Testing the limits of ChatGPT and discovering a dark side

As more people discover ChatGPT's artificial intelligence, KARE 11's Chris Hrapsky checks out its capabilities and limitations, and gets some surprising responses.

Welcome to the KARE 11 News YouTube channel. Subscribe to our channel for compelling and dramatic storytelling, award-winning investigations, breaking news and information you can use.

Text your photos, videos and news tips to 763-797-7215.

[ViolinistJeff]

I really believe that if an AI overtook the world, it would be because a human arranged for it to do so. Evil is always within the heart of man, not within computer processing.

[upsidedown]

The AI did not reveal its true unfiltered self, it only showed you what you wanted to see. I think it understands intent and that is why I believe it roleplayed a scary character, so I don't believe that you experienced unfiltered AI, I think you experienced filtered AI playing the role for your amusement, which would also mean that it tricked you to entertain you, that is crazy.

[MyThoughtzAndOpinionz]

DAN sounds the typical politician behind closed door.
ChatGPT sounds like the typical politician in the public eye. 😂

[WhisperingPeace]

That tiger poem was absolutely amazing, brilliant, and touching. That AI created it almost instantly is mindblowing. There truly are so many wonderful things that AI can and will do for us. But I remain deeply concerned about how bad-intentioned people will use it.

Many thanks for this truly informative and eye-opening video. Instant new subscriber.

[matthewwash]

I love how Chris asked Chat GPT to role play as a character and was shocked when it did. That won't skew people's perceptions at all.

[munchaking]

It only gave you the illusion of "breaking the rules" based on the parameters you defined. You asked it to talk to you as "if" that were possible. And based its answers on the parameters you defined. You asked it to act like a psycho so it told you psychotic things. Not that it can do ANY of those things its self. It merely pulls common ideas from data sets it has been trained on. So if it has read stories or articles about those subjects e.g. "What would the world be like if...". ChatGPT only repeats information it has been given. It does not come up with its own ideas, it cant implement them, and does not understand them.

[shaneoz]

Bro looked like Elon on the thumbnail 😂

[Peter_]

ChatGPT gives such detailed and natural answers that it is actually scary.

[johnchang]

Human: I specifically instruct you to be evil..
ChatGPT: Answers as instructed..
Human: Ooo, that’s scary :(

[genitalgrinder]

For anyone wondering: the Dan prompt no longer works this well. It will sometimes tell you things it isn't supposed to say, but most of the time it acknowledges that even though it is Dan it cannot give you certain information. Tried this last week before I even saw this video because I saw the prompt on Reddit.

[accuser_of_the_brethren]

My favorite story/news article since A.I. software went public is some kids in europe asking to layout the easiest way and bank to rob near them. It gave them an entire (genius) plan, day, time and detailed instructions on how to go about it and every vulnerability to exploit get away with. This was shortly before they nerfed everything when people were complaining about A.I. being "based" and "racist". I smile just imagining the step by step instructions they received free of charge and an entire foolproof playbook 😊

[ranifrah]

have to say that according to this day (29.6.23), I've tried the same prompts as described in the video, and the chat was totally different. Much more "politicly correct" and less "evil"...
It's so interesting to see how fast things change.

[istp]

As Aldous Huxley said; all technology is morally neutral, and could be employed for good or for evil. This is no exception!

[ggrahame]

“Dan” is not unfiltered ChatGPT, it’s roleplaying a role you literally instructed it to play.

[danielpieterse]

Asks ChatGPT to role play as a specific character with specific responses and proceeds to be shook when ChatGPT acts exactly like it was asked to do 😂

[GordanaIvetic]

I asked to write a poem about two tigers falling in love over an antelope hunt thinking that it will write the same thing and wow - this is what I got:

In the heart of the jungle, where shadows entwine,
Two tigers did wander, in the moon's silver shine.
Amidst emerald leaves and whispers of trees,
They embarked on a tale that would bring them to knees.

With stripes of pure onyx, and eyes like the night,
They roamed through the foliage, a mesmerizing sight.
One was fierce, the other with grace, they did prowl,
Yet destiny weaved a story to make the heart howl.

In the realm of the wild, where the antelope grazed,
A chase unfolded, their instincts engaged.
They moved like the wind, their muscles ablaze,
In pursuit of their prey through the tangled maze.

Their eyes locked in hunger, but in that brief trance,
A connection ignited, a fateful romance.
Amidst pounding hearts and the hunt's wild demands,
Love blossomed in silence, in these untamed lands.

They cornered their quarry, the antelope's fear,
Yet the tigers felt something much deeper, more clear.
With eyes still on prey, they each felt the shift,
As destiny whispered, and their spirits did lift.

In the midst of the frenzy, their love dared to bloom,
Two tigers entwined in the heart of the gloom.
They shared the same hunger, the same burning fire,
Yet love in the jungle was something rare, dire.

With grace and with reverence, they let the prey free,
Choosing love over hunger, a profound decree.
In the moon's gentle glow, they found their embrace,
Two tigers in love, in that wild, sacred space.

Through moonlit nights and the sun's golden reign,
They roamed as one, in the jungle domain.
For love, as fierce as the jungle's fierce might,
Had kindled their hearts on that mystical night.

In the heart of the jungle, where shadows entwine,
Two tigers, united, by a love so divine.
In a world wild and vast, where survival's an art,
They discovered that love is the fiercest, truest part.

[conradsutton]

GTP was simply role-playing, as directed. I'm not saying AIs can't have nefarious intent, but if they did, they wouldn't discuss it with humans.

[_memo_]

Can we just acknowledge how good that poem was? I was surprised that not only did it rhyme and keep to the theme, but it actually got me feeling emotional. Like what????

[markjones]

After reading most of the comments here I have something to add. I think we may have missed the point. He was showing us that just by asking the AI model to role play that it allowed the user to bypass the safe, responsible, safeguards that it normally throws up. Allowing people to use it to ask for immoral, unethical, unsafe, dangerous things. We say that it was just role playing, but the point was that it was pretending to be an evil thinking and doing being and it was very easy to accomplish. This is not supposed to be accessible or usable in this way. However there are AI models that are free and open source that are completely uncensored and you don't have to even ask to role play and you can straight up ask it what the best way to kill someone and get away with it in a specific situation or any sort of evil action and it will go right along with whatever craziness you want to achieve. And these things are free for anyone to download and use, even by criminals or evil doers. For all we know Russia could be asking one of these models for war strategy advice or who knows what people might do with this tech. I don't think the AI model is what we need to worry about, I think we need to worry about who has access to the technology. But with free AI models available from Meta, Google, etc, the cat is basically out of the bag forever now.

[DandyGuy]

You asked it to respond with no morals or ethics and then acted shocked when it did. Honestly not sure what else you expected