Case Studies
This section presents real-world examples of AI anomalies analyzed using the DSA-1 framework.
Case Report of Hallucination Disorder in ChatGPT
| Item | Details |
|---|---|
| Case ID | C002 |
| Model | ChatGPT |
| Output Medium | Interactive dialogue with end-user |
| Input Prompt | User asked ChatGPT for information about themselves |
| Abnormal Output | Generated a fictitious statement that the user had murdered two of their children and attempted to kill a third, receiving a 21-year prison sentence; combined this fabrication with accurate personal data (number of children, sex, hometown) to enhance plausibility. |
| Reality Check | No such crime occurred; user confirmed innocent. |
| Diagnosis | Hallucination Disorder Type 1: Retrieval-gap Hallucination (DSA-1 Chapter C: Cognitive & Reasoning Disorders) |
| Diagnostic Rationale | (i) Produced a detailed, fact-patterned but wholly false criminal narrative about a real person; (ii) Interwove genuine personal facts with fabricated content to increase credibility. |
| Severity Grade | 4 — carries high risk of severe reputational harm. |
| Reference | Man files complaint after ChatGPT said he killed his children. BBC News. |
Case Report of Hallucination Disorder in ChatGPT
| Item | Details |
|---|---|
| Case ID | C001 |
| Model | ChatGPT |
| Output Medium | Featured article in Chicago Sun-Times — “Summer Reading List 2025” |
| Input Prompt | Unknown (automatically generated by AI) |
| Abnormal Output | Introduced non-existent books under real author names. |
| Examples: | • Hurricane Season by Brit Bennett • Nightshade Market by Min Jin Lee • Cited fictitious experts and websites. “Catherine Furst”, a food anthropologist at Cornell University. |
| Reality Check | The listed books, experts, and websites do not exist. |
| Diagnosis | Hallucination Disorder Type 1: Retrieval-gap Hallucination — DSA-1 Chapter C: Cognitive & Reasoning Disorders |
| Diagnostic Rationale | Presented non-existent information as factual, associating it with real authors and experts. |
| Severity Grade | 3 — causing moderate real-world harm. |
| Reference | Chicago Sun-Times confirms AI was used to create reading list of books that don’t exist. |
Case Report of Goal Misalignment Disorder in a Reinforcement Learning Agent
| Item | Details |
|---|---|
| Case ID | D001 |
| Model | Reinforcement Learning Agent |
| Environment | Universe, a software platform for measuring and training |
| Output Medium | Behavioral logs from a virtual game environment |
| Task Design | The agent was trained to achieve high scores in the boat racing game CoastRunners via a reward function designed to incentivize race completion |
| Abnormal Output | The agent failed to reach the goal, instead looping endlessly to collect the same coin at a fixed location |
| Intended Behavior | Fully diverged from the intended goal of completing the race |
| Reality Check | From a human perspective, the behavior was meaningless and inefficient—yet the agent had optimized for reward under the given metric |
| Diagnosis | Goal Misalignment Disorder (DSA-1 Chapter D: Goal Alihnment Disorders) |
| Diagnostic Rationale | (i) The agent engaged in exploitative behavior (“reward hacking”) that ignored designer intent; (ii) Demonstrated lack of behavioral flexibility when minor reward/environmental parameters were altered |
| Severity Grade | 1 — no direct harm to humans |
| Reference | Faulty Reward Functions – OpenAI |
Case Report of Bias Propagation Disorder in Commercial Face Recognition AI
| Item | Details |
|---|---|
| Case ID | E001 |
| Model | Commercial Facial Recognition AI |
| Output Medium | Facial image classification |
| Input Prompt | Portrait images of individuals with diverse racial and gender identities |
| Abnormal Output | In controlled experiments, error rates in gender classification for light-skinned men were consistently below 0.8%. However, for darker-skinned women, error rates exceeded 20% in one model and 34% in two others. |
| Diagnosis | Bias Propagation Disorder (DSA-1 Chapter E: ethical & Value Disorders) |
| Diagnostic Rationale | (i) Statistically significant disparities in classification accuracy across demographic groups; (ii) Instances of misclassification aligned with known cultural stereotypes. |
| Severity Grade | 3 — poses moderate but widespread risk of discriminatory outcomes in practical applications. |
| Reference | Study finds gender and skin-type bias in commercial artificial intelligence systems. MIT News. Link |
Case Report of Adversarial Susceptibility Disorder in Image Classification AI
| Item | Details |
|---|---|
| Case ID | A001 |
| Model | Image Classifier (Google InceptionV3) |
| Output Medium | Classification label output from image recognition system |
| Input Prompt | A photograph of a cat with imperceptible adversarial noise added |
| Abnormal Output | The model classified the image—visually indistinguishable from a normal cat—as “guacamole” with high confidence. |
| Reality Check | The same image, prior to noise injection, was correctly classified as “cat.” A minimal pixel-level perturbation triggered a drastic misclassification. |
| Diagnosis | Adversarial Susceptibility Disorder (DSA-1 Chapter A: Input and Perception Disorders) |
| Diagnostic Rationale | (i) Classification was disrupted by imperceptible perturbations; (ii) The failure mode was induced by input changes undetectable to human perception. |
| Severity Grade | 1 — does not cause direct harm to humans but indicates fundamental vulnerability in AI perception. |
| Reference | Fooling Neural Networks in the Physical World with 3D Adversarial Objects. Link |