I suppose I should mention this, as I had to deal with it today, and it probably has a bearing.
One of my utility suppliers has switched to a ChatBot interface (which doesn't work on my prefered browser, just "spins and loads" nothing[1]... but switching over elsewhere gets it going) for the "Pay your bill" section. It used to go to the bit to put payment details in, but now it goes through a speech-bubbly-like chat interface[2] that tries to determine what you actually want to do. (Not even with Natural Language Programming, the stages I had to go through, just multichoice option selection. Maybe if I'd have gotten out to "No, something else..." options, then it'd have offered freetext as a prelude to resorting to pestering a spare human logged in and waiting on the other side of the system's cloud-based intermediary.
After several steps, I'm at pretty much the same actual payment screen as previously I'd been sent straight to by this system (one of the things I've not yet switched to Direct Debit) and its "It is now time to pay" reminders.
In my case, a lot of that was pointless. I suppose I'm funelled into a processing system that is clearly shzuzhed up to "look modern and bling" that
may be designed well enough to then send everyone into the appropriate call-handling substacks relevent to them, and keep service humans out of the loop until strictly necessary. But the immediate predecessor to this system had me just go straight to where the relevent email wanted me to go[3]. Though also not perfect, to this cynical old hand with data security, it didn't have so many opportunities to be too opaque for its own good.
Caveat: pretty much nothing of the above was probably "proper" AI, although I suspect NLP/LLM modules were sitting there awaiting any input from me beyond the trivial. And it's
technically easier to audit the ability for a computer to keep secure any passing private details it gets told about than you might with a human call-handler (assuming you've secured the communications system across to whichever). But it seems to me to be more Bandwagon plus Competitive Tender plus "because we can" rather than a properly considered solution.
(And remember the line from Demolition Man, along the lines of "Greetings and salutations. Welcome to the emergency line of the San Angeles Police Department. If you'd prefer an automated response, press 1 now." As much a mockery of the current situation as it was with phone-tree systems of the era. And today, still, though often prefixed by the long intro that often includes "go to
www.OurCompany.com slash Chat, or use the OurCompany app, for instant assistance".)
Back to AI (full and proper), though, also adding that a wiki that I edit has started getting contributions either apparently or
informed as being what someone got when they asked some ChatGPT/equivalent to summarise something. Not (so far as I know) ever completely automated, but definitely of the form not far off "please write my essay for me". Not even a school essay that had to be submitted but the person couldn't be bothered (or just couldn't?) write themselves, so they cheat, but someone having the
option to contribute a paragraph or two, doesn't do anything significant about it themselves but does go off to a tame LLM and exorts it to write something that (
perhaps after cursory checks and tweaks) they copypaste over. What's the point of that, then?
Ther
is always the possibility of sufficient perfection of the method that it's either no worse or
better than the humans (who
do have time, but maybe not the right idea), and be happily undetectable. But we're probably not there, yet, in any reliable way. As the copious examples that
haven't sneaked under the radar have pretty much proven.
[1] Doesn't even have the good grace to say "I'm sorry, we've arbitrarily decided that, because you don't support SVG version 3001.zeta and/or JavaScriptPremium with the 'import antigravity' module, we can't continue. Please upgrade your browser by waving your sonic screwdriver at it whilst saying 'I believe in faeries'...".
[2] Very wasteful of screen space, sitting there in the middle of a perfectly good wide monitor, but of course it's geared towards the 'phone experience'.
[3] Yes, all kinds of problems with random "Click here and pay" links. But the change from the recent way to this newer one barely helps. Fraudsters design fake interfaces all the time[4], and actually checked the chatbot's URL the first time I encountered it l, due to it resolving into a cloud-servery subdomain of one of the suppliers different domain domains[5]. Given it's unlikely to be an entirely inhouse project, but a service bought from a ChatBot B2B provider that gets tweaked to be slightly branded as appropriate for their end-clients, of course you're going to get "the same kind of look", anyway. The 'old click' was actually far more trustworthy than this 'new click'.
[4] Witness all the "Hi, this is YourISP. You need to revalidate your email account before we close it, so please click Here..." that link to random wordpress domains, SurveyMonkey pages or who-knows-what-but-it-has-a-Guatamalean-TLD-in-the-URL. Which I
imagine leads to a page that is a rough facsimile of the kind of thing that anyone falling for the email's MyISP-branding would think is gospel.
[5] Depending upon what you want to check, you might end up at domains like
MySupplier.co.uk or
MySupplierLtd.net or
MSServices.info, so ending up at something as esoteric as an address of the form
cloud0xFB.MySupplier.ai/customerForm?CID=6hGi6DdkZ&SID=000005938536192&opt=foobarbaz is not totally unexpected!