Named entity recognition (NER) is a natural language processing technique that scans unstructured text, locates the spans that refer to real-world entities, such as names, organizations, locations, dates, and account numbers, and labels each span with its entity type. NER models identify these entities from surrounding context rather than a fixed keyword list, so they can flag a name or identifier they have never seen before. In data privacy, NER is the detection step that finds sensitive information like PII and PHI so it can be redacted or replaced.
How named entity recognition works
An NER model reads text token by token and predicts, for each word or span, whether it belongs to an entity and which category it falls into. Early systems relied on hand-written rules and dictionaries: match a known company name, tag whatever follows "Mr." Modern NER uses machine learning models, typically transformer-based, that learn entity patterns from labeled examples and generalize to entities they were never explicitly shown. That context-awareness is what lets a model catch a misspelled surname or an address written in an unfamiliar format.
Entity categories depend on the use case. General-purpose NER recognizes people, organizations, locations, dates, and quantities. Privacy-focused NER extends this to sensitive identifiers: Social Security numbers, medical record numbers, phone numbers, email addresses, and other PII and PHI. The output is a set of labeled spans — where each entity sits in the text and what type it is — that downstream systems act on.
Why NER matters for data privacy
Unstructured text, like emails, chat logs, clinical notes, support tickets, and PDFs, is where sensitive information hides in the least predictable places. You can't redact what you can't find, and a keyword filter misses any name, date, or identifier it wasn't told to look for. NER solves the detection problem: it locates sensitive entities wherever they appear, including values and formats no rule anticipated.
That makes NER the first stage of de-identifying free text, and its quality sets the ceiling for the whole pipeline: an entity NER misses is an entity that survives into the supposedly redacted output. This is why detection is measured on its own, usually as recall: the share of true sensitive entities the model actually caught.
How Tonic.ai addresses this
Tonic Textual uses proprietary NER models to detect sensitive information in unstructured data, then redacts or synthesizes each entity to produce compliant datasets for AI development and testing. Textual is built for this detection problem specifically. Its models target PII and PHI across documents, transcripts, and free text, powering the PII detection that feeds Textual's redaction and synthesis pipeline and the broader PII redaction workflow. In Tonic.ai's PrivacyBench benchmark, Textual held 95.0% NER recall, well ahead of the 88.8% an Opus 4.8 model reached detecting on its own, a gap that matters because missed entities are the ones that leak.
