Hello, you are absolutely correct. Maybe the best way to phrase it in this
context was 'the detector checks against a secret key, and with every match
it becomes more likely that Claude edited the text'. I was trying to convey
that the system could detect that it is increasingly unlikely that Claude
didn't write the text, but i should have been more rigorous with the
phrasing. Apologies for that.

Regarding the main point about the password / login experience, I'm still
curious whether this approach could be adapted for humans, perhaps by
deriving a secret key on the fly?

On Fri, 6 Nov 2026 at 06:23, Charles Haynes via Silklist <
[email protected]> wrote:

> That was my understanding too.
>
> On Thu, Nov 5, 2026, 11:53 PM Tim Bray via Silklist <
> [email protected]> wrote:
>
>>
>>
>> On Nov 5, 2026 at 2:00:54 AM, Advaith Mohan via Silklist <
>> [email protected]> wrote:
>>
>>> Agreed, but as a backup to a regular password, I feel it would be a
>>> better use case than the forgot password flow or equivalent in case of an
>>> emergency.
>>>
>>> The way the Anthropic text watermarking works is by detecting
>>> increasingly statistically improbable sequences of words in any given set
>>> of sentences / tokens.
>>>
>>
>> Hmm, my impression is that it depends on the PRNG seed used to select
>> from the top-ranked next-token candidates; I wrote it up in
>> https://www.tbray.org/ongoing/When/202x/2026/08/24/Text-Watermarking
>>
>> -Tim
>> --
>> Silklist mailing list
>> [email protected]
>> https://mailman.panix.com/listinfo.cgi/silklist
>>
> --
> Silklist mailing list
> [email protected]
> https://mailman.panix.com/listinfo.cgi/silklist
>
-- 
Silklist mailing list
[email protected]
https://mailman.panix.com/listinfo.cgi/silklist

Reply via email to