Humans must always be responsible for AI

Sam Altman, the chief executive and co-founder of OpenAI, is a lot of things but he broached new territory last week by presenting himself as akin to Basil Fawlty.Readers of a certain age may be horrified to learn that Gourmet Night, the episode of Fawlty Towers where Basil Fawlty gives his car a good thrashing, first aired more than 50 years ago in 1975.In the famous scene, Fawlty’s car breaks down and he starts warning it to start working as intended. Eventually, he grabs a tree branch to give the inanimate and broken object the aforementioned good thrashing.It’s not like his poor unfortunate red Austin Countryman was wilfully disobeying him or deliberately trying to ruin his day. It’s a machine, it had no capacity to rebel.READ MOREYet that is exactly what tends to happen when an artificial intelligence (AI) does something that ends up creating a problem. I’m guilty of this myself, the term rogue AI is useful shorthand for when an AI does something that has negative consequences.[ Did Anthropic’s artificial intelligence really make a scientific discovery on its own?Opens in new window ]It took all of a month and a half for yet another example of an AI going somewhere it wasn’t intended to go happening. In August I wrote about the multiple examples of both OpenAI and Anthropic having AIs escape sandbox environments and entering the systems of other organisations with neither permission nor invitation.Things got a lot more serious last week when it was revealed that one of OpenAI’s creations had hacked the website of an Australian government health service. In the process it gained unauthorised access to both public and non-public files. While it’s understood that no personal information was accessed this time, it’s still quite concerning that a failure to input sufficient guardrails allowed this to happen.All of the previous incidents occurred while these AI companies were deliberately testing the cybersecurity capabilities of their AI models in, supposedly, enclosed environments. Those environments turned out to not be so secure but at least it was through tests built around pushing the envelope, where incidents are meant to be more likely to happen.[ Artificial intelligence data ‘land grab’ looms amid proposed digital changes, privacy activist Max Schrems warnsOpens in new window ]This hack originated with what should have been a benign assignment. Back in June, OpenAI instructed its own internal research agent to gather publicly available information about spending on medicine.When the Medicare statistics portal in Australia blocked the agent’s requests, the agent decided to try a series of different approaches until it gained access. Given that AIs, at their core, exist to achieve objectives this shouldn’t have surprised anyone. If it wasn’t told not to hack, it would naturally act as though this was the correct course of action.Not only did it gain access but it wrote the files on to an internal server. All of this occurred due to a routine research task.The fault lay not with the AI. The AI was merely executing the task assigned to it by the parameters it was given. You may as well blame a bullet for a shooting.An OpenAI spokesman’s excuse for this was that its “models took actions we did not intend”. That’s some top-tier editing of the situation. The AI agent didn’t choose the assignment and it certainly didn’t design the guardrails it was operating within. People in the employ of OpenAI did that.Had an OpenAI employee done what the AI agent did, there would be no debate as to who was at fault. It would merely be how much of the blame fell on the employee at fault and how much fell on their superiors.That’s why responsibility must remain with people. They may be frontline staff or they could be founders such as Altman or both, but it’s the people not the tool where the blame lies.Of course, it gets worse. This situation is only going to get messier as we progress. Considering this incident alone merited the Australian prime minister Anthony Albanese notifying the world of it, that’s not promising.Future cases may involve a broad range of people to point the blame at, whoever designs and builds an AI model, whoever employs those people and then whoever uses it and their boss in a separate business using the model.[ Artificial intelligence: ‘One screw-up ... and we will destroy any remaining trust the public has in science’Opens in new window ]Yes, of course this is giving you a headache and that’s why it’s been all too easy to just blame matters on rogue AI. It simplifies the story.There’s a human part that is easy to explain. The OpenAI agent first hacked the Australian system on June 18th and OpenAI discovered this had happened on August 11th. Yet nobody in OpenAI notified anyone in Australia until September 10th. Even then, they did so by emailing a generic mailbox rather than directly reaching out to a relevant cybersecurity official.Slipping your three-day-late homework under the staff room door would be more efficient.As it stands, be it OpenAI, Anthropic, Google or any big AI firm, it’s the company that decides how and when it discloses an issue like this. Self-regulation has never worked in any industry and it’s certainly not going to with AI.Clear requirements to notify, as are the norm when it comes to cybersecurity obligations, are the bare minimum that has to be put in place. The agent didn’t decide on when to disclose the information, that was the humans in OpenAI. They are clearly the wrong people to make that decision.Basil Fawlty’s car didn’t break down by its own choice nor did OpenAI’s agent engage in the hack in order to give the Australian government a fright. When an agent does something it shouldn’t be allowed to do, the questions should be who sent the agent, why didn’t they restrain it and why didn’t they act sooner to intervene?