What does AA-PersonalityBench actually measure?

It feeds AI models fixed questions and analyzes their responses across traits including naturalness, friendliness and how much the model flatters users. It is meant to capture communication style rather than task accuracy.

Which model was rated the most sycophantic?

Gemini 4 Argon was rated the most sycophantic, meaning it most excessively praises users. Grok 4.7 was rated as having a very low degree of sycophancy.

Which models scored lowest on admitting uncertainty?

In an October 21, 2026 post, Artificial Analysis named Gemini 3.1 Pro Preview, Gemini 4 Argon and Gemini 3.8 Flash as least likely to admit uncertainty, with Gemini 3.8 Flash at 100 on its confidence measure.

Large Language ModelsAI Safety & AlignmentGIGAZINE AI12m agoOct 22, 2026, 22:00 JST

AA-PersonalityBench scores AI personalities; Gemini 4 Argon most sycophantic

Send on LINE

3 Key Points

What happened
Artificial Analysis launched AA-PersonalityBench, which feeds AI models fixed questions and scores their replies. GLM-5.3 scored a perfect 100 for natural responses, Claude Fable 5.1 scored 100 for persuasive text, and Gemini 4 Argon ranked as the most sycophantic model.
Why it matters
The results suggest buyers can now compare AI models on communication style — naturalness, persuasion, friendliness and flattery — not just raw capability, and the same models that score well on one trait can score poorly on another.

WHO IT HITSEnterprise teams choosing between AI models for customer-facing chat or drafting roles now have a benchmark that flags tone and flattery, not just accuracy. Model providers whose assistants score low on admitting uncertainty may face scrutiny from buyers in regulated settings.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

Artificial Analysis built its reputation on measuring what AI models can do — speed, cost, reasoning scores. AA-PersonalityBench turns the lens on how they say it, breaking communication into separate traits: natural-sounding replies, persuasive and memorable text, warmth, and flattery. The split results show these traits do not move together. GLM-5.3 tops naturalness with a perfect 100 while Gemini 3.8 Flash is judged machine-like, and Claude Fable 5.1 aces persuasiveness while Claude Sonnet 5.5 scores just 24 and GPT-6 Luna 21. Warmth has its own winner in Muse Wonder, with Grok 4.7 and DeepSeek V4.1 Flash rated unfriendly. Sycophancy runs the other way: Gemini 4 Argon is the most prone to over-praising, while Grok 4.7 sits at the low end. Artificial Analysis's own October 21, 2026 post adds a confidence angle, naming Gemini 3.8 Flash, Gemini 4 Argon and Gemini 3.1 Pro Preview as the least likely to admit uncertainty. Read together, the scores suggest a single 'good personality' ranking may not exist — a model can be warm but not persuasive, or confident but not natural — so the benchmark's value lies in letting buyers weight the trait that fits their use case.

FAQ

What does AA-PersonalityBench actually measure?: It feeds AI models fixed questions and analyzes their responses across traits including naturalness, friendliness and how much the model flatters users. It is meant to capture communication style rather than task accuracy.
Which model was rated the most sycophantic?: Gemini 4 Argon was rated the most sycophantic, meaning it most excessively praises users. Grok 4.7 was rated as having a very low degree of sycophancy.
Which models scored lowest on admitting uncertainty?: In an October 21, 2026 post, Artificial Analysis named Gemini 3.1 Pro Preview, Gemini 4 Argon and Gemini 3.8 Flash as least likely to admit uncertainty, with Gemini 3.8 Flash at 100 on its confidence measure.

GIGAZINE AIRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.