Learn AI /Model Quality and Evaluation /Baseline
Baseline
- Intermediate
- Model Quality and Evaluation
- Programme & data teams
- Technical readers
Back to the milestone you came from
In five seconds
What it is
A reference method or result used to judge how a model performs against a relevant starting point.
In plain words
A baseline is the reference result you compare against before calling a model better.
An everyday example
A team in New York needs to find policy documents about flood preparedness in a large shared folder. First, one colleague tries the simplest method: search for the exact phrase “flood preparedness” and open the first ten results. They record that 4 of the 10 documents are actually useful. Then another colleague uses a better search tool that expands related terms such as “early warning” and “risk reduction” and finds 7 useful documents out of 10. The first result gives the team a starting point, so they can tell whether the new tool really improves the search.
Real model baselines can be more varied than one simple manual search. Some baselines are statistical rules, historical averages or previously deployed systems, not just a weaker version of the same tool. Also, a real evaluation may use several metrics at once, while the story uses one easy count of useful documents.
Remember thisA baseline makes a model score meaningful, provided both are tested on the same task, data split and metric.
Understand it
Understand it
Why it matters
Without a baseline, people can be impressed by numbers that do not prove real usefulness. A high score may simply reflect an easy dataset, an imbalanced class distribution or a weak comparison. Baselines help non-specialists ask better questions: Is this system better than a simple rule? Better than the current process? Better in the ways that matter for risk, cost and fairness? This supports safer procurement, clearer reporting and more reliable decisions about whether to deploy, revise or reject a system.
Baseline versus new model
Show that a model score only becomes meaningful when compared with a simple, relevant reference result.
01
Same task
Start with one clearly defined task, one evaluation dataset and one scoring rule so the comparison is fair.
At the left, draw one shared input box labelled 'Same task + same test data + same metric'. Split one arrow into two equal branches.
02
Simple reference
On the top branch, run a very simple method such as predicting the most common class or using exact keywords.
Top branch box labelled 'Baseline method'. Add small note text: 'simple, cheap, reasonable'.
03
New system
On the bottom branch, run the model or tool you want to assess.
Bottom branch box labelled 'New model or system'. Add small note text: 'more complex approach'.
04
Compare scores
Bring the two outputs together and compare their scores, not just the new score by itself.
Merge both branches into one comparison box showing example numbers such as 'Baseline: 80%' and 'New: 82%'.
05
Interpret carefully
Ask whether the improvement is meaningful for the real use case, given errors, costs and risks.
Final box on the right labelled 'Useful improvement?'. Under it, show three short checks: 'better enough?', 'safer?', 'worth the cost?'.
Two side-by-side paths compare a simple baseline with a new model on the same task and data, then bring both scores together for interpretation. Read left to right: same input, different methods, then a decision about whether the improvement is large and useful enough to matter.
Hovering over the comparison box reveals alternative baseline examples such as 'majority class', 'random guess', 'keyword search' and 'previous system'.
Describe this without the diagram
The diagram compares two approaches using the same task, data and metric. A shared starting box splits into two branches. The top branch is a simple baseline method, described as simple, cheap and reasonable. The bottom branch is the new model or system. Both branches produce scores that meet in one comparison box, with example values Baseline 80% and New 82%. A final box asks whether the improvement is useful, with checks for whether it is large enough, safer and worth the cost. The main lesson is that a model score means little without a relevant reference point.
With reduced motion: A static side-by-side comparison shows one shared input feeding a baseline branch and a new model branch, then a box comparing their scores and a final box asking whether the difference is useful in practice.
A baseline is a reference method or score that gives you something to compare against. It tells you, “With this reference approach, this is the result.” Then you compare the new system against that result.
What goes in is a task, some data and a rule for scoring performance. The rule for scoring is an evaluation metric, such as accuracy or recall. You also choose a relevant reference method, for example predicting the most common class, using keyword search, repeating the previous value or using the current process.
What happens is that you run this reference method on the same task that the new model will face. What comes out is a reference score, such as 80% accuracy, plus a clearer sense of how hard the task may be.
This matters because a score can look impressive without meaning much. If 80% of documents are routine, a system that always says “routine” gets 80% accuracy without finding anything unusual. A new model scoring 82% may sound better, but it may add little practical value.
Baseline does not mean benchmark. A benchmark is a standard test set-up used to compare systems on a defined task under stated conditions. A baseline is a reference method or result used within an evaluation. A benchmark can report several baselines.
How “Checking whether search helps at all” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Searching for the exact phrase only | A simple baseline method |
| 4 useful documents out of 10 | The baseline score on an evaluation metric |
| The improved search tool | The more complex model or system being tested |
| Comparing 4 out of 10 with 7 out of 10 | Judging whether the new system adds meaningful value |
A worked example
A desk officer in a regional bureau is finalising the evaluation annex for a document-routing pilot. The classifier assigns 300 public programme summaries to education, health or climate queues.
Starting from: Test set of 300 summaries: 180 health, 90 education and 30 climate.
- Set the baseline to always predict the most frequent class, health.
- Score that baseline on the fixed test set using the evaluation metric selected for the pilot.
- Run the candidate classifier on the same test set and score it with the same metric.
- Compare overall accuracy and class-level recall, particularly for the less frequent education and climate categories.
Result: The majority-class baseline is correct for 180 of 300 summaries, giving 60% accuracy. The classifier reaches 72%, an improvement of 12 percentage points. That gain is not sufficient on its own: if climate recall remains poor, the classifier may still be unsuitable for routing work.
In a United Nations workplace
A legal officer in an inter-agency secretariat is drafting a lessons-learned note after a pilot that grouped comments on draft guidance by legal issue. Before the findings go to the front office, the officer compares the model with the keyword rules previously used for first-pass sorting.
The keyword rules are a useful baseline only if they and the model are assessed on the same labelled comments, with the same categories and metric. The comparison shows whether the pilot improved on the existing workflow rather than merely producing a respectable-looking score.
Before you try thisComments on draft guidance may contain confidential or legally sensitive material. They must not be entered into a tool that has not been approved for that material. Use an approved environment, minimise the text retained and keep legal review in the workflow.
Where you may hear it
- A lessons-learned meeting after an AI pilot
- A procurement meeting while a vendor presents evaluation results
- An ICT governance board reviewing a proposed deployment
- Clearance of a model evaluation annex
Avoid confusion
Avoid confusion
Often confused with
Benchmark
A benchmark is a standard test set-up used to compare how well systems perform on a defined task under stated conditions.
Accuracy
The share of cases a system labels correctly out of all the cases it was tested on.
Evaluation
Evaluation checks how well a system performs on a defined task, while metrics are the specific measures and evals are the tests used to produce them.
Common misconceptions
Common misconception: A baseline must be very weak, such as random guessing.
Correction: Not always. A useful baseline should be simple but relevant. For some tasks, random guessing is too weak to teach much. A stronger baseline, such as the current manual process or a straightforward keyword rule, can give a more honest test of whether the new system is worth adopting.
Common misconception: If a model beats the baseline, it is ready for use.
Correction: Beating a baseline only shows improvement over that reference point. The model may still be inaccurate, unreliable on rare cases, badly calibrated or unsafe for the setting. You still need to check error types, robustness, human oversight needs and whether the performance is good enough for the real consequences of mistakes.
Common misconception: Baseline and benchmark mean the same thing.
Correction: They are related but not identical. A baseline is the reference result from a simple method or existing process. A benchmark is often the standard dataset or test suite used for comparison. A single benchmark can include several baselines.
Quick check
Why is a baseline useful when evaluating a classifier?
A baseline shows whether a model improves on a relevant reference method. It is not a required performance threshold, a standard test set-up or the accuracy metric itself.
Go deeper
Go deeper
The simple version
A baseline is the reference result you compare against before calling a model better.
The technical version
In model evaluation, a baseline is a reference method or performance level assessed under the same task definition, data split and evaluation metric as the candidate model. Common baselines include majority-class prediction, random prediction with known class frequencies, linear or rule-based methods, historical averages, previous production systems and human performance under defined conditions. The purpose is to compare the model with trivial patterns or an existing workflow. A baseline can reveal that an apparently high score offers little improvement over an obvious strategy, but it does not by itself diagnose problems such as data leakage. Baseline selection depends on context. For time series, a last-value or seasonal baseline may be appropriate. For retrieval, keyword search may be relevant. For classification, majority-class or simple linear models are common. A strong evaluation may report multiple baselines because one rarely captures every practical comparison.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This official framework helps readers place baseline comparisons within broader risk management and governance, and suits managers, evaluators and technical leads.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Baseline (opens in a new tab)Wikipedia
This Wikipedia search results page helps readers scan different possible meanings of 'baseline' and suits someone who needs quick orientation before deeper reading.
search results, not one pagelink checked 2026-07-31
CourseClassification: Accuracy, recall, precision (opens in a new tab)Google
This course resource helps readers see how metrics such as accuracy, recall and precision affect evaluation, and suits beginners connecting baselines to scoring.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Baseline (opens in a new tab)YouTube search
This YouTube search results page helps visual learners find multiple video explanations of baseline-related ideas, and suits readers who prefer spoken walkthroughs.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Baseline (opens in a new tab)Reddit search
This Reddit search results page helps readers see common questions and informal explanations about baselines, and suits readers who want practical discussion with caution about reliability.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A baseline makes a model score meaningful, provided both are tested on the same task, data split and metric.