Aardvark Daily

Please visit the sponsor!

This could be risky

The hype surrounding AI seemingly never ends.

Almost every industry and government on the planet is now looking at how they can leverage the perceived benefits of AI systems to their advantage... and that has to be a worry.

The latest such announcement has come from the US Department of Homeland Security (DHS) who've set up a task force specifically to investigate the application of AI to everything from the role of protecting critical infrastructure to screening inbound freight for threats and to address issues related to slave labour.

Researchers in all fields are also heavily invested in using AI to come up with new insights and solutions to problems that have hitherto remained unsolved.

Now on the face of it this all sounds great. Why not use a supercomputer loaded with machine-learning algorithms and with the benefit of huge training databases to your advantage?

Well the problem is that AI isn't always right and can't always be trusted.

There are countless cases cited where AI has done some crazy stuff when challenged with a problem.

AI has been seen to, on occasion, "make shirt up" when it feels the need to cite references in respect to its assertions.

It is also sometimes insanely confident in its results, despite being clearly mistaken and totally wrong.

One again we need only look as far as one of the most long-standing and pervasive of all AI systems (YouTube's content oversight system) to see just how often AI gets things wrong and how it is that no amount of arguing with such a system can change its assertions.

This is probably not too much of an issue if you're simply posting cute cat videos to the internet and an AI erroneously claims they're somehow unacceptably violent or violate someone else's copyright. However, if you're going to be relying on AI to protect your nation against threats or to provide lifesaving treatment for the patients in a hospital then things get a whole lot more serious.

When we're dealing with situations where AI's errors can be corrected by a human then this becomes part of the "machine learning" system. When corrected enough times, the AI should learn that a particular response or set of responses is invalid and therefore alter its processing accordingly. However, when its delivering results in a realm where no human correction is offered or where we don't even know it's made a mistake, such errors can fester and accumulate to actually pervert the processing in a very bad way.

This latter situation has been occurring for some time with YouTube.

The problem here is that YouTube now takes the word of its AI over the word of those who are obviously being hit by AI-errors and it appears that very few corrections are actually being applied to the erroneous outputs. This seems to be convincing the AI that its false conclusions are actually valid ones and that compounds the problem even more.

Perhaps the thing I fear most about a rapid proliferation and reliance on AI systems is that those using such systems will not provide the essential corrective feedback at a certain point. They'll simply become complacent and rely on AI to get it right, to such an extent that they'll stop actually checking the results for validity.

Once this happens we could very, very rapidly enter a spiral where AI disrupts the very things that it is supposed to be enhancing and improving.

Once we transition from being users of AI to becoming dependent on AI and ultimately addicted to AI in order to deliver our desired standards of living -- we become incredibly vulnerable to such a destructive spiral.

Whilst I believe AI has an immense amount to offer and it has already proven to be a huge overall benefit to mankind, I also believe that we should be doing a lot of forward planning to manage and mitigate the risks that are potentially just around the corner.

Sadly, as is so often the case when *profits* are involved, such planning and mitigation are probably seen as "non-essential costs" at this point in time and will therefore be deferred... perhaps until it's too late.

What do readers think?

Are we sailing blindly down the river of AI towards an as yet unseen waterfall?

Carpe Diem folks!

Please visit the sponsor!

Have your say in the Aardvark Forums.

PERMALINK to this column

Rank This Aardvark Page

Change Font