In December 2025, Illinois resident Katelin Cruz filed a proposed class action against Fireflies.AI Corp. after attending a virtual meeting hosted by a local nonprofit that had the company's AI meeting assistant enabled. Cruz never created a Fireflies account, never agreed to its terms of service, and never gave written consent for anything — but her voice was recorded, analyzed, and, according to the complaint, converted into a "voiceprint" through Fireflies' Speaker Recognition feature. A second, similar suit followed in March 2026. Both are built on Illinois's Biometric Information Privacy Act (BIPA) — a statute concerned specifically with biometric identifiers like voiceprints, not with recording or wiretapping as such. That distinction is worth understanding on its own terms, because it shapes every failure point below.
Failure point 1 — Transcription and biometric identification got bundled into one feature
Converting speech to text doesn't require knowing who is speaking — a transcript can just as easily read "Speaker 1" and "Speaker 2" as it can read a name. Fireflies' Speaker Recognition feature goes a step further: to label speakers by identity and track them consistently, it has to construct a voiceprint — a mathematical representation of someone's unique vocal characteristics — and retain it. That's a fundamentally different, higher-risk data type than a transcript. A transcript can be corrected or retracted. A biometric identifier, once created from someone's voice, can't be reissued the way a password can if it's compromised. The complaint's core allegation is that this happened to Cruz automatically, as a side effect of a feature she never asked for, with no distinction drawn between "produce a transcript" and "build a persistent biometric profile."
Failure point 2 — No published retention or destruction policy
BIPA doesn't just require consent before collecting a biometric identifier — it requires a publicly available policy stating how long the data will be kept and when it will be destroyed. The complaint alleges Fireflies has neither. That's a lifecycle-management gap: once a voiceprint is created, there's apparently no defined point at which it's deleted, and no visibility into how long it persists. For a data type that can't be changed once exposed, an undefined retention period isn't a minor documentation gap — it's an open-ended, indefinite exposure window.
Failure point 3 — Consent scope didn't reach the people actually affected
The entity that enabled Fireflies was the nonprofit hosting the meeting, not Cruz. She wasn't a Fireflies customer, had no relationship with the company, and had no opportunity to decline. The consent that mattered legally — a written release for biometric collection — was never obtained from the actual person whose biometric data was taken, only, at best, implied through a host organization's decision to turn a feature on.
Failure point 4 — Why a voiceprint is statistically different from a transcript, and why that's the legal crux
This is the part worth understanding properly, because it's the reason BIPA treats this category of data more strictly than ordinary recordings.
A voiceprint is typically represented as a high-dimensional embedding — a vector of numbers capturing pitch, cadence, resonance, and other vocal characteristics — engineered specifically so that each person's vector is statistically distinguishable from everyone else's. That property, uniqueness across a population, is exactly what makes speaker recognition useful, and it's exactly what makes it a biometric identifier in the same functional category as a fingerprint template, not a text transcript. Words can be common to millions of people. A voice embedding, by design, is built not to be.
That same design property creates a second, less obvious statistical risk: speaker-recognition systems, like facial recognition systems, work by comparing a new voice sample against a stored reference set and returning a match above a similarity threshold — and that comparison has the same base-rate vulnerability as any other biometric matcher. Using illustrative figures to demonstrate the mechanism, not Fireflies' actual published performance data:
- Sensitivity (correctly matching the right speaker): 99%
- False match rate (wrongly matching a different speaker): 0.01% — a very strong number for a biometric system
- Reference database of 50,000 stored voiceprints, with one true match for a given speaker
P(is correct match | flagged) = (0.99 × 1/50,000) / [(0.99 × 1/50,000) + (0.0001 × 49,999/50,000)]
≈ 0.0000198 / 0.00011978
≈ 16.5%Even with a false-match rate this tight, searching against a reference set of 50,000 stored voiceprints still means a "match" is only correct about one time in six. This is a general property of similarity-based matching against a large reference population, not a flaw specific to any one vendor's algorithm — and it's a strong argument for why any system storing biometric identifiers at scale needs a defined, minimal retention period: the larger and longer-lived the reference database gets, the worse this problem becomes, independent of how good the underlying matching algorithm is.
Why this case doesn't need to prove harm to be expensive
BIPA sets statutory damages at $1,000 per negligent violation and $5,000 per intentional or reckless violation — and, notably, plaintiffs don't have to prove actual injury resulted, only that the statute was violated. That's a lower bar than most privacy claims, which is part of why BIPA is one of the most litigated privacy statutes in the country. Applied at scale across every non-consenting meeting participant whose voice was ever converted into a stored voiceprint, the exposure math is a straightforward per-person, per-violation multiplier against a large population — with an even easier path to liability than most privacy claims, since there's no need to show that anyone was actually harmed by the collection itself.
Wayward Fowl's audit perspective
If Wayward Fowl were to audit this AI system, the audit would likely examine whether biometric-identifier creation is technically separated from ordinary transcription functionality, whether a defined retention and destruction schedule governs any stored voice embeddings, and whether consent is captured from every individual whose biometric data is processed rather than inferred from a host organization's decision to enable a feature. Potential findings could include a Speaker Recognition feature that generates persistent biometric identifiers as an unavoidable byproduct of a transcription service, the absence of any published data-lifecycle policy for a category of data that cannot be reissued once exposed, and a consent model that captures authorization from the meeting host but never reaches the non-account-holding participants whose voiceprints are actually being created. We would recommend strengthening technical controls to decouple transcription from biometric identification unless a user explicitly opts into the latter, improving data-governance practices by publishing and enforcing a defined retention and destruction schedule for all biometric identifiers, and implementing a consent mechanism that reaches every individual whose voice is processed, not only the party who enabled the tool, to reduce the likelihood of similar failures.