How to Write Original Examples for Historical AI Research

A paper reports 84 percent accuracy for a speech system. Accuracy at what? The figure means something different if it counts correctly transcribed words, correctly labeled recordings, or whole utterances. Consider an invented test of 100 short recordings: if the system assigns the correct language to 84, its recording-level language-identification accuracy is 84 percent. It may not have transcribed a single word. A small example makes the denominator visible without changing the paper’s reported result.

Original examples are purpose-built cases, not reconstructions of historical experiments. They help when accounts of early pattern recognition use familiar terms such as accuracy, training, and tracking for quite different tasks. The trick is to make the case easy to inspect while keeping its invented details separate from the research record.

Start with the distinction the reader needs

First, decide what misunderstanding the example should prevent. “Explain speech recognition” is too broad; “show why identifying a language is not the same as transcribing speech” is workable. Give two systems the same five-second recording. One returns “Language A”; the other returns a sequence of words. They may share audio-processing components, but their outputs answer different questions.

Change one important feature at a time. If the recording quality, speaker, vocabulary, and evaluation measure all change, readers cannot tell which difference matters. Keeping most details fixed creates a controlled comparison, though it is not an actual experiment.

A useful drafting prompt is: “The reader should understand why X does not imply Y.” A high score on isolated spoken digits, for instance, does not imply equally good recognition of conversational speech. Keep the language and speaker the same, then replace separate digits with pauses by an uninterrupted sentence. The added problem of locating word boundaries becomes clear without making either task sound trivial.

Choose a case small enough to verify mentally

Use numbers and inputs readers can check for themselves. A document classifier with three labels—letter, invoice, and form—is easier to inspect than one with hundreds. Two moving objects are usually enough to show why maintaining identity across frames differs from detecting objects anew in each frame.

Small need not mean spotless. If every scan is legible and every class equally represented, the example can give a false impression of research data. Add a complication when it earns its place: one blurred scan can illustrate uncertainty in document recognition; one brief occlusion can explain a tracking error. Say what the complication shows rather than implying every historical system failed under those conditions.

Two moving objects traced across successive frames

Build the example around input, operation, and output

Spell out what the system receives, what it does, and what counts as a result. In stereo reconstruction, two cameras observe the same point from slightly different positions. The system finds corresponding locations in the images; their displacement, together with camera geometry, helps estimate depth. Saying “two photos make a 3D image” skips the correspondence problem—often the difficult part.

For an original example, imagine a marked corner of a cardboard box at horizontal position 120 in the left image and 108 in the right. The 12-pixel difference illustrates displacement, not depth on its own. Without calibration and defined image coordinates, it cannot give a trustworthy distance.

Be just as clear about the output’s limits. Detecting a face in a frame does not establish whose face it is. Finding dark regions in a medical image is not a diagnosis. Describe the result in plain words, then check that your example has not quietly given the system an extra capability.

Show the intermediate step when it carries the idea

“Audio goes in, a language label comes out” may skip the very part readers need. An example can say that a language-identification system compares patterns extracted from a short recording with patterns learned from labeled recordings. There is no need to name a feature extractor or model architecture unless the source does.

When an intermediate representation matters, show a minimal version. A document-processing system might divide a scanned page into text regions before recognizing characters. If it reads a heading and two columns in the wrong order, correctly recognized words can still produce a jumbled passage. That separates layout analysis from character recognition; “the computer read the page” obscures both.

Mark invention without interrupting the explanation

Place an original example beside the evidence, not inside it. “Consider a hypothetical test set” or “For illustration, suppose…” signals that numbers and outcomes are invented. Say it once when the scenario begins, then avoid wording that makes the case sound like a finding from the paper.

This matters in historical writing. If a mid-2000s conference paper describes work on a low-resource language, an example involving ten hours of recordings might clarify data scarcity. It must not suggest the project had ten hours unless the source says so. Where the amount is unknown, contrast a system trained on many labeled recordings with one trained on few, or leave the number out.

Precise-looking improvements need the same care. “The revised method reduced errors from 30 to 12” sounds like a reported result even if it was invented to teach a point. To illustrate a rate, say, “In a hypothetical batch of ten scans, three are classified incorrectly.” To describe what a real system achieved, use its documented result.

Match the form of the example to the difficulty

Not every concept needs a story. Use a paired comparison for a distinction, a short sequence for a process, or a counterexample for an overgeneral claim. Pick the simplest form that lets readers test the idea.

  • Paired cases: Keep the input fixed and change the task. Assigning a language label to a recording is different from transcribing its words.
  • Step-by-step cases: Follow a scanned page through region detection, character recognition, and reconstruction of reading order.
  • Boundary cases: Track two people until they cross and briefly obscure one another. Detecting a person in the next frame does not guarantee the correct identity was assigned.
  • Counterexamples: Show how a classifier can score well overall by favoring the common class while missing most rare cases.

For that last case, suppose a hypothetical medical-image set has 90 images labeled negative and 10 labeled positive. A system that calls every image negative gets 90 of 100 correct but finds none of the positives. The arithmetic shows why overall accuracy cannot substitute for sensitivity. It makes no claim about any particular medical system.

Handwritten counts compare correct labels and missed positives

Keep analogies subordinate to the technical case

Comparing object tracking to following a runner through a crowd introduces the identity problem, but it leaves out how a system represents appearance, handles occlusion, or decides a track has been lost. Follow the analogy with an image-sequence example: two similar-looking objects cross, one disappears behind a post, and its next visible position could belong to either. The ambiguity is now visible in the task itself.

Watch for analogies that imply human understanding. A recognizer may classify acoustic patterns rather than “listen for meaning.” A tracker’s state may be called “memory” for convenience, but it does not remember an event as a person does. When the analogy starts implying a capability, return to the inputs and outputs.

Revise for accuracy and usefulness

Once the example is drafted, compare it with both the concept and the source. A vivid detail earns its place only if it helps the intended distinction hold up.

  1. State the point. Write one sentence describing what readers should be able to distinguish afterward.
  2. Audit the variables. Check that a comparison does not accidentally change the input, task, or measure.
  3. Verify the mechanics. Recalculate toy numbers and check technical terms, even in a hypothetical case.
  4. Separate fact from illustration. Mark invented inputs, datasets, and results; do not attribute them to a historical project.
  5. Trim decorative details. Remove dates, places, device models, or participant descriptions that add nothing and might look sourced.

An example can be memorable and still answer the wrong question. If the paragraph is meant to distinguish precision from recall, saying a model became “more accurate” will not help. Give counts of correct positive predictions and missed positives, specify each measure’s denominator, and let readers see the difference.

Let the example expose the limit of the claim

A good example shows what remains unknown as well as what a result means. Imagine a hypothetical speech test with 20 clean recordings from speakers represented in training and five noisy recordings from unfamiliar speakers. Success on the first group alone says nothing definite about performance on the second. That makes the untested condition visible without predicting failure.

As a final check, replace vague success claims with observable events. In a language-identification example, “worked well” could become “assigned the correct language label to eight of ten recordings.” If relevant, add that the recordings were clean, short, and drawn from the languages included in the test. Readers can then see exactly what the example establishes—and which recordings remain outside its claim.