The meeting transcript getting flagged as 100% AI is almost too perfect. Apparently the best way to fool an AI detector is to be a human speaking naturally in a meeting. Considering the meetings I've been in, maybe that's not too far from the truth.
I’ve been circling the same distinction from the creative side: a detector might tell you who typed something. It can’t tell you who decided it.
“Is there anything of you in this?” might be the cleanest version of that test I’ve seen.
The watermark doesn't bother me as much as the assumption that it tells you something useful about the work. The meeting transcript part of this made me laugh. I've had detectors flag my own voice as AI before, so I know the number doesn't mean anything.
The last line is the part I keep coming back to: a watermark can tell you Claude touched a piece of writing, and it can tell you nothing at all about whether your own thinking went into it.
I find it ironic to the point of amusing that governments have regulated the distinction between human or ai authorship on the one hand and yet refuse to regulate the distinction of what concepts humans contributed to the models reasoning when the users work is being diluted across the model in the name of training, and in other ways, with impunity and great benefit due to this competitive advantage.
It’s also comical how far behind Government agencies are not only in regulating it, but understanding it, in the first place. Granted there is no precedent of this scale that i know of but, but this is clearly an example of the tail wagging the dog
For me, it is probably the crazy way I throw sayings both old and new in the middle of whatever I'm doing. I have never seen AI be able to copy that, though it is very good at noticing that I do it. Also, unless I am focusing on staying purely formal, I have this odd tendency to mix slangy and formal in the way that I have never seen AI be able to copy.
This is a great piece. There’s always an interesting balance between the AI assisted and individual human writing but I think the watermark does force people to challenge themselves. I find AI writing very annoying anyway (the question no one is asking ….the loudest voice in the room . Etc) so I often will take pieces of it but usually just voice record and use AI to clean. .
The marking duty sits on the provider of the system, not on the person writing with it. Article 50(2) of the AI Act applies from 2 August 2026, and the Commission's FAQ on the transparency code, last updated 29 July 2026, adds two details that get lost in the coverage. Outputs generated and already made available before 2 August 2026 do not need to be marked retroactively. And the AI Omnibus agreed by the co-legislators envisages a transitional period to 2 December 2026 for generative systems already on the market, if adopted. So the mark is a supplier compliance artefact with a moving start date, not a judgment on the work.
I tried that too, @Joel Salinas ! One of the most obvious failures of Pangram. My voice, transcribed during a meeting, flagged as 100 % AI. Ridiculous, really. My theory is that technical language short-circuits AI detectors.
And I added the line on my post from Anthropics press release, it’s flagged as ai even if you wrote 100% of it but ask ai to fix the commas and then copy paste it from Claude.
I was very against Pangram, but I’m much more open to this as there are two different things to me. Honestly, if it touched Claude, then we are ahead because we know how to use the tool in the way it serves us and the work but interested to see how this would roll out at the organizations.
The meeting transcript getting flagged as 100% AI is almost too perfect. Apparently the best way to fool an AI detector is to be a human speaking naturally in a meeting. Considering the meetings I've been in, maybe that's not too far from the truth.
I’ve been circling the same distinction from the creative side: a detector might tell you who typed something. It can’t tell you who decided it.
“Is there anything of you in this?” might be the cleanest version of that test I’ve seen.
Haha good point, we can get pretty robotic
Well framed. The thinking behind the writing still matters more than the label.
100% agree!
The watermark doesn't bother me as much as the assumption that it tells you something useful about the work. The meeting transcript part of this made me laugh. I've had detectors flag my own voice as AI before, so I know the number doesn't mean anything.
great point, it tells you 0 about the work haha
The last line is the part I keep coming back to: a watermark can tell you Claude touched a piece of writing, and it can tell you nothing at all about whether your own thinking went into it.
Yes! If there’s no you left in it, it will be replicated as soon as it starts working
I find it ironic to the point of amusing that governments have regulated the distinction between human or ai authorship on the one hand and yet refuse to regulate the distinction of what concepts humans contributed to the models reasoning when the users work is being diluted across the model in the name of training, and in other ways, with impunity and great benefit due to this competitive advantage.
Haha 100%. So much of it is appearing like they are doing something
It’s also comical how far behind Government agencies are not only in regulating it, but understanding it, in the first place. Granted there is no precedent of this scale that i know of but, but this is clearly an example of the tail wagging the dog
Oh 100%!
Pointing at the one thing only you could have written is such a simple, useful test to run before publishing.
For me, it is probably the crazy way I throw sayings both old and new in the middle of whatever I'm doing. I have never seen AI be able to copy that, though it is very good at noticing that I do it. Also, unless I am focusing on staying purely formal, I have this odd tendency to mix slangy and formal in the way that I have never seen AI be able to copy.
This is a great piece. There’s always an interesting balance between the AI assisted and individual human writing but I think the watermark does force people to challenge themselves. I find AI writing very annoying anyway (the question no one is asking ….the loudest voice in the room . Etc) so I often will take pieces of it but usually just voice record and use AI to clean. .
Yes thats a great strategy, Chris, and it makes pieces that are fundamentally yours and helped by ai rather that the opposite
Your test shows why I don’t put much faith in AI detectors. They’ll be hit and miss at best.
I care much more now about whether there’s a decent idea in the work and whether the person behind it has actually thought it through.
Exactly! That’s also what I think savvy readers will care more about too
The marking duty sits on the provider of the system, not on the person writing with it. Article 50(2) of the AI Act applies from 2 August 2026, and the Commission's FAQ on the transparency code, last updated 29 July 2026, adds two details that get lost in the coverage. Outputs generated and already made available before 2 August 2026 do not need to be marked retroactively. And the AI Omnibus agreed by the co-legislators envisages a transitional period to 2 December 2026 for generative systems already on the market, if adopted. So the mark is a supplier compliance artefact with a moving start date, not a judgment on the work.
That last line says it all so well
I tried that too, @Joel Salinas ! One of the most obvious failures of Pangram. My voice, transcribed during a meeting, flagged as 100 % AI. Ridiculous, really. My theory is that technical language short-circuits AI detectors.
And I added the line on my post from Anthropics press release, it’s flagged as ai even if you wrote 100% of it but ask ai to fix the commas and then copy paste it from Claude.
I was very against Pangram, but I’m much more open to this as there are two different things to me. Honestly, if it touched Claude, then we are ahead because we know how to use the tool in the way it serves us and the work but interested to see how this would roll out at the organizations.
Thank you for sharing your perspective!
It is missing so much the point to add a watermark just because a text was processed by Claude. Thinking is much more important than phrasing.