Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Show HN: We fingerprinted 178 AI models' writing styles and similarity clusters (rival.tips)
78 points by nuancedev 52 days ago | hide | past | favorite | 22 comments
We have a dataset of 3,095 standardized AI responses across 43 prompts. From each response, we extract a 32-dimension stylometric fingerprint (lexical richness, sentence structure, punctuation habits, formatting patterns, discourse markers).

Some findings:

- 9 clone clusters (>90% cosine similarity on z-normalized feature vectors) - Mistral Large 2 and Large 3 2512 score 84.8% on a composite metric combining 5 independent signals - Gemini 2.5 Flash Lite writes 78% like Claude 3 Opus. Costs 185x less - Meta has the strongest provider "house style" (37.5x distinctiveness ratio) - "Satirical fake news" is the prompt that causes the most writing convergence across all models - "Count letters" causes the most divergence

The composite clone score combines: prompt-controlled head-to-head similarity, per-feature Pearson correlation across challenges, response length correlation, cross-prompt consistency, and aggregate cosine similarity.

Tech: stylometric extraction in Node.js, z-score normalization, cosine similarity for aggregate, Pearson correlation for per-feature tracking. Analysis script is ~1400 lines.



> "Models with >75% writing similarity but massive price gaps. The cheap model writes the same way. You are paying for the brand.

* > ...*

* > Gemini 2.5 Flash Lite Preview 06-17 and Claude 3 Opus: 78.2%*

As someone who has tried to use many of these models for writing assistance, you're very wrong here. It really matters whether the model can get what I'm trying to communicate well enough to be helpful, or else I'll just write it myself. If you actually play with them a bit it's very clear these models are not substitutes. This goes for many on your list!


It makes sense. The cheaper models are often distilled versions, so they may ape language but miss the connective tissue that makes the entire output coherent.


Models have their “personalities” for sure but that expensive model is better is maybe just a confirmation bias.

(There was a blind test in Wine Enthusiasist magazine - even sommeliers didn’t recognize expensive wines from cheaper alternatives.)

But ofc if you get perfect results in one shot from expensive model, it is cheaper than wrangling with cheap model for an hour…(just an example).

But what I see hard is to navigating so many models available - HuggingFace has 2,769,687 models listed…

So every comparison like this or at models.dev or arena.ai is good.


I'd bet this whole thing is vibe'd out of nothingness and no human actually thought about whether saying "you are paying for the brand" makes any sense at all.

How the hell are companies and individuals not taking reputational hits for saying blatantly wrong things in AI-voice, under their name?


Also are Gemini and Opus not both big brands? If it was some small ai shop vs opus then sure. So seems indeed to make little sense?


Also, is it "paying for the brand" or "paying for the training"?


It would be shocking to me if the large model trainers didn't have tools like this to analyze their outputs, but this is interesting work!

You can see who likely (post)trained/distilled their models or borrowed parameters from each other. I do wonder if the 32 dimensions were chosen/named from principal components or pre-selected and designed, but the tool seems like an effective discriminator in any case.

Were the prompts similarly selected for orthogonality? I've wondered how the different LLMs would respond from iterative zero-shot prompt_n generation by summary from a previous response_n to generate zero-shot response_n+1. Would it statistically converge to a more distinguishable prompt for that LLM?


I've always wondered if the "typical" AI writing style is just an unavoidable RL artifact or a deliberate fingerprint to prevent model collapse as low-effort AI-generated text floods the training data pool (the web).


Without showing the prompts and responses, it's yet another meaningless AI benchmark.

Many of those numbers do not really match what I've seen in the wild, and without clear illustration why you arrived at the number it's not a helpful number.


Ugh. subheadings were a major turn off.

I expected it to be an analysis of AI-generated writing styles. Not full of them.

;)


It's a spam account, like nearly all submissions about AI their account is just self promotion history. Together with the subject matter I'd expect nothing but lazy bullshit.


One thing I don't understand then is why Hackernews upvotes these posts when there are some other SHOW HN posts which sometimes are so much nicer and thoughtful like (https://news.ycombinator.com/item?id=47589735) etc.

our community is shooting a gun in its own feet if it continues to upvote complete self promotion, I try to upvote as many cool projects I find here but the number games is definitely frustrating

Are we entirely sure that this person hasn't used AI bot to upvote his comments to front page, I wish to much rather believe that than people upvoting it especially when most if not all comments are about how it feels extremely AI slop.

maybe such forms of (rage/click bait?) truly sells in that regards and HN isn't so invulnerable as (I) we think it is.


At least once a week, I take a few minutes out of my day to look through `/shownew` to comment on and upvote posts that:

- seem interesting but aren’t getting much traction

- are from users who actively participate in the community


No mention of any linguistic theory, some arbitrary (?*) metrics mixed together and even more arbitrary thresholds. Why does 75% "similarity" mean "writes the same"?

Low quality post imo.

*Generated I assume.


Assuming this is part of the reason I constantly see "--" whenever I prompt Claude or ChatGPT


Amazing,last time I let GPT guess Claude content,it guess GPT made it


The muted colors on a dark background makes everything hard to read.


This is the stock CC vibe-coded aesthetic - it loves to give the middle finger to even the most basic of web accessibility requirements like contrast ratios.


“you are absolutely right!” Should be at the top of the list


Has anyone else used LLMs to fact check other LLMS?

I hate to say it, but Gemini lies less frequently than paid models from OPenAI and Anthropic (Open AI is worst in my use cases).

My guess is that Google has better training data (and uses less synthetic data which might be creating training feedback loops in other models), has more of a "be calibrated" model than a "be helpful" model, but it could just be that they leverage more RAG than leveraging weights more.

But, I really shouldn't speculate the "why" as I'm out of my domain. Just curious if others use all the models they can and compare outputs as much as I do.


Besides claiming opus and gemini flash share 99% of style being suspicious, the point that you are wasting money on the expensive model is non sensical. You pay primarily for the intelligence, not the writing style.

Is this article AI slop?


I'm curious about the sorts of users who care about style but will either one-shot with default style, not providing samples or direction, or who even choose models on that style rather than, you know, substance.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: