The recent insights on generative AI and online safety by Bertie Vidgen (Former CEO & co-founder of Rewire; Former advisor to UK Govt and Researcher at Turing)  are so interesting we thought we’d include them ad verbatim: 

“The big AI labs seem to be doing a lot to make generative AI systems safe (or "helpful, honest and harmless"), from red teaming to audits, and incorporating automated safety layers. This has engendered a lot of confidence about our ability to handle the risks posed by these amazingly powerful models.

And that confidence is *somewhat* justified. Just try getting ChatGPT or Claude to respond violently or in a way that is overtly hateful. It's really not that easy. And the seriousness with which safety in AI is being considered is a sea change compared to what we saw in the early days of, say, social media.

But should we really be complacent just because the most well-known models aren't constantly spewing out toxic bile? I think that would be a serious mistake.

1. Most of the time you probably won't encounter unsafe gen AI responses, but that doesn't mean they don't exist. It all depends on who you are, which model you're using, and for what purpose. Just because you haven't seen anything harmful isn't evidence that it's safe. It's the black swan fallacy all over again...
2. Many AI safety issues are complex and nuanced, such as how gen AI models should respond to questions about abortion, gun violence, drugs, immigration and criminality. These are really questions of value preference, and are incredibly hard to resolve.
3. And there's a whole host of totally new issues to deal with, such as new types of privacy invasions; mass inauthentic behaviour; filter bubbles on steroids; hyper-realistic automated finance scams; and widespread technology addiction. In the future, malicious actors will use gen AI to launch attacks that we can't even conceive of.
4. Understandably, we all focus a lot on the big labs, who seem to be taking safety seriously (e.g. Google, OpenAI, Anthropic, Cohere). But there's loads of small start-ups, enterprise businesses, and in-house teams building and releasing gen AI models. Who knows if they'll be as careful?
5. And... finally... gen models are going to be so widely used, even a small amount of risk will scale to become a huge problem. We can't just dismiss it.

So what can we do to ensure generative AI is safe? Well there are no silver bullets but safety by design is a total slam dunk. Starting from responsible and well-informed design just has to be a good idea as it means embedding safety across the whole AI delivery process.

All of which is a very long way of saying that if you care about AI safety, you should read this blog! It's a great explainer for safety by design, outlining why it is needed and what it involves in practice.” 

Find out more about RWA's AI & data analytics recruitment solutions or browse our latest jobs.