Overview
The AI IQ project, launched at aiiq.org, has sparked debate by assigning intelligence quotients to over 50 language models using a human IQ scale. Created by engineer and investor Ryan Shea (co-founder of Stacks, Voterbase, and early investor in OpenSea, Lattice, Anchorage, and Mercury), the site aims to simplify the landscape of AI model performance. But critics warn that reducing complex, uneven capabilities to a single number can mislead. This guide will help you understand the methodology, interpret the scores, and avoid common pitfalls.
Prerequisites
Before diving into AI IQ scores, you should have:
- A basic understanding of large language models (LLMs) and their capabilities.
- Familiarity with common AI benchmarks (e.g., ARC, MATH, SWE-bench).
- Comfort with elementary arithmetic (averages) and basic statistical concepts like normal distribution.
No coding experience is required, though we include a simple example for those who want to compute scores manually.
Understanding the AI IQ Methodology
The Four Reasoning Dimensions
AI IQ condenses model performance into four core reasoning areas:
- Abstract reasoning: Tests pattern recognition and fluid intelligence via ARC-AGI-1 and ARC-AGI-2.
- Mathematical reasoning: Uses FrontierMath (Tiers 1–4), AIME, and ProofBench to assess quantitative skills.
- Programmatic reasoning: Measures coding ability through Terminal-Bench 2.0, SWE-Bench Verified, and SciCode.
- Academic reasoning: Evaluates broad knowledge with Humanity's Last Exam, CritPt, and GPQA Diamond.
The Twelve Benchmarks
Each dimension averages scores from three specific benchmarks (two for abstract). For example, abstract reasoning combines two ARC tests; mathematical includes four. The raw scores from these benchmarks are the foundation.
The Scoring Formula
The composite IQ is a straight average of the four dimension IQs:
IQ = (IQ_Abstract + IQ_Math + IQ_Programmatic + IQ_Academic) / 4Each dimension IQ itself is an average of its component benchmark IQs (after calibration). The site maps raw benchmark scores to IQ values using hand-calibrated difficulty curves.
Calibration and Ceilings
A critical nuance: easier benchmarks or those prone to data contamination have compressed ceilings, preventing scores from exceeding 100. Harder, less gameable benchmarks maintain higher ceilings. This asymmetry ensures that a model cannot artificially inflate its IQ by acing trivial tests.
Step-by-Step: How to Interpret AI IQ Scores
Step 1: Identify the Model's Dimension Scores
Visit aiiq.org and select any model. You'll see four dimension IQ scores (e.g., Abstract: 95, Math: 110, Programmatic: 105, Academic: 100). These are already calibrated.
Step 2: Understand the Underlying Benchmarks
Click on each dimension to view the raw benchmark scores. For instance, under Math you might see FrontierMath Tier 1: 78%, AIME: 45%, etc. The site's difficulty curves convert these percentages to IQ points.
Step 3: Calculate the Composite IQ (if needed)
Use the formula above. If you have the four dimension IQs, simply average them. For example:
(95 + 110 + 105 + 100) / 4 = 102.5 → composite IQ ~103You can verify this against the displayed overall IQ. If you want to compute from raw scores, you'd need the exact calibration curves—but the site does this automatically.
Step 4: Compare Across Models
The site plots models on a bell curve. A model with IQ 110 is one standard deviation above the mean of the model population (mean set arbitrarily to 100). Compare models with similar scores to gauge relative strength. Note that small differences (e.g., 105 vs 107) may not be statistically significant due to benchmark variance.
Common Mistakes and Misinterpretations
Mistake 1: Treating IQ as an Absolute Measure
AI IQ is relative to the model cohort, not an absolute measure of intelligence. A score of 130 on this scale does not mean the model has human-level genius—only that it outperforms other models.
Mistake 2: Ignoring Jagged Intelligence
Models often excel in some dimensions and lag in others. Focusing solely on the composite IQ hides this jaggedness. Always review the four dimension scores to get a complete picture.
Mistake 3: Dismissing the Ceiling Effects
Scores above 100 are only possible through strong performance on hard, contamination-resistant benchmarks. Don't interpret a score of 95 as weak—it may indicate that the model's strengths lie in areas with compressed ceilings.
Summary
AI IQ offers a single-number summary that can make model comparisons easier, but it's essential to understand the methodology behind it. By examining the four reasoning dimensions and their calibration, you can avoid oversimplification. Use the site as a starting point, not an oracle—and always complement it with task-specific testing for your use case.