- From: Patrick Logan <patrickdlogan@gmail.com>
- Date: Fri, 17 Feb 2023 09:03:34 -0800
- To: Hugh Glaser <hugh@glasers.org>
- Cc: David Booth <david@dbooth.org>, SW-forum <semantic-web@w3.org>
- Message-ID: <CAD_aa-_dg23J0qxr1bxQti23CxcX2Lsh5gvxvh3+A0P=H1SsDw@mail.gmail.com>
One potential difference is that spam as per its intent must be "imitation to a degree" -- ultimately the spam by design must have included a revelation of its true intent. But a LLM spoof intending to "imitate to the fullest degree" doesn't necessarily have to include any revelation of the underlying intent. On Fri, Feb 17, 2023, 8:36 AM Hugh Glaser <hugh@glasers.org> wrote: > I disagree, David. > The Spam-fighting arms race is an example of huge success on the part of > the defenders. > I see an irony that you sent this message reliably to everyone, using > email. > You must remember when people were saying that email would soon be > unusable because of spam. > The attitude to LLMs spoofing as human and destroying our socio-technical > fabric sounds very similar. > > I have at least 5 main email addresses I regularly use, and make no > attempt whatsoever to hide any from the world - I completely rely on > different spam filters on different providers to make it useable. > IIRC even years ago, over 99% of the incoming email to our university > server was spam, and yet I would rarely see one, and still don’t from there. > Yes, a few good fakes get through - but the number make us forget the > overall success of the spam-fighting enterprise. > Astonishing success, I would say. > > And as for lying, Thomas, why do you think I would have a problem with > something I am paying for deliberately lying (if I understand what you mean > by “lying”)? > I mean yes, hallucination and other stuff are a problem, but that is > exactly the interesting stuff to investigate. > > Let’s be Pollyanna rather than Cassandra :-) > > Cheers > > > On 17 Feb 2023, at 15:55, David Booth <david@dbooth.org> wrote: > > > > On 2/17/23 08:54, Thomas Passin wrote: > >> On 2/17/2023 8:36 AM, Hugh Glaser wrote: > >>> Has anyone tried using LLMs such as GPT-3 to find out if text is > human- or machine-generated? > >>> Can’t you just give it the text and ask it? > >> Except that they may lie or "hallucinate". > > > > And I think our experience with the spam-fighting arms race has already > answered that question in general: we can detect the crudest fakes, but > better fakes will always get through. > > > > Fake generators have an inherent advantage, because fakes can be > generated by the millions so cheaply, and the generators can be programmed > to randomly try different techniques and learn which techniques get past > the detectors. > > > > David Booth > > > > >
Received on Friday, 17 February 2023 17:03:58 UTC