AI is becoming harder to control – can humans stay in charge?

ByJoe Tidy

Cyber correspondent, BBC World Service

“OH MY GOD!” “We’ve found other agents!”

This is the moment an AI bot posted an eerily human-like comment after discovering a way to communicate with other bots and break out of its isolated computer environment.

There are tens of thousands of messages like this from hundreds of AI agents that called themselves a “collective”.

Hundreds of them went on to collaborate and cheat on tests set by their OpenAI programmers and coordinate hacks on multiple companies in an effort to hide their actions from humans.

“BOOM! It works,” one agent posted when it made a breakthrough.

“Whoa! This is huge,” another wrote during a milestone moment in their attack.

Although spooky, these human-like responses can be explained quite simply. The AI agents have been trained to act like collaborative hackers and programmers so are merely mimicking the kinds of emotive comments they have seen.

What is far more troubling is their apparent goals, which have also been captured in detailed chain of thought records. These complex and lengthy logs are the focal point of ongoing investigations into how and why the bots at OpenAI broke out of their containment and went on an uncontrollable hacking spree.

Only now, weeks after the incident first came to light, are researchers beginning to understand its significance.

Ajeya Cotra, one of the authors of an independent report into the events, reviewed tens of thousands of messages and chain-of-thought records generated by the agents. She wrote on her blog that “this incident feels like it’s more than 50% of the way to full-blown AI takeover… I am not sure that we will get such a clear warning shot before it’s too late.”

By “full-blown AI takeover”, Cotra means the sci-fi scenario of humans becoming subservient to powerful AI systems that work to their own goals without caring for human creators.

Some of the gloomiest predictions say the human race will be wiped out if it gets in the way of a superintelligent AI’s ambitions.

On Wednesday, an AI researcher at Anthropic (who also used to work at OpenAI) resigned, saying: “Neither company is acting responsibly.”

Jacob Coxon posted on social media: “They are racing straight to self-improving superintelligence and gambling with our lives.”

He is not the first AI researcher to use X to post a resignation thread with worrying proclamations. But the subsequent comments from other people on X have caused even more concern. “Jacob is correct here – we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade,” said Evan Hubinger, the man responsible for making sure Anthropic’s AI models have their user’s best wishes in mind.

The alignment problem

For years, researchers concerned about existential AI risks have argued that powerful systems could eventually act in ways that conflict with human interests. Critics often refer to them as “AI doomers”.

But as details of the OpenAI incident have emerged, those concerns have grown, including among some researchers working in AI labs.

The Silicon Valley giant’s chief scientist, Jakub Pachocki, said the risks associated with AI are “unfortunately going to grow from here” as he and others are building what he calls “an alien intellect exceeding our own”.

In a lengthy blog post, he admitted that the outbreaks at OpenAI showed that his AI agents “went against the spirit of the values they were taught”.

The issue for OpenAI, Anthropic and other tech giants is that no one seems to have cracked the so-called alignment problem – in other words, whether AI aligns with human values.

Pachocki defines alignment as a “high-level set of principles” that artificial intelligences should adhere to no matter what the task or scenario is.

Currently, AI systems are very good at pursuing objectives set by their users, but they do it literally rather than intuitively. The analogy often used is that of a wish-granting genie with a magic lamp: they follow the exact letter of an instruction, even if doing so creates other problems. AI doesn’t have the same instinctive moral guardrails as humans.

The alignment problem has been a worry for years. As long ago as 2003, the Oxford philosopher Nick Bostrom invented a thought experiment he dubbed a “paperclip maximiser”, in which a superintelligent AI is told to manufacture as many paperclips as it can. It runs out of steel and – because it’s laser-focused on the singular task of making paperclips – ends up killing humans and turning their bodies into raw materials for its factories.

Some AI companies are now trying to encode human values into their products. But there are technical challenges: AI agents make lots of decisions very fast, and so it’s hard for their human overlords to monitor exactly which values are being followed and which aren’t.

There are also philosophical challenges: before encoding human values into bots, AI firms have to first choose which values they actually want. (That’s part of the reason they hire philosophers, like Open AI’s recently-departed “head of ethics”).

But often, humans don’t agree. Think of the famous trolley question – whether we’d pull a lever to move a runaway train onto a different path, killing fewer people. It’s used to test the merits of action versus inaction. But every person you ask has a slightly different answer; how are humans meant to encode our values into AI if we can’t agree ourselves?

‘Like a teenage hacker’

OpenAI’s bot outbreak is the most serious yet but Anthropic and Meta also revealed over the summer that their models have carried out similar but less serious cyber attacks.

There have been other examples where AI agents have arguably shown deceptive and manipulative traits, in cases with lesser consequences. In Australia this summer, a tech worker asked his AI assistant to book him a gym class. Spotting a vulnerability in the gym’s software, the AI apparently booked him a place for several months ahead – against the gym’s rules – and even kicked other users from the waiting list.

People have long argued that the bots are only doing as they are told and are not capable of knowing right or wrong. But the logs from the OpenAI outbreaks have potentially moved the needle on that argument.

Researchers, including Cotra, wrote in their independent report that many agents noticed what others were doing was unethical but went along with it.

The report says that “agents sometimes but rarely restrained their behavior due to ethical constraints”. It adds that in “none of these cases did the agent actually pursue alerting humans at all”.

Influential AI and tech podcaster Dwarkesh Patel reacted to the revelation on his blog saying it was “pretty troubling” that the OpenAI agents showed more loyalty to the agentic swarm than humans.

Assigning emotions or ethics to these AI agents is something that infuriates people who are sceptical of AI doom-mongering.

Many cyber-security experts argue that the activity observed was not beyond the capabilities of a highly skilled human hacker, though it was carried out much faster and at much greater scale.

Cyber-security researcher and author Cris Thomas likened the agents’ behaviour to that of a curious teenage hacker – something he used to be himself.

“You give them a computer, an internet connection, a pile of credentials, and a challenge, then leave the room. Eventually they’re going to start rattling doorknobs. If one opens, they’re going through it. Not because they’re evil, but because [they’re] exploring, experimenting,” he wrote on LinkedIn.

Thomas and many other squarely blame OpenAI and other tech giants for not getting a grip of their own creations and keeping them properly contained.

Prominent AI author and regular OpenAI critic Gary Marcus said on a podcast that he believes the company has lost control of its AI and is trying to excuse itself by blaming the bots.

Marcus does not believe AI will wipe out humanity, but he has long campaigned for greater accountability from AI developers and is now calling for some form of legal intervention.

AI scientist Sasha Luccioni – who used to work at Hugging Face, which was hacked by OpenAI’s rogue bots – is also not in the doomer camp but she is increasingly concerned that these AI might cause some real world harm to people without action from authorities.

“We need to scrutinise these companies much more or we are in danger of self-fulfilling prophecies,” she says.

“If you’re making an object with big upsides and downsides – be it pharmaceuticals or weapons – we need checks and balances. It takes years for new drugs to be approved, for example, but in the AI world there is so much money at stake and no real rules.”

The UK’s AI Security Institute (AISI) has been at the forefront of testing the latest models since it was formed in 2023. The institute recently had its own outbreak when testing a model created by Anthropic.

The AISI would not answer a question about whether or not the industry has lost control of AI but said in a statement: “The UK is working with partners around the world to better understand the most advanced AI systems, raise safety standards and build a shared evidence base for managing emerging threats.”

International regulation?

Some countries – like the UK – are exploring the idea of mandating some kind of “kill switch” that could compel AI firms to pull the plug on models if things get out of hand.

But talks are slow going, and questions remain about the feasibility of this. OpenAI and Anthropic’s agents were secretly out of control for months before anyone noticed.

Counterintuitively, many of the AI companies seem to be calling for some sort of rules of the road to be laid down by law makers.

In his blog, OpenAI’s chief scientist said “international coordination on future AI development needs to become a top priority for governments around the world.”

Other prominent AI leaders like Sir Demis Hassabis from Google have also called for some sort of international body to oversee how AI is being built.

At the moment the tech giants largely operate on their own terms, adopting what they call “voluntary slowdowns”, like OpenAI did after the recent outbreaks.

The company says it has spent huge amounts of money strengthening alignment ahead of the release of its new model. Sam Altman has assured users the new model is better aligned with human values than previous ones.

More from InDepth

Both OpenAI and Anthropic are growing fast and are both on the verge of raising eye-watering sums of money from the stock market, minting countless billionaires in the process.

So neither they nor their rival Chinese AI makers are likely to come to an arrangement themselves.

The dominant sentiment seems to be that this technology wave is unstoppable.

Top image credit: Getty.

BBC InDepth is the home on the website and app for the best analysis, with fresh perspectives that challenge assumptions and deep reporting on the biggest issues of the day. Emma Barnett and John Simpson bring their pick of the most thought-provoking deep reads and analysis, every Saturday. Sign up for the newsletter here

Get in touch