Re: ChatGPT > Spamassassin? :)

On 2024-06-25 at 17:38:28 UTC-0400 (Tue, 25 Jun 2024 17:38:28 -0400)
Mark London <[email protected]>
is rumored to have said:

> Bill - Thanks for the response.  As an aside, it would be nice 
> (though impossible?) for a spam filter to be more suspicious of emails 
> coming from a new email address, that is not in my Sent folder or my 
> Inbox. FWIW. - Mark

Matija's mention of AWL/TxRep is correct here. While some people find it 
a nuisance when it makes one FP into an ongoing series, I think it is 
worth enabling for most sites.

However, if you do enable either of those tools, you should have a 
mechanism for  feeding FPs into both a sitewide Bayes DB and into the 
AWL/TxRep DB by using the blocklist/welcomelist options of the 
spamassassin script.



> On 6/25/2024 11:21 AM, Bill Cole wrote:
>> Mark London <[email protected]>
>> is rumored to have said:
>>
>>> I received a spam email with the text below, that wasn't caught by 
>>> Spamassasin (at least mine).   The text actually looks like 
>>> something that was generated using ChatGPT.  In any event,  I put 
>>> the text through ChatGPT, and asked if it looked like spam.  At the 
>>> bottom of this email , is it's analysis.  I've not been fully 
>>> reading this group.  Has there been any work to allow Spamassassin 
>>> to use AI?
>>
>> "Artificial intelligence" does not exist. It is a misnomer.
>>
>> Large language models like ChatGPT have a provenance problem. There's 
>> no way to know why exactly the model "says" anything. In a single 
>> paragraph, ChatGPT is capable of making completely and directly 
>> inconsistent assertions. The only way to explain that is that despite 
>> appearances, a request to answer the ham/spasm question generates 
>> text with no semantic connection to the original, but which seems 
>> like an explanation.
>>
>> SpamAssassin's code and rules all come from ASF committers, and the 
>> scores are determined by examining the scan results from contributors 
>> and optimizing them to a threshold of 5.0. Every scan of a message 
>> results in a list of hits against documented rules. The results can 
>> be analyzed and understood.
>>
>> We know that ChatGPT and other LLMs that are publicly available have 
>> been trained on data to which they had no license. There is no way to 
>> remove any particular ingested data. There's no way to know where any 
>> particular LLM will have problems and no way to fix those problems. 
>> This all puts them outside of the boundaries we have as an ASF 
>> project. However, we do have a plugin architecture, so it is possible 
>> for 3rd parties to create a plugin for LLM integration.
>>
>>


-- 
Bill Cole
[email protected] or [email protected]
(AKA @[email protected] and many *@billmail.scconsult.com 
addresses)
Not Currently Available For Hire