Biometric Training Data as M&A Liability:
How to Price the Risk in Avatar Company Acquisitions

Not reliably. The general asset purchase rule — that buyers do not assume seller liabilities — has been significantly eroded for biometric data claims. Courts applying BIPA have found that an acquirer who continues to deploy a trained model built on unlawfully collected biometric data is actively participating in the ongoing violation, not passively inheriting a past one. Under GDPR, the acquirer becomes the data controller for any biometric data that transfers with the asset. Additionally, plaintiffs’ counsel routinely name successor entities in class actions, and courts have shown willingness to allow those claims to proceed even in asset deals. An asset purchase structure reduces exposure at the margins but does not provide a complete liability shield.

No. Deletion of the raw training dataset addresses ongoing storage violations under BIPA Section 15(a) and GDPR Art. 5(1)(e) storage limitation. It does not eliminate claims for the historical period of unlawful collection or transmission. More importantly, the trained model encodes statistical representations of the biometric attributes of the training contributors. Research in model inversion demonstrates that these representations can be partially reconstructed. The model itself is, in a meaningful technical sense, derived from the biometric data of real individuals. Deploying the model post-close means deploying a product built on the contributors’ encoded biometric attributes — a fact that persists regardless of whether the underlying dataset is deleted.

Treat academic biometric datasets as equivalent to scraped internet data for M&A risk purposes. Datasets like LFW (Labeled Faces in the Wild), VGGFace2, and MS-Celeb-1M were assembled from public internet sources without individual consent, predate modern biometric privacy law, and have been the subject of removal requests and litigation by data subjects. MS-Celeb-1M was formally withdrawn by Microsoft following privacy concerns. Using any of these datasets in a commercial model’s training pipeline creates BIPA exposure for any Illinois-resident individuals whose images were included, and GDPR exposure for any EU-resident individuals. The commercial value of “pre-trained on a large public dataset” must be discounted against the inherited biometric liability from that dataset.

It depends entirely on the content of the specific contract and the jurisdiction of the gig worker. A marketplace contract that: (a) separately identifies biometric collection (distinct from general intellectual property assignment); (b) specifies the purpose of biometric use; (c) states a retention period; and (d) was signed before the first biometric capture event — may constitute a valid BIPA written release. Most marketplace contracts written before 2022 do not satisfy these requirements. The burden of proof in BIPA litigation is on the defendant to demonstrate that a compliant written consent was obtained. For M&A purposes, assume that gig marketplace contracts are deficient until a contract-by-contract review confirms otherwise.

Using a conservative Illinois demographic estimate (approximately 4% of US population is Illinois-resident), 50,000 total gig contributors yields approximately 2,000 potential Illinois-resident contributors. Under the SB 2909 aggregate model (one recovery per plaintiff): 2,000 claimants × $1,000–$5,000 = $2M–$10M exposure. Under the pre-amendment Cothron per-scan model (applicable to cases filed before the retroactivity question is resolved): if each contributor’s face geometry was scanned an average of 5 times during onboarding and data collection: 2,000 × 5 scans × $1,000–$5,000 = $10M–$50M exposure. For a $30M–$50M avatar company acquisition, this exposure range is material and should drive escrow sizing, price adjustment, and indemnification structure.