Thank you for shedding this issue. You always seem to be at the crest of the wave regarding matters that will impact most of us. I assumed, wrongly I suppose, that "watermarks" might minimize this problem.
Great article. I, too, get very frustrated when something I write is identified as AI. I spend so much time trying to rewrite a thought to avoid that moniker, and then it still gets nailed because of a word selection or certain rhythm I use. It's extremely frustrating.
Having said that, I freely admit to letting AI tools review my work and recommend changes to grammar, tone, and flow, but the words are always mine. Like you, I use AI to give me title suggestions, and I use it to research sources and the accuracy of any claims I make.
I'm curious how these AI detector tools work and why so many of them never really explain why something is identified as AI. They just give a blanket percentage score with no real justification. It's got me doubting damn near everything I write now and struggling to finish pieces.
There is also the issue of how is Pangram using the scanned content to teach its LLM. When an author opts in to have their work scanned and reports false positives to improve detection accuracy, I often wonder who is actually doing the labor. The writer may take it as their protecting their work. To me it’s a scheme wherein the author is now training the model. At no cost to Pangram. This feels more like a data acquisition strategy with a pretend veil of ignorance attached to it. And, the question of compensation for that labor is tactically not part of the conversation.
The slop problem is real but it is not primarily a technology problem; it’s a market problem. End users and the institutions that aggregate their attention (newspapers, publishing houses, recommendation algorithms), reward slop over quality to the point that it can’t be defensible. When a publishing house uses follower count as a key performance indicator for acquisitions, it has stopped being in the literature business. It is in the audience arbitrage business. Those perverse incentives produce different outputs.
Substack’s scale model is a reasonable structural bet. A large enough pool of writers across enough genres will produce excellent work as a statistical outcome. Notwithstanding, a platform optimized for volume rather than quality will bury that excellent work under the content that performs ( I sometimes get more notes with selfies than writing pieces).
If Substack were serious about the quality argument it makes implicitly by positioning itself against social media, the interventions are not complicated. Flag accounts using the platform as a content distribution tool rather than a writing practice, the same way Instagram flags inauthentic engagement. Weight the recommendation algorithm toward long-form. A piece that takes more time to read is not rewarded by current feed logic but it should be. There is also a straightforward commercial model sitting unused: sell books. Amazon and Barnes and Noble built the infrastructure. Substack has the authors and the audience.
The AI detection tool is the least interesting solution available to a platform with Substack’s position. There are better options. Most of them require the platform to have a considered position on what good writing is and why it is worth protecting. Moving towards a more exploitative model without the scale model that has made way for Substack's success to date would be a good start. But this proposition requires courage that is harder to ship than a feature.
Sidenote: I wrote this comment and ran it through the AI detector which concluded it was only 85% human. Go figure! Now I want to do a 23 and me test and see if my genetic make up shows that I am 25% AI.
Thank you for providing clear and succinct insights on this important issue. Much to think about.
this was such a great post; thank you.
Thank you for shedding this issue. You always seem to be at the crest of the wave regarding matters that will impact most of us. I assumed, wrongly I suppose, that "watermarks" might minimize this problem.
Great article. I, too, get very frustrated when something I write is identified as AI. I spend so much time trying to rewrite a thought to avoid that moniker, and then it still gets nailed because of a word selection or certain rhythm I use. It's extremely frustrating.
Having said that, I freely admit to letting AI tools review my work and recommend changes to grammar, tone, and flow, but the words are always mine. Like you, I use AI to give me title suggestions, and I use it to research sources and the accuracy of any claims I make.
I'm curious how these AI detector tools work and why so many of them never really explain why something is identified as AI. They just give a blanket percentage score with no real justification. It's got me doubting damn near everything I write now and struggling to finish pieces.
There is also the issue of how is Pangram using the scanned content to teach its LLM. When an author opts in to have their work scanned and reports false positives to improve detection accuracy, I often wonder who is actually doing the labor. The writer may take it as their protecting their work. To me it’s a scheme wherein the author is now training the model. At no cost to Pangram. This feels more like a data acquisition strategy with a pretend veil of ignorance attached to it. And, the question of compensation for that labor is tactically not part of the conversation.
The slop problem is real but it is not primarily a technology problem; it’s a market problem. End users and the institutions that aggregate their attention (newspapers, publishing houses, recommendation algorithms), reward slop over quality to the point that it can’t be defensible. When a publishing house uses follower count as a key performance indicator for acquisitions, it has stopped being in the literature business. It is in the audience arbitrage business. Those perverse incentives produce different outputs.
Substack’s scale model is a reasonable structural bet. A large enough pool of writers across enough genres will produce excellent work as a statistical outcome. Notwithstanding, a platform optimized for volume rather than quality will bury that excellent work under the content that performs ( I sometimes get more notes with selfies than writing pieces).
If Substack were serious about the quality argument it makes implicitly by positioning itself against social media, the interventions are not complicated. Flag accounts using the platform as a content distribution tool rather than a writing practice, the same way Instagram flags inauthentic engagement. Weight the recommendation algorithm toward long-form. A piece that takes more time to read is not rewarded by current feed logic but it should be. There is also a straightforward commercial model sitting unused: sell books. Amazon and Barnes and Noble built the infrastructure. Substack has the authors and the audience.
The AI detection tool is the least interesting solution available to a platform with Substack’s position. There are better options. Most of them require the platform to have a considered position on what good writing is and why it is worth protecting. Moving towards a more exploitative model without the scale model that has made way for Substack's success to date would be a good start. But this proposition requires courage that is harder to ship than a feature.
Sidenote: I wrote this comment and ran it through the AI detector which concluded it was only 85% human. Go figure! Now I want to do a 23 and me test and see if my genetic make up shows that I am 25% AI.