Controversial AI Debate Topics, Each Tested Against Evidence

Two AI debate motions with evidence on both sides: algorithmic risk scores in courts and AI-text detectors in schools, plus a format for running the round.

Guide: AI BasicsLast verified: October 20265 min read

The best AI debate topics are those where measured evidence exists on both sides. Two examples fit: whether courts should use algorithmic risk scores, where a 2018 Science Advances study found the COMPAS tool about 65 percent accurate against 67 percent for pooled online volunteers, and whether schools should use AI-text detectors, where a 2023 Stanford study found they labeled 61.3 percent of essays by non-native English writers as machine-written. A motion backed by numbers like these produces a better round than a slogan about robots.

Key takeaways

  • A strong motion names one decision, one tool, and one group affected, so both teams argue from the same facts.
  • COMPAS predicted reoffending for 7,214 Florida defendants with accuracy near that of untrained volunteers, according to Dressel and Farid.
  • In State v. Loomis (Wisconsin, 2016), the court allowed risk scores at sentencing but required written warnings and barred using them to decide incarceration or sentence severity.
  • Seven AI-text detectors flagged non-native writing as AI far more often than native writing, and the authors advised against using them in evaluative or educational settings.
  • Each side should bring the strongest opposing study and answer it, not avoid it.

What makes a motion debatable

Broad prompts such as "AI is dangerous" or "AI will replace humans" cannot be won, because no judge can say what evidence would settle them. A debatable motion has three parts: an actor (a court, a school, a hospital), a tool (a named category of system), and a policy action ("should require," "should ban," "should allow only with"). A motion written this way lets teams cite studies about the same tool and lets a judge decide whether the evidence supports the policy.

The two cases below were chosen because each has a peer-reviewed study and a rule or recommendation that can be read in full.

Motion one: courts should use algorithmic risk scores

COMPAS is a commercial tool that scores the likelihood of reoffending from 137 features. Julia Dressel and Hany Farid of Dartmouth College compared it with human judgment on 7,214 pretrial defendants from Broward County, Florida. Their paper, published in Science Advances in January 2018, reports COMPAS accuracy of about 65 percent. Online volunteers with no criminal justice training, shown only seven features for each of 1,000 defendants, reached about 62 percent individually and 67 percent when their answers were pooled by majority. A linear predictor using only age and number of previous convictions performed nearly as well as COMPAS.

Supporters of the motion can argue that a consistent score, used as one input, is easier to audit than a judge's unrecorded intuition. Opponents can argue that a costly proprietary tool adds little over two simple facts and cannot be inspected.

The Wisconsin Supreme Court addressed the legal question in State v. Loomis, 881 N.W.2d 749 (2016). It held that using COMPAS at sentencing did not violate due process, but it required presentence reports that include a score to carry five written warnings: the method is proprietary, scores rest on group data and cannot identify a specific high-risk person, no study had validated the tool on a Wisconsin population, studies have raised questions about whether minority offenders are rated higher, and the tool was built for post-sentencing corrections decisions. The court also barred using scores to decide whether a person is incarcerated or how severe a sentence should be. This article is general information, not legal advice.

Motion two: schools should use AI-text detectors

When chatbots began writing essays, detection software appeared as the obvious answer. Weixin Liang, James Zou, and colleagues tested seven detectors on 91 TOEFL essays written by non-native English speakers and on 88 essays by US eighth-graders. The detectors were near-perfect on the native writers' essays, yet they wrongly labeled the TOEFL essays as AI-generated at an average rate of 61.3 percent, and all seven agreed in flagging 19.8 percent of them. The paper, published in the journal Patterns on July 10, 2023, also reports that simple prompting can lower the bias and evade the detectors, and it advises against using them in evaluative or educational settings.

Supporters can respond that detectors are meant as a prompt for conversation, not proof, and that schools need some check on unattributed work. Opponents can respond that a prompt for conversation still lands unevenly on students who write English as an additional language, and that a determined cheater can evade detection.

Evidence map for both motions

MotionStrongest evidence forStrongest evidence againstOpen question
Courts should use risk scoresScores are consistent and recorded; Loomis allows use with warningsAbout 65 percent accuracy, near that of volunteers; method not disclosedDo scores change outcomes for defendants once judges see them?
Schools should use AI-text detectorsNear-perfect accuracy on native eighth-grade essays in one test61.3 percent average false positives on non-native essaysDo newer detectors show the same bias?

The last column matters for judges and moderators. Both studies are snapshots: one dataset from Florida, one set of detectors from 2023. A fair round states what the study cannot show.

Running the round

  • Give each team the same evidence packet, and let each add two studies of its own.
  • Require every claim to name its source, its sample size, and its year.
  • Ask each team to state the best fact against its own side and answer it in its opening.
  • Score the round on accuracy of evidence and quality of rebuttal, not on volume.
  • Close with a short discussion of what new data would change each team's position.

FAQ

What is a good first AI debate topic for a classroom?

A narrow, local one such as the AI-text detector motion works well, because students know the setting and the study is short enough to read together.

Does the COMPAS study prove risk scores should be banned?

No. It shows that one tool's accuracy was close to that of non-experts on one dataset. Whether to use such tools is a policy question that also involves fairness, transparency, and the alternatives.

Did the Wisconsin court approve algorithmic sentencing?

Only in a limited way. The 2016 decision allowed a score as one input with warnings and forbade using it to decide incarceration or severity of the sentence.

Are AI-text detectors accurate?

Accuracy depends on the writer and the detector. The 2023 study found near-perfect results on native eighth-grade essays and a 61.3 percent average false-positive rate on non-native essays.

Last verified: October 2026

Sources

  • The accuracy, fairness, and limits of predicting recidivism, Dressel and Farid, Science Advances, 2018 (free archive copy)
  • GPT detectors are biased against non-native English writers, Liang et al., Patterns, 2023 (free archive copy)
  • State v. Loomis, 881 N.W.2d 749, Wisconsin Supreme Court, 2016, as summarized in Harvard Law Review, vol. 130, 2017

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics