How Many AI Outputs to Review? Sampling for Error Rates

Q: How many AI outputs should a human review to estimate the error rate?

It depends on the precision you need, not on traffic. About 385 randomly chosen outputs give a 95% confidence interval of plus or minus 5 percentage points on any error rate, about 1,068 give plus or minus 3 points and about 2,401 give plus or minus 2. Decide the margin first, then compute the sample.

Q: Is reviewing 1% of AI traffic a good rule?

No. A fixed percentage is wasteful at high volume and useless at low volume. At 4 million outputs a month, 1% is 40,000 reviews, far more than a headline rate needs. At 2,000 outputs a month it is 20 reviews, and zero errors in 20 still leaves a 95% upper bound of about 16% on the error rate.

Q: Why can't I use the human review queue to measure error rate?

The review queue is selected on the thing you are measuring. Escalation rules send hard, low-confidence or flagged outputs, so the queue overstates errors on typical traffic and never sees confident mistakes that went straight to users. Its rate also shifts whenever the threshold changes. Measure with a separate random audit sample over all outputs.

Q: How many clean samples do I need to claim a failure rate is below 1%?

Roughly 300. The rule of three says that if you see zero failures in n independent samples, the 95% upper bound on the failure rate is about 3 divided by n. Zero in 100 supports only a claim below about 3%, zero in 300 below about 1%, and zero in 3,000 below about 0.1%.

Q: Which confidence interval should I use for AI error rates?

Use the Wilson score interval or the Agresti-Coull interval rather than the textbook plus-or-minus 1.96 standard errors. The simple Wald interval misbehaves when error rates are near zero and samples are small, and can even go negative. For example, 2 errors in 200 samples gives a Wilson 95% interval of roughly 0.3% to 3.6%.

Q: What is stratified sampling for AI review?

Stratified sampling divides outputs into slices that differ in risk or expected error rate, such as intent, language, customer tier or model version, and gives each slice its own sample size. High-risk slices get tighter margins and more reviews. The overall error rate is then computed by weighting each slice's rate by its share of traffic.

Q: Does the EU AI Act say how many outputs to review?

No. Article 14 requires high-risk AI systems to support effective human oversight, including monitoring operation and detecting anomalies, but it does not set a sample size. A defensible approach is to measure each risk slice's error rate from a blind random sample, report it with an interval, and gate releases on it, so auditors can follow the evidence.

Q: When should every AI output be reviewed instead of sampled?

When volume is low and each output is high-stakes. If a system produces a few hundred consequential outputs a month, full review is affordable and sampling only adds risk. Sampling is for when full review is impossible. It also cannot find failures rarer than the sample can see, so pair it with adversarial testing and complaint mining for rare harms.

Satyam Kumar

Back to Blog

architecture

How Many AI Outputs Should a Human Review? Sampling Statistics for Error Rates (2026)

By Satyam KumarNovember 5, 202613 min read

human review sampling ai error rate sample size calculation confidence interval wilson score interval rule of three stratified sampling audit sampling human in the loop ai quality assurance llm evaluation eu ai act article 14 ai governance review queue bias inter-rater agreement golden dataset

How Many AI Outputs Should a Human Review? Sampling Statistics for Error Rates (2026)

Frequently Asked Questions

Share this article

Twitter LinkedIn WhatsApp

Satyam Kumar

Founder & AI Architect, AppScale LLP

AI & Cloud Architect. Helping teams build systems that scale to millions.

LinkedIn GitHub

How Many AI Outputs Should a Human Review? Sampling Statistics for Error Rates (2026)

Frequently Asked Questions

Share this article

Comments

Leave a comment