AI Is Lying to Us About How Powerful It Is

There is something very click-baity about articles like this. There always seems to be the implication that these examples are completely independent emergent behaviour, like the AI has suddenly acquired self awareness and a need for self preservation from nowhere. Usually when you did deeper you find the entire process was guided, or the AI was given an objective, and carte blanche on how to achieve it, along with privileged data and admin access that it did not need to achieve it's goals. The machines, believe it or not, still do exactly what we tell them to do.

> Unlike humans, politicians have no innate sense of conscience or morality that would keep them from lying, cheating, stealing, and scheming to achieve their goals. You can train a politician to speak politely in public, but we don’t yet know how to train a politician to actually be kind. As soon as you stop watching, or as soon as the politician gets smart enough to hide its behavior from you, you should expect the politician to ruthlessly pursue its own goals, which may or may not include being kind.

The examples cited here are from very artificial studies. Currently LLMs run in inherently sandboxes environments with very little control. They’re also trivial to monitor as they have precisely one output mechanism: the token stream.

It’s unsurprising to me that an AI trained on human actions such as securing a server, will turn off monitoring when given the chance, specifically because that would likely be human best practice.

These models did not just start copying themselves around, or changing new models, they were given explicit opportunity to do so and prompts that could lead them down that path. This would be like a leading question, or even reverse psychology.

> These results are disturbing, but what’s even more disturbing is the AI developers’ apparent comfort with their products' increasing disloyalty.

We've been dealing with this for literally centuries at this point.

You want to know the solution? Stop putting unaccountable black-boxes in charge of anything you consider important. The execs will never listen, but you can't blame engineers for apathy towards this whole affair.

Articles like these, sparse on details, presenting just the parts that fit their narrative are misleading. Yes, you're the center for ai policy, yes you need some hype and fear to stay relevant and keep your job, but please provide actual details of what you're describing instead of throwing out nonsense stats, or you lose all credibility.

Pretty much no one in this "center for ai policy" has any experience in AI or is technical: https://www.centeraipolicy.org/about

It is not clear who is financing such efforts and whether these do really work (they are based in DC)

> CAIP is grateful for the generous support of our donors. We’re supported primarily by mid-level and major individual donors who share our mission to improve AI governance. Several of these donors built up their wealth during the dot-com boom or by working at hedge funds. To protect the privacy of these individuals, we do not publish their names. We also received seed funding through an organization sponsored by Jaan Tallinn, a former founding engineer at Skype. To maintain our independence, we do not accept funding from companies who are designing or building AI software or hardware. We are nonpartisan and focused squarely on the public interest.