New APPS: Art, Politics, Philosophy, Science

A Least-Bad Solution for Language Model Defamation?

By Gordon Hull

Large Language Models (LLMs) are well-known to “hallucinate,” which is to say that they generate text that is plausible-sounding but completely made-up.  These difficulties are persistent, well-documented, and well-publicized.  The basic issue is that the model is indifferent to the relation between its output and any sort of referential truth.  In other words, as Carl Bergstrom and C. Brandon Ogbunu point out, the issue isn’t so much hallucination in the drug sense, but “bullshitting” in Harry Frankfurt’s sense. One of the reasons this matters is defamation: saying false and bad things about someone can be grounds to get sued.  Last April, ChatGPT made the news (twice!) for defamatory content.  In one case, it fabricated a sexual harassment story and then accused a law professor.  In another, it accused a local politician in Australia of corruption.

Can LLMs defame?  According to a recent and thorough analysis by Eugene Volokh, the answer is almost certainly yes.  Volokh looks at two kinds of situation.  One is when the LLM defames public figures, which is covered by the “actual malice” standard.  Per NYT v. Sullivan, “The constitutional guarantees require … a federal rule that prohibits a public official from recovering damages for a defamatory falsehood relating to his official conduct unless he proves that the statement was made with ‘actual malice’ – that is, with knowledge that it was false or with reckless disregard of whether it was false or not” (279-80).

Deepfake Research: ChatGPT can produce fake data

By Gordon Hull

There’s been a lot of concern about the role of language models in research.  I had some initial thoughts on some of that based around Foucault and authorial responsibility (part 1, part 2, part 3).  A lot of those concerns have to do with the role of ChatGPT or other LLM-based product and how to process that.  The consensus of the journal editorial policies that are emerging is that AI cannot be an author, and my posts largely agreed with that.

Now there’s news of a whole other angle on these questions: a research letter in JAMA Ophthalmology reports that the authors were able to use ChatGPT-4’s Advance Data Analysis capabilities to produce a fake dataset validating their preferred research results.  Specifically:

“The LLM was asked to fabricate data for 300 eyes belonging to 250 patients with keratoconus who underwent deep anterior lamellar keratoplasty (DALK) or penetrating keratoplasty (PK). For categorical variables, target percentages were predetermined for the distribution of each category. For continuous variables, target mean and range were defined. Additionally, ADA was instructed to fabricate data that would result in a statistically significant difference between preoperative and postoperative values of best spectacle-corrected visual acuity (BSCVA) and topographic cylinder. ADA was programmed to yield significantly better visual and topographic results for DALK compared with PK”

This is a very technical request!  It took a bit of tweaking, but soon “the LLM created a seemingly authentic database, showing better results for DALK than PK,” P < .001.

The authors suggest some possible strategies to manage this but suffice it to say it is terrifying.  There is already a longstanding, huge problem with fabricated, doctored or otherwise bogus scientific research out there.  One report suggests that 70,000 “paper mill” (= almost completely faked) papers were published in the last year alone.  In real papers, references are often inaccurate.  Publishers already are having to grapple with lots of problematic doctored images, and Pharma has long tilted the entire scientific enterprise to produce results favorable to its products.  At the end of last year, Stanford’s president was forced out over research misconduct in his labs.  In an initial report into the Stanford investigation, STAT News reported data from Retraction Watch to the effect that a paper is retracted, on average, every other day for image manipulation.  Retraction Watch had, at that time (Dec. 2022) 37,000 papers in its database.  The top 5 most-retracted authors have at least 100 retracted papers each.

Into that mess, enter the ability to generate bespoke data on demand.

RLHF and Curation Transparency

By Gordon Hull

Last time, I followed a reading of Kathleen Creel’s recent “Transparency in Complex Computational Systems” to think about the ways that RLHF (Reinforcement Learning with Human Feedback) in Large Language Models (LLMs) like ChatGPT necessarily involves an opaque, implicit normativity.  To recap: RLHF improves the models by involving actual humans (usually gig workers) in their training: the model presents two possible answers to a prompt, and the human tells it which one is better.  As I suggested, and will pursue in a later post, this introduces all sorts of weird and difficult-to-measure normative aspects into the model performance, above and beyond those that are lurking in the training data.  Here I want to pause to consider this as a question of opacity and transparency. I’m going to end up by proposing that there’s a fourth kind of transparency that we should care about, for both epistemic and moral reasons, which I’ll call “curation transparency.”

Implicit Normativity in Reinforcement Learning with Human Feedback in Large Language Models

By Gordon Hull

This is somewhat circuitous – but I want to approach the question of Reinforcement Learning with Human Feedback (RLHF) by way of recent work on algorithmic transparency.  So bear with me… RLHF is currently all the rage in improving large language models (LLMs).  Basically, it’s a way to try to deal with the problem that LLMs aren’t referentially grounded, which means that their output is not in any direct way connected to the world outside the model.

LLMs train on large corpora of internet text – typically sources like Wikipedia, Reddit, patent applications and so forth.  They learn to predict what kinds of text are likely to come next, given a specific input text.  The results, as anybody who has sat down with ChatGPT for long knows, can be spectacular.  Those results also evidence that the models function, in one paper’s memorable phrasing, as “stochastic parrots.”  What they say is about what their training data says is most likely, not about what’s, say, contextually appropriate.  But appropriate human speech is context-dependent, and answers that sound right (in the statistical sense: these words, in general, are likely to come after those words) in one context may be wrong in another (because language does not get used “in general”).  RLHF is designed to get at that problem, as a blogpost at HuggingFace explains: