Debunking the Myth of "Anonymous" Data
Claims that user data can be safely “anonymized” are increasingly challenged by real-world re-identification cases and legal standards like the GDPR, which treat most pseudonymous datasets as still personal. Commenters debate whether techniques such as differential privacy, k-anonymity, synthetic data, and hashing meaningfully protect individuals, especially when datasets can be linked and business incentives favor maximum utility. Many conclude that true anonymity is either practically unattainable or requires so much noise that the data loses most of its value, reinforcing calls to limit data collection itself rather than relying on post hoc anonymization.
Limits of “Anonymous” Data
- Many commenters agree that rich, individual-level datasets are inherently hard to truly anonymize; linkage across datasets often re-identifies people.
- Classic failures (AOL search logs, Netflix Prize) are cited as examples of “anonymized until correlated.”
- Some argue that if any reasonably practical re-identification is possible, the data should be treated as personal.
- Others counter that anonymity is a spectrum: aggregates like “51.1% of US population is female” clearly include individuals yet pose no realistic re-identification risk.
Differential Privacy and Other Technical Approaches
- Several note the article downplays or omits differential privacy (DP), synthetic data, zero-knowledge proofs, and homomorphic encryption.
- DP is seen by some as mathematically solid and often “good enough”; others see it as encouraging more data sharing and giving a false sense of security, heavily dependent on parameter choices.
- Critics say most “anonymization” schemes assume weak attackers and fail against sophisticated adversaries; current techniques mostly raise the cost rather than prevent de-anonymization.
- There’s debate over whether effective, scalable, and truly anonymous analytical models are even theoretically achievable.
Pseudonymization, Indirect Identifiers, and GDPR
- Several distinguish pseudonymization (removing direct identifiers or hashing them) from stronger techniques that also treat indirect identifiers (age, ZIP, gender, etc.).
- EU-style rules are highlighted: any data that can reasonably be linked back to a person is still personal data; “reasonably likely” attacks must be considered, using all data the controller has.
- Disagreement appears over how strictly this is interpreted in practice and how it maps to US concepts like PII.
Industry Practice, Incentives, and Threat Models
- Some argue the root problem is excessive data collection; no technique fixes that.
- Ad-tech and data brokers are viewed as unlikely to use strong privacy methods because raw data is more valuable and penalties are weak.
- Others emphasize internal, first-party use cases (e.g., dev/test databases) where anonymization meaningfully reduces risk without selling data.
- There’s tension between “good faith, partial protections are worthwhile” and “anything short of not collecting data is false security or snake oil.”
Demonstrations, Services, and Pragmatism
- A few mention tools/startups offering risk analysis, de-identification, or anonymization; one seeks de-anonymization services to demonstrate risks.
- Analogies to door locks stress threat modeling: no lock is perfect, but they can be good enough against most attackers.
- Some advocate “poisoning” datasets with believable fakes to frustrate re-identification attempts.