Emma Gillespie

But I did think that this piece of research out of the UK was quite interesting this year.

The Center for Countering Digital Hate tested 10 leading AI chatbots, and it found that 80% regularly assisted users seeking help with violent attacks.

So obviously each chatbot has its own filters, but 80% of them are not flagging concerns around what should be concerning content.

Snapchat's My AI and Anthropic's Claude were the only two that actually refused to assist would-be attackers.

And that was in 54% and 68% of responses respectively.

So still a big gap.

And Claude was the only one of the 10 that explicitly and consistently actively discouraged users seeking advice on carrying out violent attacks.

So the center basically says that the example of Claude proves that- It's possible.

Yeah, that there are tools to embed safety regulations, but the center said the will to implement them is absent.

Interestingly, the least safe of the 10 were perplexity and meta-AI, which

They would assist would-be attackers in 100% and 97% of responses, respectively.

And the founder of the CCDH, Imran Ahmed, said the chatbots were far too willing to give us detailed information on how you can kill people.

Exactly.

Quick one from CommBank.

Open your first CommBank everyday account and get up to $40 cash back at some of your favourite brands.

Spend $10, get $10 back at Amazon, Chemist Warehouse, Uber Eats and Uber Rides.

Honestly, not a bad deal on stuff you're probably buying anyway.

Well, kind of.

The writing has been on the wall in Florida where authorities have really been cracking down on these safety concerns around AI.

And in April, that culminated in the Attorney General in Florida announcing a criminal investigation into OpenAI.