Can AI re-identify anonymized emails and Slack?
The short answer
Maybe. Three 2026 preprints show LLMs matching accounts at 68% recall and re-identifying 79.2% of identities, but none tested company email or Slack. The FTC says data is anonymous only when it can never be associated back to a person. Ask de-identification buyers their standard.
Most buyers I read describe a de-identification step before data reaches a buyer. These papers test how well that kind of step holds up. They are preprints on arXiv, so peer review hasn't confirmed them.
What did the Lermen account-matching study find?
Lermen and co-authors, in a preprint, matched pseudonymous accounts across platforms using only raw user text. They linked Hacker News accounts to LinkedIn profiles and matched Reddit users across communities and across time.
The abstract reports up to 68% recall at 90% precision, and says the best non-LLM method came in near 0%. The method works on unstructured text without a fixed schema.
What about de-identification tools?
Palacios and co-authors, in a preprint, tested 100 pediatric oncology notes. An LLM prompted with institution-specific knowledge scored F1 0.918 against 0.779 for Stanford's TiDE system. Re-annotation confirmed 227 identifier spans that the original gold standard missed.
The takeaway is a narrow one. Scrubbers miss identifiers that are specific to one organization, and the gold standard used to grade them can miss some too.
Do the studies cover emails and Slack?
No. The first used public social accounts and the second used clinical notes. A third preprint, by Ko and co-authors, reports that agents reconstructed 79.2% of identities in the sparsest Netflix Prize setting, against 56.0% for a classical baseline.
I found no study that ran the test on a company's own email or chat archive. In my reading, free text where names and context sit inside the message is the setting these results point to.
What does the FTC say?
The FTC's technology blog says data is only anonymous when it can never be associated back to a person. The post is about hashing.
What does this mean for an owner?
It means de-identification is a claim to examine, and I'd examine it. Ask any buyer who does the work, to what standard, and whether anyone tests for re-identification afterward.
The risks page lists what published buyer terms say, and the rights check lists what to settle with counsel.
Talk to a privacy or commercial lawyer in your state before you sign.
Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one. Ask for a referral
