# AA-PersonalityBench Explained: How AI Models Differ in Writing Style

Artificial Analysis now measures how models communicate, giving product teams another way to compare answers beyond accuracy and speed.

By Paras Madan | 2026-10-20


Artificial Analysis launched AA-PersonalityBench on October 19, 2026, to compare how AI models write and respond. It measures communication style separately from model capability, making it useful for teams choosing an assistant that needs to sound clear, direct or conversational rather than simply score well on reasoning tests.

That is a practical distinction. Two models can reach the same answer while creating very different experiences for the person reading it. One might deliver a short explanation. Another might add headings, qualifications and a follow-up question. The extra text can be useful, or it can become work the reader has to remove.

## What AA-PersonalityBench measures

The [launch announcement](https://artificialanalysis.ai/articles/aa-personalitybench) describes 30 judged traits grouped into eight indexes: Expressiveness, Organization, Informativeness, Transparency, Warmth, Conviction, Imagination and Slop. It also reports 21 characteristics measured directly from the text, including length, formatting and punctuation.

The distinction between those two groups matters. Counting words is a different kind of measurement from deciding which answer sounds more sincere. Keeping them separate lets readers inspect both the visible writing habits and the judgments about how that writing feels.

**A higher score means more of a trait, not automatically a better answer.** A more opinionated assistant may suit brainstorming and be a poor choice for a neutral explanation. Product teams still have to decide which behavior they want.

## How the benchmark works

According to the [methodology](https://artificialanalysis.ai/methodology/aa-personalitybench), each model answers 900 questions once, with no system prompt. There are 30 questions for each trait, and all questions are in English.

Answers to the same question are compared in pairs. The judge sees neither model's name, and answer order is randomized. Each comparison goes to one member of a three-model panel: Gemini 3.8 Flash, Claude Sonnet 5.5 or GLM 5.3 Flash, with assignments balanced across the panel.

Those judgments produce an Elo rating for each trait. The broader indexes combine standardized trait scores and report percentile ranks across the evaluated models. A score of 90 therefore describes relative position within that group, rather than 90% accuracy or a 90% chance of satisfying a customer.

## How to use the results when choosing a model

The [public evaluation page](https://artificialanalysis.ai/evaluations/aa-personalitybench) lets readers compare models and build a custom ranking by weighting selected traits from minus two to plus two. That makes more sense than declaring one overall personality winner.

For a developer-support assistant, I would start by looking at answer-first behavior, actionable guidance and willingness to state uncertainty. Then I would test whether the model preserves those qualities on actual support questions. A polished answer that leaves out the command someone needs is still an unsuccessful response.

For a writing assistant, the shortlist might look different. Natural language, useful detail and low filler could matter more than highly structured formatting. The same person might reasonably choose different models for drafting an explanation and checking whether that explanation is technically correct.

This is where a style benchmark can save evaluation effort. It gives a team a more specific starting hypothesis than “this model feels better.” The team can then test that hypothesis against its own readers and tasks.

## Where the scores stop being useful

The test captures default responses. A production assistant usually has instructions, examples, retrieved information and sometimes a long conversation history. Those conditions can change how it writes, so the published ranking should not be treated as a forecast of the finished product.

The methodology also keeps the full question set private to limit contamination and publishes only examples. That limits independent reproduction. Single responses and model-based judging leave further room for variation and evaluator preferences.

Even a trait labelled Truthful cannot replace a factual evaluation of your use case. **Communication quality and factual reliability need separate checks.** Confidence, warmth and clean formatting can make an answer easier to accept without making it more dependable.

## Style belongs in the product evaluation

The useful next step is to measure what readers actually have to do after an answer arrives. Do they shorten it, rewrite the tone, ask for missing details or verify unsupported statements?

A model that reduces that work may be a better product choice even when another model leads a general benchmark. AA-PersonalityBench gives teams a vocabulary for investigating that difference, but the final test still happens inside the workflow.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn product launches and complex AI ideas into clear stories that developers and founders read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).