Your AI notetaker just transcribed a meeting where you discussed next quarter's roadmap. You skim the summary, forward it to your team, and move on with your day. There's just one problem: two of the sentences in that transcript were never spoken by anyone. The AI made them up.
This isn't a hypothetical scenario. It's a documented, peer-reviewed phenomenon called AI transcription hallucination—and it's creating legal liability, HR nightmares, and discrimination risks that most organizations haven't begun to address.
The Science: AI Transcription Tools Fabricate Entire Sentences
A landmark study presented at the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT) revealed the scope of the problem. Researchers from Cornell University, the University of Washington, and other institutions examined OpenAI's Whisper, one of the most widely used AI transcription models in the world. Their findings were alarming: roughly 1% of audio transcriptions contained entirely hallucinated phrases or sentences that did not exist in any form in the underlying audio.
One percent may sound small—until you consider scale. As Science magazine reported, Whisper is used by approximately 30,000 clinicians and 40 health systems. It's integrated into Oracle and Microsoft's cloud platforms. One recent version was downloaded over 4.2 million times from HuggingFace in a single month. At that scale, a 1% hallucination rate translates to tens of thousands of fabricated sentences flowing into business records, medical notes, and legal documents every day.
The study's most disturbing finding was the nature of these fabrications. Nearly 40% of hallucinations were harmful or concerning. They included references to violence, fabricated medical conditions, invented personal relationships, and even phishing-style website links. In one case, Whisper correctly transcribed a simple sentence about fire department rescues—then hallucinated five additional sentences containing the words "terror," "knife," and "killed," none of which appeared in the audio.
Why Hallucinations Happen—and Why Cloud AI Makes It Worse
AI transcription hallucinations stem from the generative language models underlying modern speech-to-text systems. These models don't just transcribe—they predict what comes next based on patterns in their training data. When audio contains pauses, silence, background noise, or unclear speech, the model fills the gap with plausible-sounding text that may bear no relation to reality.
The Cornell researchers found that hallucinations disproportionately occur during periods of silence or non-vocal audio—the "umms," pauses, and hesitations that are a natural part of human conversation. As lead researcher Allison Koenecke of Cornell explained, silences don't get interpreted as silence; the language model picks them up and creates entire fictional sentences.
Cloud-based transcription compounds the problem in three critical ways:
- No ground truth verification: When audio is processed in the cloud, the original recording often exists only on the vendor's servers. Some services even delete the original audio for "data safety," making it impossible to verify the transcript against what was actually said.
- Automatic dissemination: Cloud AI notetakers frequently auto-email transcripts and summaries to meeting participants, managers, and integrated platforms before anyone reviews them for accuracy. A fabricated sentence can reach dozens of people in seconds.
- Persistent storage: Hallucinated content stored on cloud servers becomes a permanent business record—one that's discoverable in litigation and nearly impossible to fully retract once distributed.
The HR Nightmare: When Fabricated Transcripts Inform Employment Decisions
As Fortune reported in February 2026, AI notetakers are already creating serious HR problems. The recording and dissemination of sensitive conversations creates real legal risks—particularly when hallucinated or inaccurate transcripts feed into performance reviews, disciplinary actions, or termination decisions.
The Littler Mendelson law firm, one of the largest employment law practices in the world, published a detailed seven-point analysis in February 2026 warning employers about AI transcription risks. Among their most important findings: if employers rely on AI-generated transcripts or summaries to evaluate performance, assess candidate interviews, or inform disciplinary decisions, any systemic inaccuracies—such as consistent misunderstandings of individuals with accents, speech impediments, or other protected characteristics—could disproportionately disadvantage certain groups.
This creates a direct path to employment discrimination claims. AI transcription tools may consistently misunderstand accents, speech impediments, or other characteristics tied to protected classes—and if those flawed transcripts inform hiring, promotion, or termination decisions, employers face disparate impact liability under Title VII and analogous state laws.
The scenarios are disturbingly concrete:
- An AI notetaker misattributes a controversial statement to the wrong speaker during a performance review discussion, and the misattribution flows into that employee's personnel file.
- A non-native English speaker's comments are garbled or fabricated by the AI, making them appear less articulate or competent in meeting records used for promotion decisions.
- An AI tool records an HR grievance meeting without proper consent, then distributes the hallucination-laced summary to unauthorized recipients. One Slate reader reported exactly this scenario: an AI transcription tool at their workplace took notes at a meeting evaluating job applicants and then automatically emailed those notes to the entire company—and to the candidates under discussion.
Legal Liability: Hallucinations as Evidence
The legal implications extend far beyond HR. The Goodwin law firm warned in April 2026 that AI transcription technologies can misidentify speakers, mischaracterize speaker intent, misinterpret jargon, and produce transcripts that differ from what was actually said. These inaccuracies can lead to privacy or legal disputes when transcripts are relied upon in litigation, investigations, or regulatory compliance. Furthermore, AI tools that generate summaries or action item lists may inadvertently introduce statements that were never spoken.
If your organization is subject to a litigation hold, AI-generated meeting transcripts—hallucinations and all—are likely discoverable. As we explored in our article on AI transcripts as discoverable evidence, these records can become powerful weapons in the hands of opposing counsel. And correcting a hallucination after the fact creates its own risk: editing an AI transcript stored on a vendor's server could constitute evidence spoliation.
The broader AI hallucination crisis underscores how serious the problem has become. Damien Charlotin, a research fellow at HEC Paris, maintains an AI Hallucination Cases Database tracking legal decisions where generative AI produced hallucinated content. As of May 2026, that database has cataloged over 1,420 cases globally. While most involve fabricated legal citations rather than transcription errors, the underlying technological vulnerability is identical: generative AI models produce confident-sounding output that is entirely fabricated.
The Discrimination Multiplier: Who Gets Harmed Most
AI transcription hallucinations don't affect everyone equally. The Cornell study found that hallucinations disproportionately occur for individuals who speak with longer shares of non-vocal durations—a common characteristic of people with aphasia, elderly speakers, non-native English speakers, and people with speech disabilities.
This creates a discriminatory feedback loop. The people most likely to have their words fabricated or misrepresented by AI transcription are members of protected classes under employment law. When those fabricated transcripts feed into automated workflows—performance tracking, promotion decisions, disciplinary records—the disparate impact compounds.
Jurisdictions are taking notice. As the Littler analysis highlighted, integrating AI transcription tools into hiring or personnel decision-making may trigger AI-specific notice and audit requirements in jurisdictions including New York City, Illinois, and California. Under GDPR Article 22, individuals have the right not to be subject to decisions based solely on automated processing—a provision that could apply directly to employment decisions informed by hallucinated AI transcripts.
Why On-Device Transcription Changes the Equation
The hallucination problem isn't just about AI accuracy—it's about what happens when inaccurate output is automatically stored, shared, and treated as authoritative business records by cloud platforms you don't control.
On-device transcription fundamentally changes this risk profile:
- You control the output. When transcription runs on your device, you review the transcript before it goes anywhere. There's no automated cloud pipeline sending hallucinated content to your entire team before you've read it.
- The original audio stays with you. You can always verify the transcript against the source recording. No vendor deletes your audio for "data safety reasons" while leaving a potentially fabricated transcript as the only record.
- No discoverable copy on a vendor's server. Hallucinated content that never leaves your device doesn't become a business record stored indefinitely in someone else's cloud infrastructure, waiting to surface during litigation discovery.
- No AI model training on your errors. Cloud providers like Otter.ai may use your recordings to train their AI models. On-device processing means your data never feeds into a third-party training pipeline—and neither do the hallucinations.
Apple's approach to on-device AI demonstrates that local processing doesn't mean sacrificing capability. As Apple has stated, the cornerstone of Apple Intelligence is on-device processing—awareness of personal information without collecting personal information. Apple's Speech Recognition framework processes audio locally on the device's Neural Engine, keeping your conversations entirely on your hardware.
Basil AI leverages this architecture to deliver real-time transcription that never leaves your iPhone or Mac. Every word is processed on Apple's Neural Engine—no cloud upload, no third-party server, no opportunity for a hallucinated transcript to be auto-emailed to your entire department before you've had a chance to review it.
What Organizations Should Do Now
The convergence of hallucination risk, privacy liability, and discrimination exposure demands immediate action. For organizations already grappling with the wiretapping and consent challenges of AI meeting bots, transcription accuracy is the next compliance frontier.
- Audit your AI transcription tools. Determine what happens to meeting audio after transcription. Is the original recording preserved? Can you verify transcripts against source audio? Who receives auto-generated summaries?
- Establish mandatory human review. No AI-generated transcript should feed into employment decisions, legal records, or client communications without human verification. The Littler analysis recommends that employees always review AI-generated records and correct them before using them as a basis for business decisions.
- Evaluate on-device alternatives. For sensitive meetings—HR proceedings, legal discussions, client conversations, medical interactions—on-device transcription eliminates the risk of hallucinated content being automatically stored and shared through cloud infrastructure.
- Develop clear AI notetaker policies. Define which meetings can use AI transcription, who must review the output, how long recordings are retained, and what happens when hallucinations are discovered.
- Assess discrimination risk. If AI transcription feeds into any employment decision-making process, conduct a bias audit to determine whether the tool systematically misrepresents speakers with accents, speech disabilities, or other characteristics tied to protected classes.
The Bottom Line
AI transcription is a powerful productivity tool—but only when you control the output. Cloud-based AI notetakers that automatically store, share, and integrate hallucinated content into your business workflows create a ticking time bomb of legal, HR, and compliance risk.
The question isn't whether your AI notetaker has ever hallucinated. Based on the research, it almost certainly has. The question is whether that hallucination was caught before it became a business record, an employment decision, or a piece of discoverable evidence.
On-device transcription doesn't eliminate AI inaccuracies. But it ensures that you—not a cloud platform—control what happens with the output. And in the era of AI hallucination liability, that control is everything.
Keep Your Meeting Notes Accurate and Private
Basil AI processes transcription 100% on-device using Apple's Neural Engine. No cloud upload. No auto-shared hallucinations. No third-party access to your recordings. You review every word before it goes anywhere.
