Extracting Training Data from Large Language Models · HackerPapers discussions

Extracting Training Data from Large Language Models — mentioned by dijksterhuis

Sat, 05 Sep 2026 00:26:01 GMT

prediction is a very specific term of art in the field of machine learning. generally speaking, machine learning models like LLMs are based on probability; performing a statistical prediction of the likely y given some input x Probability(y | x) that's why we refer to outputs as a prediction. it is likelihoods and stuff. the output is never definitely correct as we're not dealing with heuristic processes. > Prediction implies there is some "truth" or event or something that you can test against there absolutely is a ground truth during training. the core predict-the-next-most-likely-token…

Paper and discussion history. A mention is not an endorsement.

Extracting Training Data from Large Language Models — mentioned by dijksterhuis

Thu, 30 Jul 2026 12:55:34 GMT

> To my knowledge there is not a single known case where this happened. > News back then were about intentionally prompting to output known copyrighted material. First, there are other cases if you take the time to dig. This is quite an old example (GPT-2) as i haven't kept up to date on this field recently, but it does show that this problem has been known about since before these systems were widely adopted: https://arxiv.org/abs/2012.07805 [0] Second, GP said nothing about the type of effort required to make it happen, just that it can be done and that the copyright owner could come along…

Paper and discussion history. A mention is not an endorsement.

Extracting Training Data from Large Language Models — mentioned by Retric

Wed, 31 May 2023 22:59:10 GMT

AI models will make 1:1 copies of training data where artists try and avoid doing so. It’s common to obscure this copying by intentionally inserting lossy steps, but making an MP3 isn’t a new work. It’s most obvious when large blocks of text are recreated, but the core mechanism doesn’t go away simply because you obscure the underlying output. “Extracting Training Data from Large Language Models” https://arxiv.org/abs/2012.07805

Paper and discussion history. A mention is not an endorsement.

Extracting Training Data from Large Language Models — mentioned by vitehozonage

Thu, 20 Apr 2023 08:33:00 GMT

Most current "AI" models essentially compress and then store the information, often to the point where training data can be extracted [1][2] [1]: https://arxiv.org/abs/2012.07805 [2]: https://arxiv.org/abs/2301.13188

Paper and discussion history. A mention is not an endorsement.

Extracting Training Data from Large Language Models — HN submission

Sun, 26 Jun 2022 12:37:21 GMT

Ask HN: GPT-3 reveals my full name – can I do anything?

Alternatively: What's the current status of Personally Identifying Information and language models? I try to hide my real name whenever possible, out of an abundance of caution. You can still find it if you search carefully, but in today's hostile internet I see this kind of soft pseudonymity as my digital personal space, and expect to have it respected. When playing around in GPT-3 I tried making sentences with my username. Imagine my surprise when I see it spitting out my (globally unique, unusual) full name! Looking around, I found a paper that says language models spitting out personal…

710 points; 346 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Extracting Training Data from Large Language Models — mentioned by qayxc

Sat, 10 Jul 2021 00:00:31 GMT

> Would it be OK if I launched a tech preview of my AI poem writer companion that would copy lines but also complete stanzas from famous poets, rock bands and singer-songwriters? Yes it would be if it only happened ~0.1% of the time and if quoting verbatim wasn't the intended function of the system but merely a side-effect. In fact, that's what artists sometimes do deliberately. It's what happens with other GANs as well and all that needs to happen is to educate users about the possibility of this. As long as you don't take ownership of the output produced by your AI (and neither do…

Paper and discussion history. A mention is not an endorsement.

Extracting Training Data from Large Language Models — HN submission

Wed, 30 Jun 2021 16:28:01 GMT

Extracting Training Data from Large Language Models

1 point; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Extracting Training Data from Large Language Models — mentioned by ma2rten

Thu, 18 Feb 2021 02:50:54 GMT

Yes, it has been shown [1] that these models can memorize information even if it only appears once in the training data. It could potentially cause issues with privacy and especially copyright. However, in this case it's not relevant because they didn't release the model weights. [1] https://arxiv.org/abs/2012.07805

Paper and discussion history. A mention is not an endorsement.

Extracting Training Data from Large Language Models — mentioned by jepler

Tue, 22 Dec 2020 02:29:54 GMT

I am guessing it was a blog about this paper: https://arxiv.org/pdf/2012.07805.pdf "Extracting Training Data from Large Language Models"

Paper and discussion history. A mention is not an endorsement.

Extracting Training Data from Large Language Models — HN submission

Wed, 16 Dec 2020 05:50:02 GMT

Extracting Training Data from Large Language Models

2 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Extracting Training Data from Large Language Models — HN submission

Tue, 15 Dec 2020 18:47:01 GMT

Extracting Training Data from Large Language Models

1 point; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.