This is actually a call for research from YC-funded OpenAI:

> Spam the Spammers

> Investigate the use of language models to remove the profit from spamming.

> Spammers generate a huge amount of undesirable traffic and attention. Their emails are merely annoying for most people, but a small fraction of users fall into their trap. Spammers receive responses from users extremely infrequently. Therefore, they manually reply to each email.

> The task is to build a bot that automatically replies to spam emails. Such a bot shouldn't be easy to detect, which could be achieved by use of a powerful language model.

https://openai.com/requests-for-research/#spam-spammers

Could be an extension from Graham's A Plan for Spam, which basically called for a DDoS on spam servers:

> As I mentioned in Will Filters Kill Spam?, following all the urls in a spam would have an amusing side-effect. If popular email clients did this in order to filter spam, the spammer's servers would take a serious pounding. The more I think about this, the better an idea it seems. This isn't just amusing; it would be hard to imagine a more perfectly targeted counterattack on spammers.

> So I'd like to suggest an additional feature to those working on spam filters: a "punish" mode which, if turned on, would spider every url in a suspected spam n times, where n could be set by the user.

http://www.paulgraham.com/ffb.html

Considering that a lot of spam is sent via botnets, it would not take much for spammers to either protect themselves by hosting their pages on botnets, or use it to cause others to run a DDOS for them.

This is a DDOS of the humans, not their computers.

I was responding to the Paul Graham quote:

> So I'd like to suggest an additional feature to those working on spam filters: a "punish" mode which, if turned on, would spider every url in a suspected spam n times, where n could be set by the user.

EDIT: But note that the problem is the same if you reply to the e-mails. If this is automated and enough people do it, then sending spam with fake from addresses becomes an effective way of attacking peoples mail accounts. I've seen this happen first hand by a spammer that used incendiary content to trigger manual responses by sending out what claimed to be an ad for child porn sent in the name of an anti-spam activist. It was scarily effective, even though in this case it required getting people to take manual action (in this case bad enough that he needed police protection as a result of a range of credible threats)

Why is it a call for research? A simple n-gram model would be enough. It only needs to be almost grammatical, the spammers can't read all the messages anyways, if they want to fight that they need another machine classifier.

> Why is it a call for research?

I assume it's their catch-all term for things that they don't expect to be profitable as a business.

I think for this to work, they will have to make something that can pass the Turing test, at least for a certain period of time.

Anyone that thinks scammers won't generate their own version of the Voight-Kampff test and it won't quickly spread is likely to be disappointed.

Can the AI get away with imperfect grammar and sentence construction to signal that it is a less educated human? You'll be fighting fire with fire.

An interesting area of linguistics and natural language research: creating an AI that makes the same mistakes as a human.

This is discussed in Brian Christian's The Most Human Human (a book about participating in a Turing test contest); there was a chatbot that claimed to be a young person from another country and made a lot of mistakes. People found it pretty convincing and forgave some of its linguistic comprehension errors because they had a plausible explanation for why it wouldn't be good at English.

Sounds like "Eugene Goostman". It's common to "cheat" on the Turing test by pretending to be an incomprehensible human, like "PARRY" which "simulated" a paranoid schizophrenic, about which Daniel Dennett relates this anecdote in his essay "Can Machines Think?":

"My favorite commentary on it was Joseph Weizenbaum’s; in a letter to the Communications of the Association of Computing Machinery (Weizenbaum, 1974, p. 543), he said that, inspired by Colby, he has designed an even better program, which passed the same test. His also had the virtue of being a very inexpensive program, in these times of tight money. In fact you didn’t even need a computer for it. All you needed was an electric typewriter. His program modeled infant autism. And the transcripts--you type in your questions, and the thing just sits there and hums--cannot be distinguished by experts from transcripts of real conversations with infantile autistic patients."

It's well worth a read!

http://www.nyu.edu/gsas/dept/philo/courses/mindsandmachines/...

Eugene Goostman is exactly right -- and apparently genuinely fooled people.

Thanks for the reference.

Edit: It's very interesting and Dennett anticipates a number of Christian's observations, even in the part of the paper written before Dennett was associated with the Loebner contest.

Lenny is apparently quite effective against telemarketers:

http://toao.net/595-lenny