New AI Labels Your Company's Sensitive Data Automatically
Unlabeled data is a lawsuit waiting to happen — this AI hunts it down.
Every big company stores information in giant databases made of tables and columns — 'customer_email', 'order_total', 'patient_diagnosis'. Someone has to write down what each column means and flag which ones hold private information. In practice, nobody keeps up. Data piles up faster than people can label it, so companies end up with mystery columns nobody understands and sensitive fields nobody protected. That gap is how data leaks happen, and how companies fail privacy audits like GDPR or HIPAA.
Glyph, a system from three researchers, tackles two jobs at once: describing what each column holds, and tagging it with the right classification label. It does this with AI agents — software that can take actions on its own — arranged like a flowchart. One agent, the Descriptor, doesn't guess from the column name. It goes and fetches the actual computer code that creates the column, reading it on demand from the company's code repository. That's a version of RAG (letting AI look things up) where the AI decides what to look up.
The second agent, the Tagger, picks labels from a company's official list of 275 data categories. Rather than trusting one method, it runs three in parallel: one reads column descriptions, one uses pattern-matching rules, and one compares each column's technical details against similar labeled columns stored in a searchable database. Their answers are then merged, like ranked votes, into one result. The team fine-tuned a small language model for that third method, which raised its matching accuracy score from 0.55 to 0.92 — roughly the difference between 'often wrong' and 'usually right'.
Why should you care? Two reasons. First, Glyph never looks at the actual data values inside your records — it works from code and metadata — so your private information isn't fed into an AI. Second, it shows its work: every label comes with a reason attached. That matters because regulators and auditors don't accept 'the AI said so'. Making automated cataloging explainable and reliable is what turns it from a demo into something a bank or hospital can actually run.
- Glyph automatically writes descriptions and privacy labels for database columns, replacing slow manual work that companies rarely finish.
- It uses three AI methods in parallel and combines their answers, boosting tag-matching accuracy from 0.55 to 0.92 in testing.
- It never reads the actual data values — only code and metadata — and records why each label was applied, which matters for privacy audits.
Why It Matters
Faster, more accurate data labeling means fewer breaches, cheaper compliance audits, and less sensitive information slipping through unnoticed.