Can ChatGPT Grade Essays Accurately for the HKDSE?

If you feed your practice past paper into an uncalibrated prompt, the short answer is no: ChatGPT cannot reliably grade HKDSE essays out of the box. While it quickly spots comma splices and suggests elevated vocabulary, standard Large Language Models (LLMs) consistently suffer from severe mark inflation, award phantom points, and struggle to apply the strict marking schemes published by the Hong Kong Examinations and Assessment Authority (HKEAA).

Surveys by Save My Exams reveal that roughly 93% of secondary school students now use generative AI for revision. Yet randomized control trial data cited by Third Space Learning underscores a sobering reality: unguided revision with generic LLMs can lead to approximately 17% lower exam outcomes compared to structured revision. With Form 6 students across Hong Kong entering intense winter mock preparation, relying on raw ChatGPT feedback can build false confidence, leaving candidates shocked when mock examiners award a Level 3 to an essay ChatGPT rated as a Level 5.

The Test: Putting ChatGPT Against Official HKEAA Mark Schemes

To see how generic AI handles real local standards, we ran a series of past paper submissions through free-tier and default LLMs. We tested essays across two primary HKDSE formats:

    1. HKDSE English Language Paper 2 (Writing): Both Part A (compulsory short task) and Part B (extended elective task) responses evaluated across Content (7 marks), Language (7 marks), and Organization (7 marks).

    2. Humanities Extended-Response Questions: Historical and economic evaluative essays from HKDSE History Paper 1/2 and Economics Paper 2, testing causation, data synthesis, and counter-arguments.

We compared the AI-generated scores against verified HKEAA sample scripts graded by seasoned markers. The results exposed three recurring failure modes that every DSE candidate must watch out for.

Why ChatGPT Misjudges DSE Scripts: 3 Critical Failure Modes

1. Mark Inflation and the "Polite Examiner" Bias

Base AI models are fine-tuned via Reinforcement Learning from Human Feedback (RLHF) to be helpful, encouraging, and diplomatic. When asked, "Grade this HKDSE English Paper 2 Part B essay," ChatGPT instinctively praises superficial eloquence. It routinely awards 6/7 or 7/7 for Language simply because the prose flows smoothly and contains ambitious vocabulary like "plethora", "quintessential", or "myriad".

However, HKEAA markers penalize unnatural collocations, register inconsistencies, and formulaic "canned" template sentences that do not fit the specific audience or genre (e.g., a formal letter to the editor versus a school magazine article). In our tests, essays demonstrating severe register clash still received top-band marks from generic AI, inflating predicted scores by as much as two whole levels.

2. Phantom Content and Contextualization Points

In HKDSE humanities subjects like History, Economics, and BAFS, candidates earn marks only when specific syllabus-aligned concepts and historical facts directly answer the prompt. When grading an 8-mark or 12-mark essay, ChatGPT frequently hallucinates points by rewarding rhetorical sophistication instead of factual precision.

For instance, on an economic question regarding tariff policies, ChatGPT awarded full marks to an answer containing a vague explanation of price mechanisms without referencing deadweight loss diagrams or statutory trade barriers. Because the explanation sounded cohesive, the AI assumed the content criteria were met. In a real HKEAA marking room, that response would be capped at Level 2 for lacking concrete syllabus knowledge.

3. Stimulus and Data Misinterpretation

Many HKDSE extended writing questions rely heavily on data files, scenario prompts, or bulleted guidelines. In English Paper 2 Part A, straying from the situational context—such as misidentifying the persona, the school club, or the target demographic—triggers immediate point deductions in Content.

Generic LLMs experience prompt drift: as they evaluate a 500-word essay, they lose track of the strict constraints given in the original prompt. They grade the essay in isolation as a generic piece of creative writing rather than measuring how thoroughly every required bullet point was addressed.

The 3-Step Rubric Calibration Protocol

You do not need to abandon AI entirely during your DSE revision. When you pair structured frameworks with effective AI-powered study methods, AI can become a sharp proofreader. To prevent mark inflation, use this 3-step prompt calibration protocol before submitting your next draft:

Step 1: Feed the Exact HKEAA Band Descriptors

Never ask AI to simply "grade my essay." Instead, copy and paste the exact Level 1 to Level 5 band descriptors directly from the HKEAA subject examination report. Force the model to adopt the persona of a senior DSE marker.

Example Prompt Directive:
"Act as a strict HKEAA senior marker for English Language Paper 2. Do not praise my work. Grade this response strictly using the 3 domain criteria: Content (max 7), Language (max 7), and Organization (max 7). Penalize overused idioms and incorrect tone."

Step 2: Enforce Evidence-Based Mark Justification

Stop the AI from providing a single aggregate score. Instruct the model that for every mark awarded or deducted, it must quote exact lines from your essay and cite the corresponding criterion from the rubric.

Example Prompt Directive:
"For any domain score below 6/7, quote the exact sentence that caused the deduction and specify whether the error relates to register, grammatical accuracy, cohesion, or task fulfillment."

Step 3: Run a Counter-Factual Strictness Check

LLMs tend to drift toward generosity over long conversations. Counteract this by running a secondary prompt immediately after receiving your initial feedback:

Example Prompt Directive:
"Now act as a chief HKEAA moderation panel reviewing this grade. Identify 3 concrete reasons why a strict examiner would mark this essay one level lower. Check specifically for repetitive sentence structures, missing prompt elements, and lack of depth."

Building an Exam-Ready Revision Routine

Using AI to support your past paper drills can save dozens of hours, but blind trust in unguided chatbots is one of the quickest ways to stall your score growth. Effective revision requires systematic error analysis, reliable syllabus boundaries, and targeted drills. If you are compiling revision materials across core subjects, explore our curated DSE revision notes and guides to ensure your conceptual foundations match real exam requirements.

For educators designing mock papers and calibrating rubrics across entire Form 5 and Form 6 cohorts, see how educators use dedicated AI assessment tools for schools to maintain standard-aligned grading without manual mark inflation.

To turn your writing weaknesses into reliable Level 5 execution before exam day, start targeted practice on the Thinka adaptive exam preparation platform, where practice questions, instant feedback, and mark schemes are calibrated directly to official assessment standards.