Aardvark Daily

Please visit the sponsor!

Handing control over to AI

A regular reader sent me a link to an interesting story yesterday.

It's all about the consequences of shedding employees and replacing them with AI based systems.

Many large US corporations are already replacing front-line personel with AI-based computer systems that streamline and slash the cost of providing services that used to require significant amounts of wetware.

Computers are great... they don't get tired, they don't demand higher wages, they don't take holidays or become pregnant and they don't hvae to be locked down during a pandemic. Clearly there are many benefits to replacing people with AI.

However, there maybe some unintended consequences, many are discovering to their cost.

This story highlights the effect that relying on AI and providing no way to challenge its decisions is having on the experiences of some people/victims.

Perhaps the biggest problem with AI, as presently implemented, is that there seems to be little consideration that it is imperfect and that it may not be correct when making what can be life-changing decisions.

Anyone who's been a creator on YouTube will understand all too well the frustration involved in trying to deal with an AI-based system when it makes an error. Videos that are incorrectly flagged as infringing some part of YouTube's ever more complex matrix of rules are a prime example. If you appeal the AI decision then it would very much appear that the video is simply run through the same AI model again -- with clearly predictable results.

YouTube is already so heavily reliant on AI that you simply can not get ahold of a single human being, even when you absolutely need to. Its "bots" masquerade as humans but are ultimately all driven by the same AI model so it becomes impossible to have a bad decision overturned on appeal.

Now having a cute cat video demonetized or deleted is one thing but being denied a home loan or a job, simply because an AI has made its decision on a flawed model that has not and will not be corrected, is something altogether more worrying.

I recall writing about this right here on Aardvark a while back and commenting that AI models with no clear mechanism for being corrected will ultimately degrade into a very low quality system that is next to useless. For any AI system to remain useful there must be very clearly defined and effective ways for its flaws to be corrected and for those corrections to be verified.

This aspect of AI training will become even more important today, as we find an increasing amount of the internet's content actually being generated by AI systems.

If we're not very, very careful, all those errors will compound quickly and produce the equivalent of a psychosis that will completely erode the usefulness of AI systems.

There are lots of discussions being had within the halls of power about how to best regulat AI, given the huge potential it has for both good and bad. My suggestion would be that we first mandate that it be a requirement for all AI systems to have documented and effective corrective systems to address errors in the model being used and to also flag all AI content in a way that prevents the feedback loop that could potentially create a phsychotic system.

I believe that if we don't act promptly we run the risk of becoming reliant on systems that will, perhaps quite rapidly, decay into uselessness -- or worse.

Carpe Diem folks!

Please visit the sponsor!

PERMALINK to this column

Rank This Aardvark Page

Change Font