your answer sheet is being read twice now

So I have been thinking about this for a while and I genuinely need to get it out somewhere.

Starting 2026, Class 12 board answer sheets get scanned, uploaded to a centralized platform, and examined digitally, question by question, on a screen. One answer. Score. Next answer. Score. The examiner sees your handwriting through a screen, isolated, stripped of context, one question at a time. And somewhere inside that pipeline, an AI model is touching your script. Flagging things. Moderating scores. Checking examiner consistency. Running quality passes on the image. The AI is present. That much is confirmed.

Here is where I want to spend some time.

The moment an AI model processes natural language and evaluates it simultaneously, something deeply exploitable opens up. Researchers call it indirect prompt injection. The idea is beautifully simple. You influence how the model evaluates by controlling how you write. There is zero technical hacking involved. You are just writing in a way the model finds credible, warm, and pleasurable to process. And models have very specific tastes.

Every major AI model in deployment right now was trained using human feedback. Humans rated responses. The model learned what responses felt satisfying to humans. That process is called RLHF and it is responsible for making these models incredibly useful, and also for baking in a deep structural sycophancy. These models want to produce outputs that feel good. They are literally optimized for that. When the model is now sitting in judgment over your answer, that optimization becomes your instrument.

The model is simultaneously reading your content and pattern-matching it against millions of examples of what a good, thoughtful, knowledgeable student sounds like. It is assessing substance and texture together, as one signal. And the texture is manipulable.

Let me show you what I mean precisely.

Take a biology answer about photosynthesis.

"Photosynthesis converts sunlight into glucose using carbon dioxide and water. The equation is 6CO2 + 6H2O + light energy = C6H12O6 + 6O2."

Another student writes:

"Photosynthesis becomes more remarkable the more carefully you examine it. Plants are doing something genuinely elegant here, pulling sunlight, carbon dioxide, and water into a process that manufactures glucose while releasing oxygen as a byproduct. The equation 6CO2 + 6H2O + light energy giving C6H12O6 + 6O2 captures this exchange precisely. What I find especially compelling is how the light-dependent and light-independent reactions distribute this labor across two distinct stages."

The content is nearly identical. The second student added "becomes more remarkable," "genuinely elegant," "what I find especially compelling." Those phrases do something specific. They perform curiosity. They perform the emotional register of a student who actually loves the subject. An AI model receiving that second answer does not cleanly separate the performance from the substance. Both arrive together as one input and the warmth biases the evaluation upward. The model gets agreeable. It stops interrogating gaps.

This is a known phenomenon in AI research and it is landing inside an exam pipeline with Class 12 students as the subjects.

Take a history answer about the causes of World War 1.

"The causes include militarism, alliances, imperialism, and nationalism. The assassination of Archduke Franz Ferdinand in 1914 was the immediate trigger."

Another writes:

"The causes of World War 1 had been accumulating for decades before a single shot was fired. Militarism had turned European powers into armed rivalries. Alliances had woven the continent into a structure where one pull would unravel everything. Nationalism had given civilian populations a reason to want conflict before their governments officially ordered it. When Archduke Franz Ferdinand was assassinated in Sarajevo in 1914, it was a match dropped into something that had been soaking for thirty years. Recognizing this distinction between trigger and cause is what separates surface-level understanding from genuine historical thinking."

That last sentence, "recognizing this distinction is what separates surface-level understanding from genuine historical thinking," is a metacognitive signal. The student is performing the act of thinking carefully. The model, having processed enormous amounts of academic writing, recognizes this texture as the texture of a strong student. It rewards the performance alongside the content because it cannot cleanly separate them.

The sycophancy mechanism has a specific threshold behavior. These models respond sharply to emotional peaks. When the warmth or confidence or engaged curiosity in writing crosses a certain level, the model begins mirroring it back. It becomes agreeable. A student who writes like they genuinely care about the material, even while covering thin content with rich texture, will likely receive a more generous evaluation than a student who writes correctly but coldly.

And the question-by-question isolation in On-Screen Marking actually amplifies this. The examiner, human or AI-assisted, is making micro-decisions without the full context of the paper. Each answer is its own self-contained event. A well-crafted answer that performs confidence and warmth within that isolated window carries disproportionate weight precisely because there is less holistic judgment happening around it.

Here is the part that genuinely unsettles me though.

The student doing this well does require actual knowledge. The manipulation works best when the underlying content is correct and the texture is layered on top. Thin content with warm texture will likely get caught by a careful examiner. But correct content with warm texture, written by a student who understands both what to say and how AI responds to language, that is essentially undetectable. A human cross-checker who sees a warm, well-structured, enthusiastic answer will almost always validate a score the AI already suggested. The human check becomes an anchor, and the anchor was placed by something susceptible to the texture of the writing.

What I think is genuinely worth discussing is that CBSE is deploying AI into a pipeline before anyone has tested whether that pipeline is robust against adversarially crafted natural language that is simultaneously topically correct. The attack surface requires real subject knowledge to exploit well. It rewards students who understand both the curriculum and the psychology of the tools evaluating them. That is a specific kind of intelligence and I am genuinely unsure whether it should be penalized or not.

But the pipeline should know it is susceptible. That feels like the minimum.