back to article Could OpenAI's 'too dangerous to release' language model be used to mimic you online? Yes, says this chap: I built a bot to prove it

A machine-learning software engineer has trained OpenAI’s too-dangerous-to-release language model on personal Facebook messages to show how easy it is to create a bot that can attempt to impersonate you. Last month, researchers at OpenAI revealed they had built software that could perform a range of natural language tasks, …

  1. Pete 2 Silver badge

    Too late!

    > A machine-learning software engineer has trained OpenAI’s too-dangerous-to-release language model on personal Facebook messages to show how easy it is to create a bot that can attempt to impersonate you.

    I have a facebook account that I never use (in fact, I have several). Therefore even a 'bot that does nothing is already impersonating me.

    1. Anonymous Coward
      Anonymous Coward

      Re: Too late!

      I have a facebook account that I never use (in fact, I have several). Therefore even a 'bot that does nothing is already impersonating me.

      Well, PROVE IT! don't even have a facebook account, but I'm sure that a well-crafted message to my dis-believing friends, explaning, in my own words, how I changed my mind and click on this link to my all-new facebook page...

      Long term, increasingly advanced remote impersonation might force radical changes into how people communicate remotely, and how actual remote communications is performed...

      p.s. forget about pigeons, with current rate of drone development you'll soon have them impersonate pigeons on the fly!

      1. Glen 1 Bronze badge

        Re: Too late!

        On the fly... I see what you did there :3

  2. petethebloke

    In English

    I'm vaguely surprised that his conversations are in English.

  3. Anonymous Coward
    Anonymous Coward

    That's not "intelligence", it merely shows boredom

    Come on, let's face it: most people on FB live perfectly normal, not-terribly-exciting lives so there are not that many variables you have to take into account to spin up a viable "conversation" - I'd say that an AI is overkill to cope with the variables, but it is much easier to let an AI sort out for itself what to pick up from conversations that to design the filters and replies.

    Two more things to make it real for the average FB user:

    - some randomisation (and store per simulated identity so it keeps track of what it has told to whom)

    - update thispersondoesnotexist.com so it generates faces + changing backgrounds (btw, the latter now has an explanation with the pics).

    If you get that up, nobody would be able to tell the difference. As a matter of fact, it would be the *only* way I'd ever have a FB entry: a wholly fake one.

    1. herman Silver badge

      Re: That's not "intelligence", it merely shows boredom

      A https://thispersondoesnotexist.com/ FB profile will only require a very simple text bot. Something that can say: 'Yes', 'No' and 'Where's my tea?'.

  4. Denarius Silver badge

    if not already said before

    Why does Hitchikers Guide come to me where Arthur is told the Mice want his brain and Zaphod says, not much to emulate or words to that effect. Its been years ^H^H^H^H^H decades since I read it, OK ?

  5. herman Silver badge

    This type of Bot could make a good newspaper sports reporter. Lots of sports reporting is already automated - soccer and horse race scores for example. A half decent text bot could make the result more humane and user friendly.

    1. Anonymous Coward
      Anonymous Coward

      "This type of Bot could make a ..."

      ... typical TheRegister commentard - just need to train a set of standard responses to key words like Apple, Google, MicroSoft, Windows, Android, iOS, Linux, etc and it would probably pass a Turing test around here.

      1. MiguelC Silver badge
        Coat

        On the other hand, some commentards already seem to have some difficulty in passing the Turing test...

        1. Anonymous Coward
          Anonymous Coward

          MiguelC, that's EXACTLY the user I was thinking of - as well as the one that only ever seems to comment on the first one's post...

      2. This post has been deleted by its author

  6. Anonymous Coward
    Anonymous Coward

    "The output wasn't particularly coherent..."

    another proof it was authentic...

  7. ratfox Silver badge
    Paris Hilton

    "What are you wearing?"

    1. Paul Crawford Silver badge
      Gimp

      "Are you sure you want to know?"

    2. Anonymous Coward
      Anonymous Coward

      Y-fronts, a string vest and I'm eating a scotch egg seductively, how about you?

      Let's see AI emulate that.

    3. GrapeBunch Bronze badge

      "What are you wearing?"

      The cats broke the string holding the purple disco ball up in the light of the living room window. But my purple balls have never fallen off.

  8. Anonymous Coward
    Anonymous Coward

    Keeping things unreleased hinders research

    yes, a great shame!. Think of all the offers you might get if you had some brains and access to that research. Three letter agencies, mass-marketing, darkweb, and let's not forget the actual good-for-humanity research!

    1. Mark 85 Silver badge

      Re: Keeping things unreleased hinders research

      Be careful and afraid of what you wish for. Imagine for a moment that it's used to impersonate a politician.

      1. Anonymous Coward
        Anonymous Coward

        Re: Keeping things unreleased hinders research

        It already is. The real Trump lives at Mar-el-Lago revising his monumental history of the Thirty Years War and running his charitable foundation that supports refugees, while having dinner parties with Al Gore, Bernie Sanders, Elizabeth Warren and Nancy Pelosi.

    2. Michael Wojcik Silver badge

      Re: Keeping things unreleased hinders research

      I still believe this is 90% marketing fluff, and at best 10% a relatively unexciting improvement in the state of the art. Anyone who actually pays attention to Natural Language Processing research knows that we already have models that do a very good job of generating usable, believable prose in a number of domains.

  9. Doctor Syntax Silver badge

    It can always be used for good as well as evil. We could have a new, even funnier Lenny.

    1. JetSetJim Silver badge

      I want it to reply to all my boss's emails

  10. Anonymous Coward
    Anonymous Coward

    Sleep

    "was too hard to sleep last night"

    Funniest when you accidentally read it with an I at the beginning. DAMMIT FREUD!

  11. BrownishMonstr

    Damn it's a scary world. Let's ban AI..and computers.

    You know, to protect the children and all that.

  12. rab_au

    Keep a lid on it - LOL

    We all know how well that works. It's only a matter of time and everyone will have access to similar functionality.

    This could actually end the Internet.

    1. Michael Wojcik Silver badge

      Re: Keep a lid on it - LOL

      Everyone who cares to learn about Natural Language Processing and do some work has "access to similar functionality" now, and has had for a while. This whole Open AI kerfluffle is a lot of marketing bullshit.

  13. ThatOne Silver badge
    WTF?

    Sorry, what's the problem again?

    So, what's new? That you can automate spamming? I personally don't care if the spam is made by humans or computers, it remains spam and is treated like any other spam.

    The scamming perspective is more troubling, but scammers will always go for quantity over quality, so taking the time to train an AI with victim-related data is highly unlikely - except for high-profile spear fishing, and that would be better handled by a human anyway.

    IMHO it's like saying burglars can use a hydraulic press to break the door to my apartment. While true, it's not my biggest concern. Maybe I missed something?

  14. Starace Silver badge
    Flame

    It's garbage

    Like many attempts at AI stuff it's a pretty feeble attempt that only superficially works - the model can chuck out stuff that's statistically about right but any sort of close look at the output shows how badly it actually functions.

    Keeping the model secret is more about it being easier to maintain hype for something that can't be tested independently rather than any risk of it being misused.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019