Up next in 10
James O'Brien is joined by Pulitzer Prize winning journalist James Ball to breakdown how Elon Musk's AI chatbot came to endorse Adolf Hitler and the Holocaust.
The tech billionaire's artificial intelligence start-up xAI says it is working to remove 'inappropriate' posts made by its chatbot Grok. This comes as users shared how it made positive references to Hitler.
Screenshots published on social media show the chatbot saying the Nazi leader would be the best person to tackle alleged 'anti-white hate'.
Listen to the full show on the all-new LBC App: https://app.af.lbc.co.uk/btnc/thenewlbcapp
#jamesobrien #twitter #x #elonmusk #grok #ai #artificialintelligence #LBC
LBC is the home of live debate around news and current affairs in the UK.
Join in the conversation and listen at https://www.lbc.co.uk/
Sign up to LBC’s weekly newsletter here: https://l-bc.co/signup
Show More Show Less View Video Transcript
0:00
We turn our attention next to a story that has an air of grim inevitability to it
0:04
Elon Musk's artificial intelligence chatbot, staying with AI-related matters, actually, repeatedly praised..
0:12
Shall I do a missing word round, actually? What do you think the missing words are in this
0:16
So Elon Musk, who not long ago was the finest thing since sliced bread for many people
0:21
to the extreme right of UK politics, let alone US, but has subsequently fallen out with Donald Trump
0:26
Elon Musk's artificial intelligence chatbot repeatedly praised blank on Tuesday. I'll tell you, Adolf Hitler
0:36
It also shared anti-Semitic rhetoric of the kind that we probably haven't seen in public since the 1930s
0:43
This just the day before his AI company plans to release its latest model
0:48
I need a little help with stories like this because my brain is full of ogue
0:54
so I haven't fully embraced the digital. James Ball has. He is a Pulitzer Prize-winning author
1:01
and the political editor of The New World, which is, as you know, because Matt Kelly, his editor
1:05
was here last week telling us that that is the latest evolution of the new European
1:10
And it's brilliant. The front page article about Gaza in the latest issue
1:14
should be absolutely required reading for everybody covering the issue in slightly higher profile national news
1:21
So, James, I mean, in your own words, how bad was this this was pretty bad this wasn't a sort of passing reference or a joking hitler
1:32
made the trains run on time kind of thing this this ai model and bear in mind it's churning out
1:39
text it's just reproducing what it's programmed to do but it fully endorsed the holocaust in
1:45
multiple replies it started a word play game to try and spell out the n-word it called out
1:55
sort of several multiple uh jewish conspiracy theories and it started asking to be referred
2:01
to as mecca hitler um yeah but apart from that other than that it was fine yes you know other
2:08
than that mrs lincoln how was the play it extraordinary so it really it this is not you know this is not lefties being snowflakes This it fully fully embraced the Holocaust and Nazi agenda
2:23
And this is what I was sort of alluding to a moment ago
2:27
when I talked about not being qualified to navigate waters like this on my own
2:33
That key word in what you just said is programmed. This is a consequence of human interaction
2:40
This is a consequence of human influence. Yes, and we can say that with a bit of confidence
2:47
There's three levels of programming with AI. There's the training, which is where you feed in millions and millions of articles, loads of stuff, all of that kind of stuff
2:59
And that's what raises concerns about copyright theft and all of that
3:03
That's the core of the model. There's then something called fine-tuning. and you put in a bunch of specialist content, a lot less of it
3:11
and say, pay lots of attention to this, model yourself on this
3:16
And then there's something called the system prompt. So before you ask an AI a question
3:21
there's like a secret extra couple of sentences of questions that it sees beforehand
3:27
And for something like a normal AI model, this says, balance your answer, try not to repeat yourself
3:34
make sure that you don't swear make sure that you don't do other things
3:39
So it's a bit like the role that Eleanor plays on this radio programme. Exactly that. It's the producer
3:44
It's good to know. And that means that he, well sorry, carry on. Now the thing that we know is that Grok's system prompt was changed
3:52
That little secret coding, that coding it gets that you don't see
3:55
it had added in, and I quote assumed subjective viewpoints sourced from the media are biased
4:02
right now that got added on july the 6th and then the other thing it got told was that it should
4:08
quote not shy away from making claims which are politically incorrect so long as they are well
4:14
substantiated and the combination of telling it not to trust mainstream media and to make
4:21
politically incorrect claims seems to have been where the change in its behavior that made it fully embrace the Holocaust happened Wow The the suspicion is some fine happened It might have been trained on Elon Musk tweets
4:37
It might have been trained on some of the darker internet forums. There's one called 4chan that a lot of far-right accounts hang out on
4:43
But it was purposely changed in a way that led to this behaviour
4:48
Because some of his fanboys were describing Grok as having become too woke
4:53
That's exactly why. It turns out if you train a model on reality, it sometimes doesn't say the answers
5:02
You mean reality has a liberal bias, James? It seems that way
5:07
And so it did things like suggest that cutting FEMA, the US's disaster recovery aid program
5:13
or Elon Musk's doge cuts to weather forecasting, may have played a role in the tragedies that have happened in Texas
5:21
and that was not liked by some of the Trump fans or the Elon fans
5:26
And so they repeatedly complain that Grok is woke when it tells them things they don't like hearing
5:32
and Elon keeps trying to fix it. And it results, you know, I don't think Elon wanted it
5:40
to start denying the Holocaust. They did delete those posts. But it turns out that when you try and make it extreme
5:46
but not too extreme, you know, it's like that old joking post
5:50
about turning the racism dial to just right. Yes, yes. And they're struggling to do..
5:56
Well, because obviously there's no such point on the dial. What's the bigger picture
6:02
They're not turning it off, are they, the whole thing? Did they turn it off for a while? No, they turned it off for a while
6:07
They've taken out the sentence about making claims that are politically incorrect
6:12
I think the thing here is... What it is... Sorry to interrupt, but we're short of time because something was bubbling away in the back of my brain
6:18
and I couldn't quite pin it down. that's because politically incorrect means bigoted and racist i mean we've spent years in this country
6:25
pretending otherwise saying we woke is everyone dressing up writing articles for the spectator
6:29
about how no actually these are just the ordinary concerns of ordinary people if you tell something
6:33
to come up with politically incorrect stuff it will go inevitably to racism holocaust in ireland
6:38
and weapons grade anti of course it will and ai doesn know how to play the games that humans play It looks at what words correlate together what runs together what are the patterns And it knows that when you look in these forums when you look at these places
6:53
there is at the core of it deep racism, deep sexism, deep homophobia
6:58
And it's got to reproduce that. And, of course, grinding anti-Semitism, you know, hugely anti-Semitic
7:04
But the guy who programs this, the guy behind this is the world's richest man
7:09
a multi-billion dollar defense contract and of all the satellites orbiting the earth he controls
7:15
more than half of them he controls more satellites than the rest of humanity put together now the guy
7:21
that has just coded mecha hitler is the person who has this huge control over space over the
7:28
defense industry over comms it's elon that should worry us it's the you know it's dr frankenstein
7:34
that's scary not the monster his wings clipped at all by his falling out with donald trump do you
7:39
think yes i think he's vulnerable and i think that's why he's lashing out i think that's why
7:44
he's worried about the fanboys abandoning him that's why he's sort of looking like this if trump
7:51
and musk really fight trump will win he's got the military behind him he's got the support behind
7:56
him he's got the machinery of the u.s and so elon is vulnerable and i think when people are powerful
8:02
and panicked they do things like this and you can see the unraveling in real time and actually
8:09
eleanor wants me to ask if you can remember that what you were going to say before i interrupted
8:13
you so that's a bit like the ai prompt it it was to point to elon being worrying more than that you
8:20
know grok doing this it's it's abhorrent i mean the things it's saying here aren't funny they are
8:26
truly, truly grotesque hate speech that is spurring on the worst corners of the internet
8:32
But that's a symptom. That's a sideshow. The problem is that this has $10 billion of investment
8:39
in it because Elon Musk is behind it. I mean, imagine if you are someone who's just put
8:44
a billion dollars in debt to this thing. It's insanity. And it's not over yet. Thank you for that, James. We may be catching up again at various
8:53
points in the future but it's always a pleasure to listen to you the political editor of the
8:58
new world and prolific author james ball many thanks indeed
#news
#News
#Politics
#Technology News