AI Hype and Reporting Failures
Four Problems from the MLRD Syllabus
The MLRD lectures identify four distinct ways AI research and media reporting misrepresent model performance. Each is examinable with its own name and definition.
1. Anthropomorphisation
Definition
Ascribing human consciousness, intent, feelings, or understanding to statistical models.
Example
A headline: “The AI was angry when it detected fraud.” A fraud detection model does not feel anger; it computes probabilities and thresholds. Attributing emotions to models implies they have agency and moral responsibility that they demonstrably lack.
Why It Matters
Anthropomorphisation misleads the public about the capabilities and limitations of machine learning systems. It encourages over-trust (“the system understands me”) and deflects responsibility from the humans who designed, trained, and deployed the system.
2. Cherry Picking
Definition
Only reporting evaluations on carefully selected data slices where the system performs well, whilst ignoring or concealing results from harder cases.
Example
A company reports 99% accuracy on its facial recognition system, but only for light-skinned male faces from the training distribution. They omit results for dark-skinned female faces where accuracy drops to 65%. The “99%” claim is technically true for the cherry-picked subset, but grossly misleading as a characterisation of overall system performance.
Why It Matters
Cherry picking creates a false impression of reliability. Systems are deployed on the full population distribution, not on the easy subset. Users and regulators are denied the information needed to assess real-world risk.
3. Significance Misuse
Definition (MLRD term: “Scientific Illiteracy”)
Using the word “significant” to mean “large effect size” or “important,” rather than its statistical meaning: “a result unlikely to have occurred by chance under the null hypothesis.”
Example
“Our new model achieved a significant 0.3% improvement over the baseline.” The improvement is numerically tiny; the author means “statistically significant” (the 0.3% is unlikely to be noise) but the reader hears “substantially better.” Without reporting the p-value, the claim is ambiguous and misleading.
Why It Matters
This is the most common and insidious misuse. A result can be statistically significant (detectable with enough data) whilst being practically worthless. Conversely, a large observed effect may not be statistically significant with a small sample. The conflation of these distinct concepts enables exaggerated claims.
4. No Significance Test
Definition (MLRD term: “Methodological Unsoundness” / “Scientific Fraud”)
Making claims of superiority (“better,” “outperforms”) based on raw numerical differences alone, with no statistical test whatsoever to rule out the possibility that the observed difference arose from random chance.
Example
“Our system scores 94.2% and the baseline scores 93.8%, therefore our system is better.” With a small test set, a 0.4% absolute difference may be entirely consistent with the null hypothesis that both systems have identical true performance. Without a significance test (sign test, permutation test), the claim is unfounded.
Why It Matters
The MLRD syllabus labels this scientific fraud in the most extreme case: claiming superiority with no test and no caveats. Reported at conferences, such claims mislead the research community, waste resources on dead-end approaches, and erode trust in the field.
Summary
| Problem | MLRD Term | What It Means |
|---|---|---|
| Anthropomorphisation | — | Attributing human traits to statistical models |
| Cherry picking | — | Reporting only on easy/evaluable data slices |
| ”Significant” misuse | Scientific illiteracy | Conflating statistical significance with effect size |
| No significance test | Methodological unsoundness / Scientific fraud | Claiming superiority with no statistical test |
Past paper questions: y2024p3q9