I was thinking about this after a discussion at work about large language models (LLMs) - the initial scrape of the internet before Chat GPT become publicly usable was probably the last truly high quality scrape of human-made content any model will get. The second Chat GPT went public, the data pool became tainted with people publishing information from it. Future language models will have increasingly large percentages of their data tainted by AI-generated content, skewing the results away from how humans actually write. To get actual human content, they may need to turn to transcriptions of audio recordings or phone calls for training, and even that wouldn’t be quite correct because people write differently than they speak.
I sort of wonder if eventually people will start being influenced in how they choose to write based on seeing this AI content. If teachers use AI-generated texts in school lessons, especially at lower levels, will that effect how kids end up writing and formatting their work? It’s weird to think about the wider implications of how this AI stuff will ultimately impact society.
What’s your predictions? Is there a future where AI can get a clean, human-made scrape? Are we doomed to start writing like AIs?
Writing is not easy, people go to college for years to learn how to do it, unless the actual skill of writing can be instilled into an LLM, they won’t replace people.
The companies that try and use them to replace writers now are the companies that will feel the repercussions first: poor quality, no experienced employees, and lack of business.
An LLM will never be able to replace writers, they lack an understanding of the core concepts that are actually involved.
Besides, who is going to write things to train the AI?
It’s not going to replace actual dedicated writers, but it’s definitely going to hinder people learning to write and make up a large portion of the text online. It may also make it harder for actual writers to be found in all the noise. I heard a little while back about a scifi magazine which had to close its submissions because it was getting too many AI-written stories and sorting through the real versus fake was becoming difficult for them.
As for who’s going to train the AI, that’s part of what I’m arguing here - future LLMs are going to wind up being trained on AI-generated text because there will be so much of it online that screening it out becomes near impossible. Reddit mods already have challenges screening out chat GPT bots from their comments. When a future LLM scrapes the web for writen words, it’ll come back with lots of garbage AI text which will taint its learning pool. AIs will learn from AIs and become worse for it.
Yes but many industries, writing included, already suffer from fraudulent activity.
One of the largest reasons why entry-level software engineering jobs require five years experience is because consulting firms train developers (making the juniors) and have them interview as senior developers.
The firm manages the job search, helps the consultants during interviews, and they have teams helping with the actual work as well as fitting in with the rest of the, more experienced (or also fraudulent), staff.
There are currently industries which make a profit off of fooling other companies and consumers, which are arguably more frightening.
If anything, this will increase demand for better human writers and ways to authenticate their work. If it doesn’t, we’ll get sick of the content AI creates before it gets too bad.
This is a classic feedback problem: you use a microphone to amplify your voice, but If the mic picks up the amplified sound it creates audio feedback + a sharply increasing wail.
I can’t imagine what LLM feed back ‘sounds like’, but a guarantee you it ain’t pretty.
I think it could end up being a problem that we face in the future, but probably not an insurmountable one.
For one, I suspect that clean data sources will always be available, though it could become a lot more expensive to obtain. As an extreme example, you could always source your data by recording in-person conversations.
Also, as AI improves, I’m guessing it will be able to handle bad data more gracefully, and that it should be able to train to the same effectiveness while using a smaller dataset.
I feel like if you tried to train an LLM on spoken conversational English the output would just be “yeah um yeah um yeah um”
But on a more serious note spoken English is very different than written.
Either way you can find validated sources of human written text it just won’t be as easy.
Maybe an LLM that can have a normal sounding spoken conversation will be a next step. The Turing test but speaking instead of typing. I assume the neural networks could learn things like intonation.
I suspect the quality LLM development teams will pursue the same in-depth data sourcing & cleaning techniques that quality ML researchers are developing today. Or rather, they’ll do something similar in principle to mitigate this issue.
I still agree with your conclusions. It will be a bigger consideration and less scrupulous teams will be more effected.
You have to remember one thing, writing or speaking of a language is not a fixed scientific law or math formula that will stay true through out history. A living language is always moving and evolving in most of its components, be a vocabulary, grammar, or even meaning of words/phrases. We are just entering an era where AI generated content someone might feel appealing and follow that style, compare to copy a contemporary popular writer.
Indeed. As long as the language is still expressive and we understand what is being communicated, I don’t see why it would matter if it “sounds like” AI or not.
If it really becomes a problem then just curate the training data better to exclude the stuff that “sounds like” an AI. Doesn’t matter if it’s actually written by an AI or not, just select the training data that matches what you’d like the AI to learn and go with that. There’s not some kind of magical ghost present in human-written words that’s absent in AI-written words, if the words are the words you want then that’s all that matters.
I’ve heard this theory. Feels like unrealistic hopeful wishes of people who want AI to fail.
LLM processing will be a huge tool for pruning and labeling training sets. Humans can sample and validate the work. These better training sets will produce better LLMs.
Who cares is a chunk of text was written by a human or not? Plenty of humans are shit writers who believe illogical or clearly incorrect things. The idea that human origin text is superior is a fantasy. chatGPT is a better writer than 80% of humans todat. In 10 years LLMs will be better than 99.9% of humans. There is no poison to be avoided.
chatGPT has an apparent style when used in the default mode, but you can already get away from that with simple prompt tweaks. This whole thing is a non-issue.
LLM generated text can also be easily detected provided you can figure out which model it came from and the weights within it. For people training models, this won’t be hard to do.
I agree with the take that getting better and better datasets for training is going to get easier over time, rather than harder. The story of AlphaZero is a good example of this too - the best chess AI quickly trounced any AI trained on human games simply by playing against itself. To me, that suggests that training on LLM output will lead to even better results, since you can generate so much more of it.
The chess engine’s training is anchored by the win/lose outcome of the game. LLM training is anchored by what humans like to read and write. This means that a human needs to somehow be in the loop.
I think OpenAI’s own chatGPT detector had double digit false negative and positive rates. I expect as diversity of LLMs proliferates, it will become increasingly harder to detect.
Well thankfully we have tons of archived conversations before ChatGPT on archive.org.
Many forums with hundreds of thousands of posts are just waiting to be scraped.
thanks, it’s terrifying
Thanks, I hate it
There has already been jokes of AI being used to create well crafted correspondence, then another AI translating that into a short summary.
I think you are going to see AI as something people lean on more to talk to others, and that is going to create its own language where AI talks to AI.
That’s not a joke – that’s exactly how a lot of the smaller open-source LLMs are trained. Orca (paper) is trained between GPT-4 and GPT-3.5-turbo
I’m not sure this is true. They could be trained based on published works prior to a certain date as the formal writing style, eg Project Gutenberg, then layer on the recent internet to better capture modern stylistic trends.
Ultimately, the models will always require fine tuning, and selecting which data set you use for early training has a very large impact on the overall performance of the model. Additional knowledge and trendiness can be learned after the fact.
This sounds like what an ai would write /s
I think that while LLMs are going to get worse, the AI software will get better to the point of strong AI, and it will do a lot of “apple-esque” changes to mass produced speech that will ultimately be for the better… The cynical possibility is that it will further taint human dialogue even though it could provide a better way.
There’s usually a context difference that might might be significant. People don’t write the same way way for an email, like they would a letter, text message, or tweet.
They might write more like an LLM for things like essays and reports, but your usual writing is probably still fine. Then classics that inspire people to write are still around, and I doubt that they would be supplanted by an LLM any time soon.
We might start being in trouble if people start republishing books with them, but that’s unlikely to to happen any time soon, considering the current state of copyright around AI works.
There are some in the research community that agree with your take: THE CURSE OF RECURSION: TRAINING ON GENERATED DATA MAKES MODELS FORGET
Basically the long and short of that paper is that LLMs are inherently biased towards likely responses. The more their training set is LLM generated, and thus contains that bias, the less the LLM will be able to produce unlikely responses, over time degrading the model quality throughout successive generations.
However, I tend to think this viewpoint is probably missing something important. Can you train a new LLM on today’s internet? Probably not, at least without some heavy cleaning. Can you train a multimodal model on video, audio, the chat logs of people talking to it, and even other better LLMs? Yes, and you will get a much higher quality model and likely won’t get the same model collapse implied by the paper.
This is more or less what OpenAI has done. All the conversations with 100M+ users are saved and used to further train the AI. Their latest GPT4 is also trained on video and image recognition, and they have also been exploring ways for LLMs to train new ones, especially to aid in alignment of these models.
Another recent example is Orca, a fine tune of the open source llama model, which is trained by GPT-3.5 and GPT-4 as teachers, and retains ~90% of GPT-3.5’s performance though it uses a factor of 10 less parameters.
I believe I read/heard somewhere that future AI training will take place using less data and will potentially pay field experts to better curate signal from noise
