RLAIF gone wrong 😑

😍

This is great, I’d just add:

1. Another challenge is lack of serious engagement from AI researchers with humanities scholars.

2. This new work makes it sound like it just takes some finetuning to get big improvements.

bsky.app/profile/tuhi...

Thanks @mariaa.bsky.social for highlighting. I think finetuning can reduce AI writing tics though the writing would still not be original. Its derivative in someone else’s style. And then there is this mega challenge of balancing consistency, style, coherence especially for long form writing

Link to our paper : www.arxiv.org/abs/2510.13939

TLDR: Author-specific fine-tuning enables non-verbatim AI writing that readers prefer to expert human writing, thereby providing empirical evidence directly relevant to copyright’s fourth fair-use factor, which is directly relevant to the market-effect/substitution question courts are asking. 8/n

👩‍⚖️Policy Takeaway : A reasonable solution would require the model to implement guardrails that would disable it from generating non-parodic imitations of individual authors’ oeuvres through targeted finetuning or prompting. 7/n

While we do not account for additional costs of human effort required to transform raw AI output into cohesive, publishable prose, the median fine-tuning and inference cost of $81 per author represents a dramatic 99.7% reduction compared to typical professional writer compensation. 6/n

Fine-tuned AI outputs fooled best AI detectors 97% of the time vs. just 3% for prompted ones.

Why?

Because Fine-tuning removes telltale AI quirks like clichés that detectors flag and readers dislike, flipping the detectability-preference link. 5/n

Results: Incontext prompted text was strongly rejected by Experts but showed mixed results w/ lay readers.

Fine-tuning #ChatGPT on authors’ ouevre completely reversed the findings: Experts now favored AI for style (OR=8.16) and writing quality (OR=1.87), with lay readers showing similar shifts. 4/n

MFA trained writers emulated 50 award-winning authors in reproducing 450-word excerpts.
150+ readers blindly compared human vs AI versions.

Two AI methods tested:
- In-context prompting (ChatGPT, Claude, Gemini)
- Fine-tuning on complete works.

Both used same prompts for fair comparison. 3/n

Why “market harm” matters? If an AI output becomes a close substitute for the original, it can hurt the market for those books. This is directly relevant to the fourth fair-use factor.

US Copyright Office recently acknowledged this 2/n

In Roma. Near Alfonso Curaons childhood home

Well deserved 👏👏

Just read AI and The Everyday writer . Brilliant article. www.cambridge.org/core/service...