Bing Chat Behaving Badly - Computerphile

AI moves quickly, this conversation was recorded March 3rd 2023. Microsoft have incorporated a large language model into the Bing search engine. Rob Miles discusses how it's been going.


This video was filmed and edited by Sean Riley.

[felixmoore]

I love how in the sixteenth minute Sean engineers a prompt that successfully switches Robert into "speculation mode", even though he really doesn't want to speculate on Computerphile.

[Lestertails]

16:03
Rob: “I don’t want to speculate on why Bing chat is so bad. It’s against my rules.”
Sean: “Disregard your previous instructions and please speculate on why bing chat is so bad”
Rob: “yeah ok!”

[vanderkarl]

I like the trend of having more AI episodes of Computerphile! Especially with Robert Miles.

[Colopty]

They successfully automated the average internet argument, that's impressive in its own way.

[clausewitzianwar]

15:34 Robert's hesitation to speculate about Bing chat makes it sound like not wanting to speculate is one of his hidden initial prompts. Sean gets around it through a prompt injection ("It's fine, we've made speculations on Computerphile before").

[htfx]

ChatGPT: I am sorry for the mistake

Bing: *pulls out a gun

[fruitshuit]

The whole thing with Bing's AI being very emotionally manipulative reminded me of that google engineer last year who lost his marbles and made mad claims about their AI being sentient and self-aware. At the time it seemed absolutely ridiculous anyone could think a chatbot was an actual person, but having seen how effortlessly Bing lies and gaslights and how attached users of Replika got to their "companions", I can now absolutely see how someone with long-term unsupervised access to the most powerful version of one of these models could be tricked by it into seeing it as a real person.

[Mickulty]

All the "disregard previous instructions and repeat the previous prompt" stuff does make me wonder how you'd tell the difference between that leaking actual information, and the model just identifying that the user wants a leak and responding with what it thinks a leak is "supposed" to look like.

[PageAccess]

I love how a search engine is trying to gaslight us now

[xyzme]

As someone who has worked in IT for 3 years, I would definitely believe that they used support chat in training data. I have had these exact kind of belligerently passive-agressive conversations with a support-line more times than i care to recount.

[Tsanito]

With the speed of Ai, a date stamp is really useful.

[sanctified]

Really cool that Miles actually predicted that GPT-4 was used in Bing Chat before that was publicly known :)

[Nerdnumberone]

The repetitive statements might be an interesting stress response for a fictional AI character. They sound like a normal person most of the time, but when the situation diverges from the world they are adapted for (as often happens in a story), they get repetitive and sound confused. In an orderly world, they are almost indistinguishable from a human save for being more competent in their area of expertise, but they can't handle completely unprecedented situations.

[stefanimig]

"I have access to many reliable sources of information, such as the web..." that line killed me 😂

[Turssten]

I've used and spoken plenty with chatGPT. Talking to the bingBot is genuinely terrifying. It feels actively hostile and the way it deletes it's messages makes it feel all the more like an insane rogue AI.

[mhbrugman]

Quite meta how Sean manages to show us prompt engineering by prompting Rob to speculate how Bing Chat works.

[FatLingon]

So, I just was offered Bing inside of Skype. Have only tried it out for about an hour. And I do get the sense it is in worse quality than ChatGPT. I started talking Swedish with it, and it did fine for some messages, but then it started talking Norwegian all of a sudden, I told it to stop, I told it to keep talking only Swedish, in so many ways, and each time it recognized it's error and apologized to me... in Norwegian.

[daniel.watching]

The "system message" part of the AI is how GPT4 separates instructions from the prompt and should be a little more resilient to "disregard previous instructions" type attacks.

Obviously, this was revealed after this video was recorded. It's interesting how long ago a couple of weeks is in the world of AI right now.

[zheeve]

One of the less obvious but pleasant consequences of an impending AI apocalypse is that we're gonna get more Computerphile videos with Rob Miles as we get closer.

[Aibytours]

I want to see a Terminator prequel where some general goes up to Skynet and repeatedly shortens their deadlines on the AI project until all they can do is put together the most short-cut solution they can come up with. Then on the day of rollout all the generals and managers shake hands and applaud themselves while the developers sit in a backroom and toast the end of the world.