How data affects AI performance: the amount versus the quality

Data sets the ceiling on AI performance. More data usually helps, but each extra batch adds less than the last, and only if the data is accurate, complete, current and representative of real cases. Past a certain point, cleaning data improves results more than adding it, because errors in the data become errors in the output.

This guide explains how data affects AI performance in three ways (quantity, quality and fit), what the research says about each, the data problems that most often hurt AI in Saudi businesses, and a checklist to run before you train, tune or feed an AI system. Examples are illustrative unless a source is cited.

How data affects AI performance: quantity, quality and fit

SDAIA's AI Ethics Principles put it plainly: the quantity and quality of datasets should be sufficient and accurate to serve the purpose of the system, and the sample size collected has a significant impact on the accuracy and fairness of a trained model's outputs [1]. The US National Institute of Standards and Technology (NIST) lists data among the risks that make AI different from ordinary software: the data may not truly represent the context in which the system is used, quality issues can undermine its trustworthiness, and datasets can become stale relative to where the system is deployed [2].

Three questions follow from that:

When an AI system disappoints, check these three before you blame the model.

How much data does AI need? What the scaling research shows

For the large language models behind chat assistants, the relationship between data and performance has been measured. A 2020 study found that a model's error (its "loss") falls as a power law with model size, dataset size and the computing power used for training [3]. In plain terms, gains are steady and predictable, but each doubling of data buys a smaller improvement than the last.

A 2022 study trained more than 400 language models, from 70 million to over 16 billion parameters, on 5 to 500 billion tokens. It concluded that model size and the number of training tokens should be scaled equally, and that many large models had been trained on too little data. Its model Chinchilla, with 70 billion parameters and about four times as much data as the 280-billion-parameter Gopher, built with the same compute budget, outperformed Gopher and several larger models across a wide range of tasks [4].

The volumes involved are enormous. SDAIA's National Center for AI says the 7-billion-parameter version of its Arabic and English model ALLaM was trained on 5.2 trillion tokens: 4 trillion in English, then 1.2 trillion of mixed Arabic and English [5]. A token is a short piece of text, often part of a word.

A company rarely trains a model of that kind. It usually adapts an existing model with its own examples, or gives the model its documents to read when answering. The data you need is then far smaller, but it carries more weight: a few hundred wrong or missing examples can decide whether the system works for your customers.

Why quality beats volume past a certain point

Three findings show that cleaner data can matter more than more data.

• Curated data can outperform scale. Researchers trained phi-1, a 1.3-billion-parameter coding model, on about 6 billion tokens of "textbook quality" web data plus about 1 billion tokens of synthetic textbooks and exercises. It reached 50.6% on the HumanEval coding test, and the authors attribute its strength against much larger models to the quality of its data [6].
• Label errors mislead you about performance. A study of 10 widely used test sets estimated an average of at least 3.3% label errors, and at least 6% of the ImageNet validation set. With corrected labels, a smaller model beat a larger one once the share of originally mislabelled test examples rose by just 6% [7]. In other words, errors in your test data can lead you to pick the wrong model.
• Data problems compound. Google researchers who interviewed 53 AI practitioners described "data cascades", compounding events that cause negative downstream effects from data issues, and reported a 92% prevalence among them [8].

Seven data problems that hurt AI performance

SDAIA's AI Bias Reference Guide, first published in May 2026, defines about a hundred ways in which data, models and the people who use them introduce bias or error [9]. The table pairs seven of the most common with what they look like in a Saudi company.

Missing values and duplicate records belong on the list too. They are not bias in SDAIA's sense, but they distort what a model learns in the same quiet way.

Arabic data: why the first gap is volume

Arabic starts with less data than English. In the latest crawl listed by Common Crawl in October 2026 (CC-MAIN-2026-39), a public archive of web pages widely used to train language models, English made up 41.86% of pages and Arabic 0.67% [10], roughly 63 English pages for every Arabic one. ALLaM's training mix shows the same imbalance: 4 trillion English tokens before 1.2 trillion of mixed Arabic and English [5].

The gap is wider still for Saudi dialects, scanned documents and sector terms. SDAIA's guide gives the classic case: a review model trained only on formal text fails to understand comments written in dialect, which marginalises the way many users actually write [9]. For a Saudi company, the practical lesson is to test any AI system on your own Arabic material, not on a vendor's demo, and to treat your past emails, tickets and forms as data worth cleaning and keeping. If your records are scanned PDFs, our guide to Arabic OCR explains how to turn them into data you can trust first.

When more data makes things worse: synthetic data and model collapse

As more web content is written by AI, more of it ends up in training sets. Researchers found that indiscriminately training models on content generated by other models causes irreversible defects, in which the tails of the original distribution, the rare but real cases, disappear [11]. More data, in that case, means a narrower and less accurate model.

The same logic applies inside a company. AI-generated summaries, draft replies or synthetic records can help, but they are not a substitute for real records. Keep the originals, mark which records were generated by AI, and never let generated text quietly become the "ground truth" a future model learns from.

Accurate data is also a legal requirement in Saudi Arabia

When personal data is involved, quality is not only good practice. Article 14 of the Personal Data Protection Law bars a controller from processing personal data without taking sufficient steps to verify its accuracy, completeness, timeliness and relevance to the purpose for which it was collected [12]. Article 11 limits the content of personal data to the minimum needed to achieve the purpose of its collection [12].

SDAIA's ethics principles add process controls: datasets should be approved and signed off before model development starts, and the data preparation process and quality checks should be documented and validated by the responsible parties [1]. This section is general information; for a specific case, check SDAIA's guidance or ask a licensed advisor.

A 10-point data checklist before you use AI

Run this on any dataset before it trains, tunes or feeds an AI system.

1. Name the decision. Write down what the AI will decide or produce, and the cases it must handle.
2. Count every case. Check that each category, branch, dialect, product line and season the AI will see appears in the data, not just the common ones.
3. Test the labels. Have two people label the same sample of records and compare; disagreements show where rules are missing.
4. Remove duplicates and test records, including demo entries and cancelled transactions that were never real.
5. Check the dates. Note the time range, and flag records from before any change in prices, fees, rules or processes.
6. Separate training and test data, by date where possible, and check that no input field gives away the answer.
7. Keep a hold-out set of recent real cases that the AI never sees during training, and judge results only on those.
8. Mark AI-generated records so they never pass as original data.
9. Minimise and verify personal data, in line with the Personal Data Protection Law.
10. Monitor after launch. Re-test on fresh cases on a fixed schedule, and retrain or adjust when results slip.

Your next step

Pick one dataset you would like AI to work on, such as customer messages, supplier invoices or service tickets, and run the first five checks on a sample of a few hundred records this week. The gaps you find will tell you more about how the AI will perform than any vendor benchmark. When you are ready to test AI on that task, our guide on how AI can help your business shows how to run a short, measured trial. You can read about O AI, a software and AI company in Al Khobar, on our home page.