Straight up disingenuous
Altering logprobs to encode a key is neither random nor transparent with respect to output
This post is just
"It is a watermark that is completely invisible except that's impossible because then it wouldn't be a watermark"
Altering logprobs to encode a key is neither random nor transparent with respect to output
This post is just
"It is a watermark that is completely invisible except that's impossible because then it wouldn't be a watermark"
We’ve written an FAQ to answer some of the questions we've received about watermarking.
In summary:
• We’re implementing watermarking to comply with the EU AI Act. Other major model developers have signed the same Code of Practice and will also be implementing watermarking;
• Our watermarking method doesn’t have any practical impact on the quality or content of Claude’s outputs;
• The difference between watermarked and un-watermarked text will not be distinguishable to readers;
• Nothing is added to the text and there are no hidden characters;
• Watermarking doesn’t require extra tokens, and will not be more expensive;
• Watermarks can’t be traced to a specific person, organization, or chat.
Read more: anthropic.com/news/claude-te…
1
0
6
434
@itsyourcode Do you understand how it works? Are you familiar with SynthID?
1
0
1
24
@jon3k Yes. As I said it alters tail logprobs in a narrow distribution, thereby changing the output.
Aug 16, 2026 · 1:51 PM UTC
1
0
0
16
@jon3k And just to be clear, I don't care as much about watermarking itself. What annoys me is the attempts to market is something that is "doesn't change the output" and "imperceptible to humans" when that is obviously not true.
1
0
0
16
@jon3ksearch community is defending it at a technical claiming by saying things like "it just changes the seed (the secret is the seed)
But think about that: if you lock a seed, instead of randomizing one, you are locking in a subset of pathways. That's the point of seeds. Even the longest key is subset of all possible seeds.
This is why they have to disable it entirely on low entropy responses, otherwise it would totally break the output. This also means it's not a real watermark. It's a best-effort-high-entropy-distribution-bias.
Again I don't really care about the watermarking itself. It's a very weak mechanism. It's the marketing that's annoying.
But think about that: if you lock a seed, instead of randomizing one, you are locking in a subset of pathways. That's the point of seeds. Even the longest key is subset of all possible seeds.
This is why they have to disable it entirely on low entropy responses, otherwise it would totally break the output. This also means it's not a real watermark. It's a best-effort-high-entropy-distribution-bias.
Again I don't really care about the watermarking itself. It's a very weak mechanism. It's the marketing that's annoying.
0
0
0
14