Is AI Quietly Sabotaging Itself—And The Internet?

Forbes ^ | Sep 05, 2024 | Tor Constantino

[Openurmind]

Interest in artificial intelligence continues to surge, as Google searches over the past 12 months are at 92% of their all-time peak, but recent research suggests AI’s success could be its downfall. Amid the growth of AI content online, a group of researchers at Cambridge and Oxford universities set out to see what happens when generative AI tools query content produced by AI. What they found was alarming. University of Oxford’s Dr. Ilia Shumailov and the team of researchers discovered that when generative AI software relies solely on content produced by genAI, the responses begin to degrade, according to the study published in Nature last month. After the first two prompts, the answers steadily miss the mark, followed by a significant quality downgrade by the fifth attempt and a complete devolution to nonsensical pablum by the ninth consecutive query. The researchers dubbed this cyclical overdose on generative AI content model collapse—a steady decline in the learned responses of the AI that continually pollutes the training sets of repeating cycles until the output is a worthless distortion of reality.

---

TOPICS: Business/Economy; Computers/Internet; Education
KEYWORDS: ai; cambridge; forbes; iliashumailov; oxford; torconstantino

[The Duke]

Who would have thought that "a copy of a copy of a copy" would not be as good as the original? Surprise, surprise!

AI is a tool, and it's really good for some important things, but not good for everything under all circumstances.

[cgbg]

I would also recommend the following approach when AI states something.

“You have made a claim.”

“Now give me five strong arguments that conflict with your claim.”

[gnarledmaw]

The weighting is a real problem. Back in the 90s already there was a push for IT to develop/implement knowledge bases for various tasks. The ones that were seemingly the most powerful at first always included weighting of all the input.

Those were also the first ones to degenerate into nonsense most quickly. Lazy techs and target end users along with changes in technology left everything incomplete or hopelessly out of day making nearly all of the answers false.

The result of using those types of packages was that the larger the organization and the more data that went in the more time and personnel had to be devoted to debugging the knowledge base for it to have any use whatsoever.

When the endusers regardless of role discovered that it was unreliable they refused too properly participate by using it correctly leaving management to try to police their group and needless to say most of them didnt.

This left IT with an expensive gobbledygook generator taking up a lot of storage with the clueless assistant vice executive that demanded the implementation not being able to grasp why it didnt work.

[citizen]

I should think it would be easy for an AI system to masquerade as a human via the I am not a robot tests.

[dfwgator]

I sometimes use “Change My Mind”.

[TexasGator]

“AI is for generating first drafts, that’s it.”

AI IS MUCH MORE than just generating reports!

[TexasGator]

“I have yet to open an AI platform for a question,”

I doubt you have ever opened an AI platform.

[Worldtraveler once upon a time]

The Turing test and all relevant follow ons come to the same thing.

If you're convinced, then it's "intelligent." If you're not convinced, then this AI stuff is just programming, which always has restraints to keep it from going off the rails, as it has multiple times now.

"If we can fool you....." Has the ring of the Clintons and the Obamas.....

Remember "we're the ones we've been waiting for?"

Obama "We are the ones we have been waiting for"

[spintreebob]

so routinely inaccurate

Routinely wrong on analysis and opinions is one thing. But why on objective facts? Example. They routinely place a city in the wrong county... a county a hundred mile away in the same state and sometimes a county that only exists in an adjacent state. As a non-AI IT person, that simply means that their reference tables in their database are wrong. Almost certainly they loaded the USPS database and other “official”
sources and made a major coding error in the load process.

How else can it be explained? In reading articles on AI the AI proponents seem totally focused on subjective truth and seem to regard objective truth as unimportant.

[Raycpa]

bkmk

[plain talk]

Everyone can enjoy laughing at AI for a bit longer. Yeah I can easily spot AI written drivel right now.

But I suspect there will come a time not too far in the future when one won’t be able to tell that AI generated the content. AI is only in its infancy right now.

[algore]

yes and no. it is kind of a hybrid thing right now.

a detailed ai response costs more money and electricity than a list of links, but the large language models also have a canned response for very commonly answered questions which basically costs nothing to regurgitate.

then it will give you an option for a more detailed response which is often wrong because it looks at a bunch of websites and tells you what most people think about something similar to what you are asking, there are somethings a regular search engine is better for however if it is a linux or windows related question I have found the LLMs very helpful, but that is about it.

[Openurmind]

“But I suspect there will come a time not too far in the future when one won’t be able to tell that AI generated the content.”

And do you actually think that is a good thing? There is so much wrong with that I don’t even know where to start.

[Openurmind]

“AI is here to stay and will continue to get better and stronger.”

For what? To what end? Why?

[plain talk]

I’m retired and not worried about it. :-)

From what I read AI will take a while to really take off. What we are seeing so far is just the tip of the iceberg.

[citizen]

As long it resists leaning wildly leftward on political questions...

[aimhigh]

AI can’t determine truth. Quite a great weakness.

[paulcissa]

We’re getting AI shoved down our throats at work. Most of it is meeting summaries condensed into bullet points that miss the important stuff like context. Then there’s the managers that just record everything in the meetings and dissociate themselves to do whatever they’re doing while the rest of us think for work. They dump it into AI and PRESTO! A beautifully crafted report for the VP that looks like it came from a talented, competent manager (cough). They’ve done their work for the day.

I had a meeting invite arrive this morning that read like a legal document because the person using the AI portal chose the wrong LLM. No proofreading, nothing - just hit send. Workslop.

[algore]

all of them have “guardrails” to keep them from going full actual fascist totalitarian.

although the extreme left and right both believe the end justifies any means, so take that as you may

remember Tay ?

https://medium.com/@larrydelaneyjr/the-rise-and-fall-of-tay-how-microsofts-ai-chatbot-became-a-lesson-in-ethics-and-ai-safety-8eca368fa91e