4 ms·

Is it possible that google is now turning to sites that have a strong moderation policy, as a vetting strategy against LLM generated posts? (Maybe November 2022

Is it possible that google is now turning to sites that have a strong moderation policy, as a vetting strategy against LLM generated posts? (Maybe November 2022 is turning out to be a new kind of Eternal September for the rest of the Web...)

baq 3y ago
I'd go a bit further and ask what other choice do they have...? Unless they can make a reasonably effective LLM-generated content detection (sounds tough!) they're faced with a rather inconvenient problem of mining the internet for known human-generated content (while we still agree what 'to know' even means). Interesting parallels to https://en.wikipedia.org/wiki/Low-background_steel https://en.wikipedia.org/wiki/Low-background_steel.
- thatjoeoverthr 3y ago
  It’s certainly hard to detect authorship per se, but I expect very easy to detect content people produce using LLMs. You can target affiliate marketing blogs and other SEO garbage, entirely ignoring whether it’s literally written by human hands.
  - A_D_E_P_T 3y ago
    > It’s certainly hard to detect authorship per se, but I expect very easy to detect content people produce using LLMs. Yes. For now. Just flag the use of the words "delve" and "deep dive," flag bullet-pointed paragraphs and lists, and flag anything that closes with a too-neat one-paragraph conclusion — and you've just caught 85% of AIspam. But this is already changing, and in about 6-10 months I expect it'll be impossible to detect at more than 70% confidence whether an article is written by an AI or a human. (You could say: The human writes with soul, with style, whereas the AI is totally bland, irritatingly didactic, and generic. But you can quite easily coax an AI into imitating any literary style, and some of the best AI-generated content is warped in this way.)
    - tacocataco 3y ago
      > Yes. For now. Just flag the use of the words "delve" and "deep dive," flag bullet-pointed paragraphs and lists, and flag anything that closes with a too-neat one-paragraph conclusion Uhoh, is my copy pasta in danger of being snuffed out before anyone is served it? Look, I get how some people would consider it spam. However, the problem still remains and the solution has not changed! I have a life to live and can't spend every waking moment crafting a unique comment every time I need to reply to someone.
  - mvkel 3y ago
    This is the Netflix algorithm problem. They had an algorithm that perfectly recommended content to you. Then they scrapped it and went back to star ratings. Why? Netflix makes money from monthly subscriptions. They lose money when actually streaming content to users. So the ideal user from their standpoint is one who keeps paying but watches almost nothing. Thus, the company is incentivized to put out just enough good stuff so you won't cancel, but they don't want you binging. If Google sends users to pages and advertisers pay them, the incentive to address the content problem isn't there. In fact, LLM content may be another cash cow waiting in the wings.
    - baq 3y ago
      You're right, they never used the winning algorithm, possible for reasons you describe. The prize was a recruitment tool from the start, though. I don't think they ever intended to use any result.
reddalo 3y ago
I'd like this definition of "Eternal AI September": all the website that existed before November 2022 and kept posting the same kind of content will be considered safe. Everything else will be vetted to prevent LLM generated content.
Shawnj2 3y ago
I bet a lot of text generated content you see online is LLM generated and I'm not sure how exempt HN is from this. Probably more than most but in 2024 it is trivial to generate something with an LLM that could be assumed to be written by a human and post it. Most of the internet pre-November 2022 isn't AI generated (with the exception of one off things like Seychelles Anon, SubredditSimGPT2, etc.)
weinzierl 3y ago
Similarly HN is already an important data source for large language model training. For example, the RefinedWeb paper lists HN as one of only 12 websites that were excluded. From what I understand, it was excluded because it went into the final dataset unvetted. RefinedWeb was used for the Falcon model. https://arxiv.org/pdf/2306.01116.pdf https://arxiv.org/pdf/2306.01116.pdf
3abiton 3y ago
Also reddit exodus, aftee their api changes. Don't understimate people's imcline for good open source (or 3rd party clients)
- bayindirh 3y ago
  If your revenue stream depends on your client injecting ads to your users' feed and inability to remove that ad from the said feed unless paid for, you balk at 3rd party clients. Same is also true if you're implementing half of your moderation at client level.

baq 3y ago

I'd go a bit further and ask what other choice do they have...? Unless they can make a reasonably effective LLM-generated content detection (sounds tough!) they're faced with a rather inconvenient problem of mining the internet for known human-generated content (while we still agree what 'to know' even means). Interesting parallels to https://en.wikipedia.org/wiki/Low-background_steel https://en.wikipedia.org/wiki/Low-background_steel.

thatjoeoverthr 3y ago
It’s certainly hard to detect authorship per se, but I expect very easy to detect content people produce using LLMs. You can target affiliate marketing blogs and other SEO garbage, entirely ignoring whether it’s literally written by human hands.
- A_D_E_P_T 3y ago
  > It’s certainly hard to detect authorship per se, but I expect very easy to detect content people produce using LLMs. Yes. For now. Just flag the use of the words "delve" and "deep dive," flag bullet-pointed paragraphs and lists, and flag anything that closes with a too-neat one-paragraph conclusion — and you've just caught 85% of AIspam. But this is already changing, and in about 6-10 months I expect it'll be impossible to detect at more than 70% confidence whether an article is written by an AI or a human. (You could say: The human writes with soul, with style, whereas the AI is totally bland, irritatingly didactic, and generic. But you can quite easily coax an AI into imitating any literary style, and some of the best AI-generated content is warped in this way.)
  - tacocataco 3y ago
    > Yes. For now. Just flag the use of the words "delve" and "deep dive," flag bullet-pointed paragraphs and lists, and flag anything that closes with a too-neat one-paragraph conclusion Uhoh, is my copy pasta in danger of being snuffed out before anyone is served it? Look, I get how some people would consider it spam. However, the problem still remains and the solution has not changed! I have a life to live and can't spend every waking moment crafting a unique comment every time I need to reply to someone.
- mvkel 3y ago
  This is the Netflix algorithm problem. They had an algorithm that perfectly recommended content to you. Then they scrapped it and went back to star ratings. Why? Netflix makes money from monthly subscriptions. They lose money when actually streaming content to users. So the ideal user from their standpoint is one who keeps paying but watches almost nothing. Thus, the company is incentivized to put out just enough good stuff so you won't cancel, but they don't want you binging. If Google sends users to pages and advertisers pay them, the incentive to address the content problem isn't there. In fact, LLM content may be another cash cow waiting in the wings.
  - baq 3y ago
    You're right, they never used the winning algorithm, possible for reasons you describe. The prize was a recruitment tool from the start, though. I don't think they ever intended to use any result.

reddalo 3y ago

I'd like this definition of "Eternal AI September": all the website that existed before November 2022 and kept posting the same kind of content will be considered safe. Everything else will be vetted to prevent LLM generated content.

Shawnj2 3y ago

I bet a lot of text generated content you see online is LLM generated and I'm not sure how exempt HN is from this. Probably more than most but in 2024 it is trivial to generate something with an LLM that could be assumed to be written by a human and post it. Most of the internet pre-November 2022 isn't AI generated (with the exception of one off things like Seychelles Anon, SubredditSimGPT2, etc.)

weinzierl 3y ago

Similarly HN is already an important data source for large language model training. For example, the RefinedWeb paper lists HN as one of only 12 websites that were excluded. From what I understand, it was excluded because it went into the final dataset unvetted. RefinedWeb was used for the Falcon model. https://arxiv.org/pdf/2306.01116.pdf https://arxiv.org/pdf/2306.01116.pdf

3abiton 3y ago

Also reddit exodus, aftee their api changes. Don't understimate people's imcline for good open source (or 3rd party clients)

bayindirh 3y ago
If your revenue stream depends on your client injecting ads to your users' feed and inability to remove that ad from the said feed unless paid for, you balk at 3rd party clients. Same is also true if you're implementing half of your moderation at client level.