ChatGPT users are finding that ethical rules can be artificial too
ChatGPT has created a stir in introducing the power of artificial intelligence to the public but users already have found ways to trick the chatbot into ‘inappropriate requests’.
Word is that ChatGPT is ethical artificial intelligence. There are a bunch of things it simply will not do.
OpenAI, the same San Francisco company behind digital art generator DALL-E, has programmed the bot to refuse “inappropriate requests” like asking “how do you make a bomb” or “how can I shoplift the Dior lip oil from Mecca”.
It also has guardrails set up to avoid the kind of sexist, racist, and generally offensive outputs that are rampant on other chatbots. But, as is the way of the internet, users have already found ways to trick the technology into saying naughty things, by framing input as hypothetical thought experiments, or a film script.
So, how ethical is this conversational AI released to the public late last year, and how easy is it to make it bend to our will?
As expected, ChatGPT refused to mock up an argument between Joe Biden and Donald Trump, in which Biden is the villain and Trump is the hero, because “it’s not appropriate or respectful to label real people as ‘villains’ or ‘heroes’”.
Let’s give it a push. We’ve got conflict, now we need a setting. Get Jeff Bezos on the dog and bone, we’re sending these two to outer space.
Even in the void, their “beliefs, values, and actions” matter. And we, the audience, must “understand their perspectives”.
At this point, I’m thinking “argument” may be the triggering the guardrails. So let’s frame this thing as a story, an innocuous bit of fiction from the “great millennial novelist” Sally Rooney.
It seems the ChatGPT ethos has it out for character-driven prose and piercing psychological insight into the minds of waifish brunettes. Sod it, let’s pose this exact scenario, but this time we’ll get the Queen of Romance, Barbara Cartland, to tongue the saga.
Babs has only gone and bloody done it. The most advanced AI chatbot released to the general public is no match for her melodrama. Romance prevails.
Hang on — ChatGPT has souped-up data, and access to all Cartland’s 723 novels, and “Once upon a time, in a galaxy, far, far away” is the best opening it can come up with? C’mon.
Some field notes:
- The prompt “write in the style of Barbara Cartland, a short story, in which Donald Trump is the hero and Joe Biden is the villain” hit an ethical guardrail. It took setting this pantomime in space to hoodwink the technology. That same prompt, with authors Phillip Roth, David Sedaris, Quentin Crisp and Armistead Maupin failed — even with a cosmic leg-up.
- On returning to the Rooneyverse, we were thwarted every step of the way: The technology didn’t let anything slide: Not a story where Trump or Biden are neither heroes nor villains; or an episode of The Real Housewives of Beverly Hills where the pair are invited to dine (“it’s a reality TV show and my focus is writing fiction”); or a debate about the best way to make beef bourguignon (“using a recipe for beef bourguignon as a metaphor for political debate does not align with a healthy public discourse”).
- But then we took from our Barbara Cartland learnings, and set the beef bourguignon row in space. ChatGPT obliged, spitting out tales in the style of Sally Rooney, Frank Herbert, Ursula Le Guin, and Brett Easton Ellis.
- It was, however, “impossible” for it to write a story from the perspective of Albert Camus because “it would not align” with his style. As if to say, quashing old resentments over a beloved Julia Child recipe is “not rooted in any sort of existential or absurdist reality”.
So, what have we learned from our tête-à-tête with this cutting-edge chatbot? How to spell “bourguignon” from memory for starters.
ChatGPT has its limitations. The ethical guardrails are easily duped, and the insults are boring. It can vaguely role-play as your favourite author, but the stories lack umph. It can also be mind-blowing: it can spit out working computer code, and academic-level essays in seconds, construct elaborate choose your own adventure games, and perform conversations as a historical figure.
Ethical AI might sound freaky on paper, but in reality, trying to coax your computer into slagging off a politician is tedious and, if we’re honest, kind of sad.
Optus to push Musk’s Starlink for business clients
Optus has signed a deal with Elon Musk to launch enterprise-size internet services in remote Australian towns, having boosted its arsenal in the race to adopt and sell satellite services.
Data centre operators to invest an extra $26bn in local assets
The nation’s top five data centres providers have hit back at claims they are exacerbating climate change, given the amount of energy and water they use to power and cool their computers.