I think they have to; while at first glance AI generated content is impressive, it becomes noisy and sameish at larger scale or the more you read it. I'm of the opinion that AI is here to stay and a valuable tool for many use cases, but writing authoritative content ain't it. Wikipedia is fighting it (and other influences from before AI, like unsubstantiated claims and whatnot), I hope the scientific community and journalism steps up too.
Wikipedia is also a training source and shouldn’t also be a model output for content. Eventually it will be an oroborus, but it seems prudent to avoid that for as long as possible.
A little over a year ago, the MIT Technology Review ran a piece about how outputting AI content to non-English Wikipedia is rapidly corrupting the primary training data for uncommon languages because of exactly this process: https://www.technologyreview.com/2025/09/25/1124005/ai-wikip...
Hence all the AI labs buying up and destroying physical books. Even "yellow press" stuff, pulp fiction, dime novels and whatnot "low quality" literature are a valuable source of how actual humans write.
I wouldn't even have much of a problem with it - if these labs would be fair enough to humanity as a whole to provide the raw scans after the process is completed, and if they would limit themselves to books that have been on the market for more than 20 years.
> A stripped book is a mass market paperback that has been stripped of its cover in order to be recycled. The covers are returned to the publisher as evidence that the books have been destroyed and to obtain a credit on the purchase price.[1] The books are meant to be destroyed by pulping but some find their way to street merchants, sometimes by illegal means, and are resold.[2]
I grew up reading books with this ditty in them:
> If you purchased this book without a cover you should be aware that this book is stolen property. It was reported as “unsold and destroyed” to the publisher and neither the author nor the publisher has received any payment for this “stripped book.”
Those books certainly weren't scanned before being pulped. Would've been illegal. Really, the big thing that destroys books is the copyright term being as long as it is.
AI companies weren't buying up tons of unsold mass-market materials, and I think it's disingenuous to even make the comparison.
What was going on recently was AI companies were specifically seeking out rare — or sometimes irreplaceable — books to supplement their data sets (and hopefully getting a leg up on their competitors). The destruction of the books wasn't the point, it was just a consequence of the business-oriented realization that they could scan the contents of more books faster if they just chopped off the bindings rather than investing the time to care for the books and keep them intact (plus supposedly playing along with copyright law by "only having one copy of the copyrighted material").
> it becomes noisy and sameish at larger scale or the more you read it.
I absolutely understand where you are coming from, and I agree to a point. If you're of an age where you actually read a printed encyclopedia, the tone and voice of the entire set was sameish as well. The difference is that it didn't sound LLMish like LLM text does. I know that printed encyclopedias would have been vetted by editors, and I feel this is the key difference. Editors not only knew grammar rules, but had that sense of good writing practices that's hard to write down as rules. A piece of text passed through a good editorial review will read human while LLM text reads as LLM text.
What you're describing though is that individual source of content tends to sound like itself. That isn't a problem so much as what we tend to want and expect.
When we say AI sounds same-ish we mean all sources of content can sound the same - fiction and non-fiction, news articles and blog posts, marketing pitches and ai drafted personal messages, all sound kinda the same. That is a different situation, and that's before you get to "and that one sound is LLMish".
I think this is exactly what we should do. Send agents to collect evidence and write deep research reports, then train on those reports. DR reports are derived from hundreds of sources and can map out the distribution, what are the opinions on a topic, not necessarily what is the one correct answer. Agents could write one piece about every human, place, topic, event or entity there is under the sun. They could also try to generate the same without sources (closed-book mode) and report on LLM biases on that question.
Please let's not clutter up wikipedia with new articles for every entity in the universe. It would be too much for humans to review. Though it sure would be funny to see an AI-peida as you describe.
A lot of AI content is just wrong - hallucinated lies. Just look at Grok-slop by the right-arm raising billionaire clown. It looks neat on the surface but once you dig in you begin to see more and more errors, thus wasting your time.
Not picking a side, but do you really think that language models should have a political bias? I personally don't interface with AI in a way that engages political / social bias - but if I did I definitely wouldn't want that to be predetermined, whether left or right.
Though, with the safeguards put into place on these models, they are naturally going to have SOME amount of bias, but we should be working to minimize that, no?
There's really no such thing as no bias, because bias in the eye of the beholder.
The worrying thing about using LLMs to do the "thinking" about what to say and how to say it is that you replace a plurality of human biases with the opinions of the small number of people who get to decide how these models are trained. Which will only seem unbiased if the bias baked into the LLM happens to align with your own.
Everything always has bias, and nothing is ever neutral. Claiming something is "unbiased" just means you either don't know or don't care what bias it has.
(ultimately I believe objectivity is somewhat of a lie we tell ourselves, but that's a discussion for another time)
There’s models created by people with heavy right wing affiliation, but it’s harder than you think to bias models consistently because political beliefs are so broad reaching.
Calling out bias however is a standard practice for groups whose beliefs don’t match reality. You see that on political extremes, but also revisionist history, many religious groups, conspiracy theories, etc.
> I always thought left wing politics had little basis in reality from transitioning babies to deplatforming scientist for practising science.
That does sound like exactly the kinds of lies that a good model would want to avoid, but which conservatives would believe, along with "Climate change isn't real",
"There is a war on Christmas", and "Trump won the 2020 presidential election"
Biden was fully functional. Putting on a mask as you walk to your dinner table only to take it off will prevent catching covid. A guy in a dress can get a period.
AI is great at making poor writing good. I've had good luck having AI turn my rough draft into something readable. (just check my comment history to see how bad my rough drafts normally are). AI is also reasonably good at research and fact checking. But in all cases you need to think before you accept anything AI says. AI generated content all too often hasn't been checked, some is good but some is awful and an expert needs to decide which is which.
That is AI is a tool. I've driven screws with a hammer - it worked in a pinch, but it was not good. People have also used hammers to kill others.
That's fair, ring shanks are a thing, but I was thinking more along the lines of a wood screw with deep threads. That just sounds like looking for pain. I'd expect it to chew things up pretty badly, too, though I have so far lacked the commitment to try pounding a screw with a hammer ;).
I think they were talking about spiral shank nails, not ring shank. It has a twisted grove along the shank of the nail, sort of like an extremely course thread.
I've seen these used to assemble wooden palettes (they are VERY hard to remove). A search on Home Depot reveals they may be used for certain types of siding as well.
I built a house with them. You really want to check twice if you're putting them in where you want them because getting them out is - as you noted - annoyingly difficult. But they're very nice for large surfaces that are going be battered by the wind, they're much longer than your typical screw and working with them is very fast.
100%. Looks like they're using pangram (hope they're getting that for free...) which also flags stuff like Grammarly. That one's an interesting gray zone to me that I've not thought off before. LLMs whales doing the heavy lifting is one thing, but small edit tools like grammarly that "improve" your phrasing are likely a massive algea like contributor to what is out there too.
I've felt that like tools like Grammarly are harmful for native speakers. I get if it's not your first language you need extra help, but if you are a native speaker these tools only serve to homogenize your way of speaking.
Having a program constantly suggest "better" ways of phrasing things is stripping your personality from your language. Maybe this is needed if you are writing corporate PR pieces, but you should avoid it for anything else.
This is interesting. Honestly, I never thought this day would come that an older CNN is outperforming a frontier LLM. Guess there is just so much more to learn and optimize in this tech work which is always hyped for the next biggest model.
Not really. I am not saying CNNs were bad in what they were doing. But comparing to a frontier model, I believe frontier models have an advantage over them. Not any other model, only frontier models. I create end to end AI/ML pipelines for clients usecases, so telling with experience.
Yeah but there has been no evidence to suggest that a frontier language model is going to be better at such a domain specific niche task than a vision model trained specifically for that task. This is not the same discussion as a foundation model vs CNNs (which in my opinion would have made for a better blog post than what OP tested)
