Suggestion: AI posts should be quarantined from the rest of the forum

[Wetzelrad]

I would like to discuss my ideas with only humans, and I'm sure I'm not the only forum user that feels this way. If I wanted to argue with a chatbot I am perfectly capable of opening a conversation with one myself.

In recent months I have demonstrated repeatedly that ChatGPT frequently generates Holocaust "hallucinations", which I define as new, false information. Examples:

• Me calling out a hallucination which reversed the truth. https://www.codohforum.com/viewtopic.php?p=9736#p9736
• Me calling out (conservatively) 5 hallucinations. https://www.codohforum.com/viewtopic.php?p=10690#p10690
• Me calling out (conservatively) 11 hallucinations. https://www.codohforum.com/viewtopic.php?p=12956#p12956
• Me calling out a fabricated quote. https://www.codohforum.com/viewtopic.php?p=13146#p13146
• Me calling out continued usage of previous hallucinations. https://www.codohforum.com/viewtopic.php?p=13331#p13331
• ConfusedJew admits to a hallucination, going further to say, "The frustrating thing is that it isn't honest about why it makes mistakes". https://www.codohforum.com/viewtopic.php?p=13322#p13322

Hallucination is a behavior common to all AI. I've observed it with X's Grok and Gab's Arya on unrelated topics. This UNESCO paper supports my argument with additional examples:
https://www.gcedclearinghouse.org/sites ... 178eng.pdf

1. AI automated content may invent facts about the Holocaust.
AI models have produced misleading or false narratives about the Holocaust. Data voids and biases have led to “hallucinations” in generative AI systems, producing incorrect or invented content that never occurred. Without AI literacy and research skills, users may not know how to verify AI-produced texts, or recognize the unreliability of the data.

AI has invented events, personalities and even whole phenomena relating to the Holocaust; one example of such hallucination on ChatGPT is the concept of the ‘Holocaust by drowning,’ which assumes that mass murder campaigns took place in which Nazis and their collaborators drowned Jews in rivers and lakes. Although historically no such campaigns took place, AI invented them based on the concept of the Holocaust by bullets – i.e. large-scale murder by shooting. Google’s Bard has, on other occasions, hallucinated fake quotes from witnesses to support distorted narratives of Holocaust massacres (Makhortykh et al., 2023c).

Text-generative AI systems (e.g. Google Bard) may generate false information about lesser-known events during the Holocaust, basing instead their data sets on better-known histories of places and events for which information is more widely available on the internet. For example, Bard’s responses to prompts concerning the 1941 massacres in Liubar in Ukraine reiterate details from the more well-known narrative of Babyn Yar, resulting in historically inaccurate claims, both concerning how victims were killed and whether Ukrainian collaborators were involved in the killing.

While I'm not an AI expert, my understanding is that, when the AI takes a side on an issue but has insufficient justification for it, it's forced to invent and misinterpret evidence to support its beliefs. Since AIs are trained on a wealth of Holocaust-affirming material, and since Holocaust affirmation may even be baked into their secretive core prompts, it is not surprising that hallucination occurs frequently with this topic.

Among the problems with allowing AI posts to run wild on the forum:

• Users avoid accountability for the claims and accusations they make.
• Pollution of forum search results with false information.
• Pollution of web search results.
• Pollution of future AI training material.
• Time wasting repetition.
• Degeneration of discussion to instead talk about AI.

Possible solutions:

• Require all AI posts to be very explicitly designated as such.
• Only permit AI posts in threads that are tagged for AI, perhaps in the topic title.
• Only permit AI posts in a subforum designated for that purpose.

Any of these would alleviate the problems.

The question could be asked, how is a moderator to determine if a post is or is not using AI? After all, ConfusedJew was making AI posts without acknowledgement from his very first post. To me, AI posts are fairly obvious, so I don't think any complicated guidelines are needed here. Judgement can be used on a case-by-case basis. Otherwise, I'm sure there are objective ways of checking for AI usage, like the common trick of looking for em dashes (—).

[HansHill]

Could not agree more. On top of the suggestions Wetzelrad has made which I endorse, I have one additional suggestion, which unfortunately will only add more work to the Mod's current duties.

In the thread Wetzelrad has linked to, I made this (somewhat similar) point regarding absurd claims which could be extended to AI - see below. In that context, it could be:

- User flags the suspect post
- Poster has 1 chance to provide export of entire prompt for transparency
- Failure to do so gets the post shadow-realmed

This will disincentivise prompt-farming and hold the original poster somewhat to account for their material. Either way, great post Wetzelrad and I do think we can make progress on this.

For this i have a mod suggestion - if we (revisionists) were to flag a post which makes an absurd claim (such as the Kremas are no longer able to be examined) - the author is given one chance to offer a citation, and if this is not produced, the post is moved to a quarantine, or removed entirely.

This will not only help keep threads lean, but will over time incentivise these people to start hedging their words like "its possible that X is the case".

[Archie]

Clearly this has been a huge problem with CJ from the beginning. He is in over his head and lacks the knowledge necessary to debate, so he uses LLMs to try to make up the gap.

We have not had an explicit rule about presenting AI output as if it were an original post, but it should go without saying that posts shouldn't be plagiarized. Text that is not your own should be quoted. Everyone has noticed CJ's brutally obvious copy-pasting, and he's been called out for it repeatedly. I have given him multiple warnings.

For example,
https://codohforum.com/viewtopic.php?p=8788#p8788
https://codohforum.com/viewtopic.php?p=9749&#p9749

He completely ignored these. Given his refusal to change, the only way to make it stop would be to ban him. I have held off on doing that as it seemed some posters wanted to continue engaging with him. If anyone feels strongly about letting him stay, please send me a private message.

I don't have a full-fledged policy worked out on all the potential AI issues. Here are some basic guidelines which we can adapt over time.

• No plagiarism. Feel free to report posts that seem to be plagiarized.
• Quoting AI output that is relevant to the discussion with attribution is okay but please use discretion. Make sure it is interesting and do it within reason.
• Threads about AI (developments and news and so forth) are completely fine as it is a useful productivity tool, and it relates to issues of censorship and control of information (one of our major themes).
• It can be tempting to try to use AI as an arbiter on contested points but this doesn't seem to work well since people can just shop around and tweak the prompts to get one to agree them. Moreover, it seems the more recent models have been calibrated to be extra agreeable and flattering to the user.

[Archie]

For this i have a mod suggestion - if we (revisionists) were to flag a post which makes an absurd claim (such as the Kremas are no longer able to be examined) - the author is given one chance to offer a citation, and if this is not produced, the post is moved to a quarantine, or removed entirely.

This will not only help keep threads lean, but will over time incentivise these people to start hedging their words like "its possible that X is the case".

We have something along these lines in the current rules.

"No Dodging" If you make a controversial claim without support, others have the right to request support. You must respond in some fashion, either by explaining your basis for the claim (whatever it may be) or by conceding that support is lacking. All posters need not agree on whether the support provided is convincing, but outright refusal to respond will be considered "dodging." Citing overly vague support (like a lengthy source with no explanation) may also be considered non-responsive. You are only obligated to respond to demands regarding claims you yourself have made.

The idea here is that you must, upon request, be prepared to explain the basis for any claims you make. I might add a sentence in here saying that "AI told me" is not an acceptable source. You have to track down a real source.

This rule is a more relaxed version of a rule on the old forum. Under the old rule, you were, in theory, obligated to respond to any challenge presented to you. But I don't agree with that since I think posters are at their best when they participate in topics where they are strongest. And I do not see why any other user has a right to dictate what another poster posts about. But if you make poorly supported claims, then you are opening the door to be challenged.

[HansHill]

it seemed some posters wanted to continue engaging with him.

I am guilty here but please don't take my continuance as an endorsement! In fact, I've been putting in so many hours in here, were the unthinkable to happen, it just might save my marriage

[Stubble]

Concerning the CJish question, banning isn't necessarily the only option is it? Perhaps a suspension with an explanation that AI output needs to be presented in quotes would bring some kind of reckoning to the mush between his ears. Might even result in self deportation.

He claims he values the board, and perhaps in his own queer way he does. Ultimately the CJiberish he presents to the forum is a type of pestilence built brick by brick on misconception, blithe dismissal and straight up lies, not even forged by his own tongue and there is a real miscarriage or travesty in that.

I accept that I don't always do a good job 'tending the garden' either. I do at least put my own thoughts on the table and I don't simply vomit regurgitated tripe spoonfed by AI.

[Wetzelrad]

He completely ignored these. Given his refusal to change, the only way to make it stop would be to ban him. I have held off on doing that as it seemed some posters wanted to continue engaging with him. If anyone feels strongly about letting him stay, please send me a private message.

I will PM to oppose only because I prefer less strict moderation generally. With the implementation of one of my suggestions, I would take his posts less seriously, as would everyone else.

I don't have a full-fledged policy worked out on all the potential AI issues. Here are some basic guidelines which we can adapt over time.

• No plagiarism. Feel free to report posts that seem to be plagiarized.

I find it odd to call this behavior plagiarism since it treats AI as if it has rights that can be violated. I could make some arguments here, like how can AI be plagiarized if AI output according to some U.S. courts cannot be copywritten? But putting semantics aside...

Yes, this is an important rule to deal with AI posts. However, I will make the argument in my next post that this issue is best remedied by standardization.

• Quoting AI output that is relevant to the discussion with attribution is okay but please use discretion. Make sure it is interesting and do it within reason.

This would be the most important rule, but is "please use discretion" enough? That is basically an open policy for anyone to enter any thread and paste arguments they borrowed from AI. If I want to have a serious debate thread, I don't want slop replies, nor do I want to read slop replies to other people's threads. Not even if they do it politely and discreetly.

Plus, what if some users want to engage in an all-out AI battle? Can they not mark their thread accordingly and dispose with discretion?

• Threads about AI (developments and news and so forth) are completely fine as it is a useful productivity tool, and it relates to issues of censorship and control of information (one of our major themes).

Agreed.

• It can be tempting to try to use AI as an arbiter on contested points but this doesn't seem to work well since people can just shop around and tweak the prompts to get one to agree them. Moreover, it seems the more recent models have been calibrated to be extra agreeable and flattering to the user.

This is true but it's totally unavoidable. People bringing AI into a discussion are more likely to select an AI that agrees with them, and there is no such thing as an unbiased AI that you can force them to use. So this guideline wouldn't do anything.

[Wetzelrad]

Reworking my previous list, these are the options. They are to be read as discrete policies.

1. Require all AI posts to say they are quoting AI. (current policy)
2. Require all AI posts to be tagged conspicuously as AI.
3. Only permit AI posts in topics that are tagged as AI in the topic title.
4. Only permit AI posts in a new AI subforum.
5. Ban or substantially restrict AI usage.

After some consideration, I lean toward either 3 or 4 as my preferred policy, but let me make the argument for 2 as well.

If a user wants to avoid AI posts, then they are forced to scroll through AI posts. Are they to just suffer through it? If they make a thread or post, are they meant to ignore posts that do reply to them with AI? Do they have to mark their own posts with "don't respond with AI!" and would that be respected? Or what of the forum more generally -- if someone like CJ decides to offer a challenge in the debate section, is this user meant to ignore that challenge because it was clearly written by AI?

These are all issues that either disadvantage the user or make the experience less user friendly. The best solution would be one that makes it so that AI posts are either clearly marked or separated from the rest.

Solution 4, though somewhat extreme, would handily solve the problem. No one would have any question what is or isn't AI. The major advantage with this solution is that people could go hog wild, pit AI against AI directly with no concern for human attention spans or other limits. This could even be a motivator for building a specialized AI with the capability to debunk any Holocaust narrative.

Solution 3 has the benefit of informing the user whether or not the thread is worth clicking on ahead of time. Users that don't want to waste time arguing with a chatbot can simply regard it as beneath them.

Solution 2 seems somewhat less satisfactory, but if there was a standardized tag at the top of every AI post, it would be easy enough to regard it appropriately and scroll past it. Without this requirement, people might hide it at the bottom of the post or otherwise disrespect the purpose of the rule.

Some added thoughts. This forum uses a linear format in which all posts are treated with equal respect. Other websites like Reddit or X rank posts algorithmically so you can expect slop replies to be seriously downranked where no one will see them. Because of this difference, the rules may need to be different here, perhaps more strict. However, even those websites have some rules for the use and labelling of AI, though not very consistent or clear ones. See for example the many X accounts marked as "Automated" like @PossumEveryHour, or the Reddit subforums that have banned all AI content.

I think special consideration should also be given to how this website may be used by human guests and AI. This forum does show up on Google web search results. Example:

search results example.png (73.01 KiB) Viewed 322 times

This has also been the case in the past, when search engines had sometimes brought me to the old CODOH forum. Links from websites like Metapedia may also bring people here. Likewise, users may search the forum directly for information on a specific topic. For all of these use cases we should want the posting quality to be high. AI ruins that.

There is also the distant use case in which AI could be trained on material from this forum. We might even make that AI ourselves. This becomes more possible if the forum is not polluted with nonsense.