[ 10 / comparison — limina ]
Synthaveil and Limina (formerly Private AI).
What it is
A commercial service that detects and removes personal data from text, documents, images and audio. It runs as a hosted API or as a container inside your environment, and it can put the original values back into a model's response.
Where it wins
Limina finds personal data in 52 languages, and it reads images, including scanned pages, and recorded calls, which we do not do today. It also publishes accuracy figures. If you need detection beyond English, or in images or audio, Limina is ahead of us today.
Where we differ
Limina’s default swaps a detail for a label such as [NAME_1]. It can also make realistic replacements, which its documentation says are created with an ML model and are non-deterministic: the same person can come back with a different fake name each time, by chance. With Synthaveil you decide. A stand-in can stay the same for as long as you need, so one customer’s chats stay one customer and one patient’s letters, notes and results still line up. Or it can change each day or for each project, so separate pieces of work cannot be joined up. It is a setting you choose, not chance. The stand-ins also keep the shape your analysis relies on: an age stays close to the real one, and a postcode stays a real postcode that does not point back to the person. No generative AI or large language model ever sees your real data.
| Capability | Synthaveilthis site | LiminaCommercial detection and redaction |
|---|---|---|
| Runs inside your own environment | yes | partly |
| No generative AI or large language model reads your real data | yes | partly |
| Output keeps its shape — a name looks like a name, a date like a date | yes | yes |
| Consistent results without a stored table of real-to-fake values | yes | partly |
| Nothing held at the vendor that could reverse it | yes | partly |
| Replacement matched on statistical character, not only on format | yes | not verified |
| Location stand-ins that keep the area useful without pointing back to it | yes | not verified |
| Stand-ins go stale on a schedule | yes | not verified |
| Reversible on your side when you need the real value back | yes | yes |
| An audit record for every policy decision | yes | partly |
| Works on live traffic, not only batch jobs | yes | yes |
| Word, Excel and browser add-ins | yes | not verified |
| Stand-ins built from official public statistics | yes | not applicable |
| Detection across many languages and file types | partly | yes |
| Free to use | no | not verified |
Source: Limina documentation
Can the stand-in values be turned back into the real values?
Yes, on your side. When an AI tool replies using stand-ins, Synthaveil puts the real values back, so the answer reads with your real names and is ready to use. Alias, Cover and Legend all restore values; each product page sets out how.
Is pseudonymised data still personal data under GDPR?
For you, yes: you hold the key, so it remains personal data in your hands. The EU Court of Justice (EDPS v SRB, 2025) and the ICO (2025) both look at whose hands the data is in: a recipient with no key, no mapping and no other route to the person may not hold personal data at all. That is what Synthaveil sends an AI provider: stand-ins with no route back. Your duty to tell people about the sharing still applies when you collect their data.
Does Synthaveil make my data anonymous?
No. Synthaveil makes no claim that its output is anonymous. In your hands it remains personal data, because you hold the key. What changes is what the recipient gets: stand-ins with no key or mapping to link them back. Review the remaining context before you share.
Do you have to store a mapping table to keep pseudonyms consistent?
No. Synthaveil keeps stand-ins consistent without storing any real-to-stand-in pairs. Products that restore values keep that information on your side, in your browser, workbook or n8n environment, and it needs protecting.