Juha Saarinen: How can we keep AIs aligned with human values?

AIs are only as trustworthy as the humans who add the data for the machines' learning. Photo / 123RF

OPINION

Artificial intelligence will be humanity’s saviour, able to figure out how to end the climate crisis, come up with new miraculously effective medicines, and many more wondrous things that don’t exist now.

AI will also harm and hurt humanity, cause mass unemployment and become an existential threat to us all by generating devastating new weapons.

Unfortunately, both of the above statements could be true and possible, as AI systems grow bigger and better.

Who had that on their bingo card: a heady brew of neural networks, a technology that was in the “doesn’t work well” basket until researchers started feeding it giant amounts of internet data to be processed by powerful graphics cards?

That’s the power of pattern recognition for you. An AI is able to process images, text, sound and other data tirelessly and in minute detail, to spot patterns that humans might miss or not see at all. Like in medical research data.

What about the risks of AI, then? Computer scientists say that to not cause harm, an AI should be in alignment with the goals and values of its human creators and users, and adhere to them.

The machine does what it’s told, and we keep an eye on it to make sure it stays within safe boundaries and doesn’t go off the rails hallucinating, for example.

However, OpenAI’s co-founder and chief scientist Ilya Sutskever noted in an interview recently that today’s alignment methods won’t work once the machines become as intelligent as people, or even more so.

AI is developing fast, with Sutskever and other scientists now believing we’ll have an Artificial General Intelligence (AGI) soon. An AGI that we might not be brainy enough to fully understand and control.

The answer to that is superalignment, which means AIs being able to reason about and contribute to human values, so as to be more adaptable to people in order to serve them better. Would an AGI with cognitive faculties superior to people stay locked down in servitude to us, though?

Furthermore, alignment is a difficult concept to grasp.

OpenAI’s ChatGPT and Google’s Bard AI both say alignment is “challenging” to measure and trot out some general principles about being good machines, for the benefit of humanity.

Similarly, they will deny being misaligned with humans, giving the same reasons. “No rogue AI tendencies here!” as ChatGPT put it.

Anyone who has tried out an AI will have prompted it to hallucinate at some stage. ChatGPT’s claim of having no rogue tendencies is best seen as an example of it generating a response that’s not anchored in reality but which the AI thought the user would like to see.

Ensuring alignment in AIs can be problematic due to their design and the goal of creating human-like content. That is, the generated material should be realistic, but it mustn’t infringe on copyright or AI customers could face legal consequences.

Basically, an AI must know what it can’t show.

Tech entrepreneur Simon Willison’s experiments with ChatGPT-4, which can now drive the OpenAI Dall-E 3 image generator, found a “hidden prompt” which tells the AI not to create images in the style of artists whose works are from within the last 100 years.

Willison said this is because “they have obviously trained the model on all sorts of copyrighted images”.

Understandably enough, there are artists who are on the warpath against AI mimicry, and don’t accept it as the sincerest form of flattery.

Instead, they have turned to technology for protection.

The University of Chicago has launched The Glaze Project, which lets content creators “poison” their copyrighted work. Humans will see one thing, but AIs trained on Glazed works will get it wrong, and produce unexpected results.

Data poisoning can be done for entirely malicious purposes too. Across the Tasman, researchers at the Cyber Security Cooperative Research Centre have warned about fake and misinformation being added to machine learning data to subvert it.

The CSCRC noted that labelling data is often outsourced to countries with cheap labour, in “AI sweatshops”. Poor workers and corrupt officials could be exploited by malicious parties to manipulate data labelling to attack AIs.

It doesn’t take much tampering to cause damage either, with a group of researchers noting that just 0.01 per cent is required - and it’s cheap, just US$60 ($102).

Think about it: AIs that speak, write, see, hear and create images, based on human-generated data that may have been corrupted without us having any idea. We really do live in the future.