use inputs for model improvement.

Courts evaluating whether trade secret protection has been lost will examine whether the company took reasonable steps to maintain secrecy. Using an AI tool whose terms of service explicitly permit the provider to retain and use inputs is difficult to reconcile with “reasonable measures to maintain secrecy.” If the AI vendor’s model has been trained on your trade secrets, and the vendor’s outputs to other users could potentially reflect that training, the argument that you maintained adequate secrecy becomes very difficult to sustain.

The trade secret risk is particularly acute for technology companies and startups, where competitive advantage is tied directly to proprietary information. Employees working remotely, under time pressure, and habituated to using consumer AI tools for convenience will often not think twice about pasting source code, internal strategy documents, or customer data into a chat interface. Without policy, training, and technical controls, this happens constantly — and the legal consequences are irreversible.

Spoliation: When AI-Generated Records Become Evidence

AI notetakers and transcription tools create records — detailed records, often stored on third-party servers outside the company’s control. When litigation arises, those records may be discoverable. When they have been deleted — or when a company failed to place a litigation hold on them before they were automatically purged — the company may face sanctions for spoliation of evidence.

Spoliation occurs when a party to litigation destroys or fails to preserve evidence that it had a duty to retain. Courts have broad authority to sanction spoliation, including instructing juries to draw adverse inferences — that is, to presume that the destroyed evidence would have been damaging to the party that destroyed it.

For employers, the spoliation risk from AI meeting records takes two forms. First, if an AI notetaker recorded meetings relevant to litigation and those records were automatically deleted after the tool’s default retention period, the company may have failed to preserve discoverable evidence. Second, if an employee used shadow AI tools to process information related to a dispute — generating summaries, analyses, or drafts — and those tools purge records automatically, the company needs to have a mechanism to preserve them.

Many AI transcription tools retain meeting records for 30 to 90 days by default before automatic deletion. Employees who used shadow AI tools for months or years before a legal dispute arose will have a fragmented and potentially incomplete record. The employer’s obligation to implement a litigation hold upon reasonably anticipating litigation extends to AI-generated records in third-party systems — but only if the employer knows those records exist.

This is another reason why shadow AI is a legal problem, not just an IT problem. When AI tools are used without authorization or disclosure, the employer has no visibility into the records they create. No visibility means no litigation hold. No litigation hold means potential spoliation sanctions.

What a Shadow AI Policy Must Address

Addressing shadow AI risk requires more than a generic “authorized tools only” policy. The specific risks described above — recording consent violations, privilege waiver, trade secret compromise, and spoliation — require specific policy provisions.

Explicit tool authorization. Identify and publish a list of approved AI tools for workplace use. For each tool, specify the authorized service tier (enterprise only, never consumer/free), the approved use cases, and any data classification restrictions. Employees should understand that using non-listed tools is prohibited, not merely discouraged.

Meeting recording and AI notetaker protocol. Establish a clear policy on AI notetakers and meeting transcription. At minimum, the policy should require disclosure to all participants before recording begins, and prohibit use of AI transcription tools in calls involving legally privileged discussions, customers who have not consented, or participants located in all-party consent states unless all-party consent is confirmed.

Data classification rules tied to AI use. Establish data classification categories (confidential, proprietary, trade secret, legally privileged) and specify which categories cannot be processed through AI tools — or can only be processed through enterprise-tier tools with contractual confidentiality commitments.

Vendor review and approval. Any AI tool that will process company data should go through a vendor review process before deployment. The review should examine the vendor’s data handling terms, security certifications, and contractual obligations. Enterprise agreements should include NDAs or confidentiality terms covering processed data.

Litigation hold procedures that include AI records. When a litigation hold is triggered, the procedure should explicitly require employees to identify any AI tools they used to process information related to the matter, and to notify the relevant AI vendor to suspend automatic deletion of relevant records. Many enterprise AI tools have legal hold features; consumer tools typically do not, which is an additional reason to restrict use to enterprise tiers.

Training and awareness. Policy without training is ineffective. Employees — especially managers and executives who use AI tools most heavily — need to understand the specific legal risks: that recording without consent can violate state law, that sharing privileged communications with AI can waive privilege, that transmitting trade secrets to an AI vendor can compromise trade secret protection.

The Employer’s Bottom Line

Shadow AI is not a niche IT risk. It is a set of overlapping legal exposures that sit at the intersection of employment law, intellectual property law, evidence law, and privacy law. Every medium-size business that has employees who use AI tools — which, in 2026, is every medium-size business — is exposed to some combination of these risks right now.

The Otter.ai class action is a preview of what the litigation landscape looks like when unauthorized AI recording becomes a legal claim at scale. The Heppner ruling is a warning about what happens to privilege when legal communications pass through consumer AI tools. The trade secret cases are beginning to develop. The spoliation issues will become visible when the first wave of AI-related employment and commercial litigation moves into discovery.

The window to get ahead of these risks is open. It requires clear policy, employee training, enterprise-tier tool procurement, and litigation hold procedures that account for AI-generated records. The companies that act now will have defensible positions when disputes arise. The companies that wait will be litigating in the dark.

This post is for general informational purposes only and does not constitute legal advice. Reading this post does not create an attorney-client relationship. If you have questions about your specific situation, consult a qualified attorney.