AI assistants that say they do not train on your data by default
Which AI assistants do not use my conversations to train their models?
Updated · 18 assistants ranked
Of the 47 assistants in our catalog, 21 record that they do not train on your data by default, 6 record that they do (Alfi, Instinct, Muse, Pally, Poke and Wajo / Fo) and 20 are unclear. This report lists the 21 and ranks the 18 that have a score on Permissions & privacy or Memory; Orchid, Folk and Airtap have neither. A policy statement is not a test result: the scores show how each assistant behaves on access, rules and deletion, and most are thin, with one test each for 6 of the 18.
Top picks
- 110.0HarkTop pick
Permissions —Memory 10
- 28.0NeoBest free option
Permissions 8Memory 8
- 38.0OllieFastest that's still good
Permissions 8Memory —Time to finish 7m 25s
- 47.0
Permissions 7Memory 7
- 57.0
Permissions 8Memory 6
Top pick: highest average here. Best free option: highest among assistants you can use for free. Fastest that's still good: quickest median time to finish among those averaging 7 or more.
Full ranking
| # | Assistant | Permissions | Memory | Average |
|---|---|---|---|---|
| 1 | Hark | — | 10 | 10.0 |
| 2 | Neo | 8 | 8 | 8.0 |
| 3 | Ollie | 8 | — | 8.0 |
| 4 | Bantr | 7 | 7 | 7.0 |
| 5 | Familiar | 8 | 6 | 7.0 |
| 6 | Vellum | 7 | 7 | 7.0 |
| 7 | Ask Ranger | 7 | 6 | 6.5 |
| 8 | Norton Family Assistant | 5 | 7 | 6.0 |
| 9 | Friday | 6 | 6 | 6.0 |
| 10 | Del | 8 | 4 | 6.0 |
| 11 | Caddy | — | 6 | 6.0 |
| 12 | Luke | — | 6 | 6.0 |
| 13 | Persona | 6 | — | 6.0 |
| 14 | Nessi | 6 | 4 | 5.0 |
| 15 | Today | 4 | 5 | 4.5 |
| 16 | The Shadow | 5 | 4 | 4.5 |
| 17 | AVA | 4 | 4 | 4.0 |
| 18 | Joshu | — | 3 | 3.0 |
How this report works
Included are assistants whose catalog entry (checked 2026-10-26) records no training on your data by default, taken from each product's privacy policy. They are ranked by the average of Permissions & privacy and Memory, with at least one taken; Memory also checks that you can see and delete what it stored. Scores come from the published score sheet and are dated 2026-10-27. 14 of the 18 ranked assistants (Neo, Vellum, Bantr, Familiar, Ask Ranger, Friday, Norton Family Assistant, Luke, Persona, Nessi, The Shadow, Today, AVA and Joshu) carry pre-filled scores, which are placeholder estimates the site has not yet replaced with its own test runs, so treat their places as provisional. Hark, Ollie, Del and Caddy are not pre-filled.
How to choose
- Read the policy line yourself. Our catalog records what the privacy policy says on the day it was checked, 2026-10-26, and policies change.
- Use Permissions & privacy and Memory as the behaviour check: they test what it can access, whether it follows your rules, whether it lets go when you disconnect and whether you can see and delete what it stored.
- Mind the data it still handles. Not training is different from not storing: check the product's own deletion and export steps, which our catalog also records.
Which AI assistants do not train on my data?
21 of 47 catalog entries record that they do not train on your data by default: the list on this page. 6 record that they do (Alfi, Instinct, Muse, Pally, Poke and Wajo / Fo) and 20 are unclear.
Does a no-training policy make an assistant private?
No. It covers one use of your data. Our Permissions & privacy test checks what the assistant can access and whether it lets go when you disconnect, and Memory checks that you can see and delete what it kept.
Which of these have the best privacy and memory scores?
Hark has a Memory score of 10 on one test, Neo averages 8.0 over two and Ollie has a Permissions score of 8 on one. Hark and Ollie are not pre-filled; Neo is, so its place is provisional.
Frequently asked questions
Which assistant is best for this?
Hark leads with an average of 10.0, ahead of Neo and Ollie.
Is there a good free option?
Hark is the highest-ranked assistant with a free plan, averaging 10.0.
How did you test these assistants?
Every assistant got the same published tests, word for word, scored from three checkpoints and the number of corrections it needed. Each test is on the Dimensions page.