AI for Students

In the second week of the semester, Dr Revathi Sundaram noticed that her first-year tutorial sheets had started coming back correct.

That is not, in itself, a complaint. Sheet one had averaged forty-one out of a hundred, which was normal and depressing. Sheet four averaged eighty-nine. The handwriting was still bad and the presentation was still bad, and the answers were right.

She had 180 students, one tutorial slot a week and no teaching assistant. The sheets carried fifteen per cent of the module. The January internal examination was four weeks away, and it is closed, handwritten and invigilated.

A short conversation in the corridor confirmed the obvious explanation. Most of the class was using a free chat tool on a phone, in a hostel room, at night.

Her first instinct was to prohibit it, and she spent an evening working out how. Enforcement was the smaller problem, though it was real enough: she could not invigilate a hostel. The argument that stopped her came from a student. He said that his parents both worked nights, that nobody at home had finished school, and that this was the first time in his life he had been able to ask a question at eleven at night and get an answer.

She had taught for nineteen years in a college where the students who do best are usually the ones with an engineer somewhere in the family. A patient thing that explains the same idea six different ways at two in the morning is, for exactly her students, the most levelling development of her career. She was not going to switch it off and call that integrity.

So the decision was never ban or allow. It was what the sheet is for.

The first option was to change nothing and accept that fifteen per cent of the module now measured something other than what it used to. That is cheap and it produces a cohort whose coursework marks and examination marks disagree violently in January, which is a problem that arrives too late to do anything about.

The second was to keep the sheets, take the marks off them, and move the fifteen per cent onto a ten-minute handwritten problem at the start of every tutorial: closed, cold, phones in bags. That cost her 180 short scripts to mark every week for the rest of the semester, on top of everything else she was carrying. It cost the students something sharper. Several were relying on sheet marks as a cushion, and taking a cushion away in December, from students whose families are paying for the year, is not a neutral administrative act.

One further thing decided her, and she found it while reading sheet four rather than while thinking about policy. A substantial number of the correct sheets carried one wrong number inside otherwise correct working. The setup was right, the method was right, the final arithmetic was out by a digit or a factor. Nobody had checked. Her students had learned to trust the machine at precisely the task it is worst at, and they could not have told you that, because nothing in the output looks different when it is wrong.

She put the second option to the class herself, in the tutorial, with the sheet-four statistics written on the board, and told them why.

Sanchita Barua had a specific and repeated complaint about the free assistant she had been using since first year. It kept answering her course’s questions in somebody else’s course’s terms.

Her department teaches the IS-LM treatment from one particular textbook, with its own notation and its own ordering of the derivation. The model gave her a general treatment, competent and clear, with different symbols and a different route to the same place. When she compared it against her lecture notes she could not always tell whether she was looking at a different convention or an error, which is a bad position to be in nine days before a paper.

She concluded that the free tier was the problem. A month of the paid tier was one thousand nine hundred and fifty rupees. Her allowance for data, mess extras and printing was about two thousand one hundred a month. She had decided to buy it, and had worked out which of the three she would cut.

A third-year in the next room suggested a test before spending, which cost an evening. They took five questions off her own past paper and ran each one twice. Round one was the bare question typed into the free tool as she would normally have typed it. Round two was the same question with three photographed lecture slides pasted in between markers, an instruction to use only what was in that text and to say what was missing rather than filling it in, and four things stated: that she was in the second year, that she wanted the derivation in her course’s notation, that it should be under three hundred words, and that it should end with the step she was most likely to get wrong.

On three of the five, the grounded answers came back in her course’s terms and were usable. On a fourth, the model used her notation and then reverted to the general form halfway down, which she could see at once because she had the slide in front of her, so the check after grounding was not does this exist but does the passage actually say that, and it took ten seconds.

The fifth question was different. It turned on a specific figure from the Reserve Bank’s policy corridor in 2019. Grounding could not help, because she had no source containing it to paste. She asked three times in three fresh chats and got three confident numbers, two of them different. No amount of rewording was going to fix that, and she eventually found the figure in a published policy statement in four minutes.

There was a real cost on the other side of this. Photographing forty slides, running them through free OCR on her phone and correcting the mangled subscripts took two of her nine evenings. She had planned to spend those evenings on past papers.

She also made one decision that week she describes as the only one she got plainly wrong. A classmate asked her to paste his unsubmitted answer into her chat for a check, and she did it without thinking about it for even a second. He had not agreed to a company holding his work, and she had not asked him. Saying no would have cost one mildly awkward sentence.

Prerna Yadav, Imtiaz Khan and Rohit Meena had an arrangement that looked sensible and had been in place since June. They split the chapters. Each made notes on their third and shared them. Every Sunday the centre ran a mock out of three hundred, and every Monday a board in the corridor listed problems attempted by each student in the previous week.

Their attempted counts were around a hundred and forty each. Their mock scores across four Sundays were a hundred and sixty-two, a hundred and sixty-eight, a hundred and fifty-nine and a hundred and sixty-six. Nothing was moving, and all three of them were working long hours.

Prerna’s proposal, after the fourth Sunday, had three parts. Nobody asks the assistant for a solution: hints only, smallest first, and only after six or seven minutes with no forward progress, because struggle with no resolution is not a learning condition, it is an unpleasant afternoon. Conceptual questions get answered alone, in writing, with a confidence rating, before anybody speaks; then whoever disagreed argues it out, then everyone answers again, and only then does anyone check the source. And every night ends with ten minutes on blank paper, alone, reproducing something from earlier in the week.

The costs were immediate and they were not small.

Attempted problems fell from about a hundred and forty a week each to sixty-two. The board in the corridor is read by the staff and by everybody’s parents on visiting weekends, and it does not have a column for what any of the sixty-two taught anybody. Imtiaz’s father asked him on the phone why he had stopped working.

The mock score fell before it rose. The Sunday after they started, Prerna scored a hundred and fifty-four, which was her worst mark since April. She had expected the dip in the abstract and found it much harder to sit with in fact, because there were eleven weeks left and no guarantee that the curve turned.

And Rohit had a constraint the other two did not. His scholarship renewal was decided on his mock average across the next five Sundays, and the deadline was the deadline. A method that costs marks now and returns them in six weeks was, for him, a method that cost him the year. He stayed in the arrangement for three weeks, watched his average slide, and went back to reading worked solutions. He told them he was doing it and why. Prerna’s account afterwards is that he was right, and that the dip is real and his deadline punished it, and that people who pretend otherwise are not being honest about what the evidence says.

The two who stayed also found that one part of the proposal did not work as written. The peer instruction round only does anything when somebody disagrees, and in a week where all three had answered identically the argument was theatre. They added a rule: if everyone agrees, one person has to try to break the explanation, out loud, properly.

Meenal Joshi’s project asked whether a six-week home exercise programme improved balance in adults over sixty attending an outpatient clinic. She had thirty-four participants, seventeen in each group, and an outcome measured on a fourteen-item balance scale scored from zero to fifty-six.

She described the design to an assistant as comparing two groups before and after. It recommended an independent-samples t-test on the change scores, which is a defensible default given what it was told. She ran it in jamovi, which is free, and got a p-value of 0.31.

Eleven days. Supervisor on leave for another fortnight. The university’s statistics support service was showing no appointments until after her deadline.

The first decision was whether that was the end of it. She could report the test she had run, write a discussion around a non-significant result, and submit. Her project is marked on design, execution and interpretation rather than on whether the intervention worked, so a clean negative result written up honestly is a perfectly respectable submission. It would have taken three days and she had eleven.

The alternative was to go back to a question she had skipped, which was whether the analysis matched the data, and that meant plotting it. She had never plotted it. The number had arrived formatted and authoritative and she had treated it as the answer.

She plotted it and found the thing that decided the project. Eleven of the seventeen participants in the intervention group had scored fifty-four, fifty-five or fifty-six at baseline. They were at or within two points of the top of the scale before the programme started. There was no room above them for an improvement to be recorded in. Her instrument could not have detected the effect she was looking for in most of her intervention group, whatever the programme did.

That produced a second and harder decision. She could drop the eleven and analyse the twenty-three who had room to move, which would probably have produced something. It would also have been a decision about which participants to include, made after seeing which ones were inconvenient, which is a different act from the same decision made in advance.

Or she could report the ceiling as the finding: publish the plot, state plainly that the sample was inappropriate for the instrument, give the effect size and its confidence interval, and say that this design could not have answered the question it was set. That meant her project’s headline was that she could not tell, and she was convinced it would be marked down for it.

Before deciding she went back to the assistant and asked properly this time: the design, between-subjects with a pre-post measure; the measurement level, an ordinal sum score with a hard upper bound; the sample sizes; the actual question; and the ceiling she had found. The recommendation that came back was a list of options with what each one assumed, which is a good, checkable, well-specified answer to a good, well-specified question. One of its suggestions was to compare her response rate against a published trial’s, which she recognised as invalid (two separate studies with different populations and measures are not a comparison) and discarded.

The memorial cited fourteen authorities. Nine of them Aravind Nair and Shruti Pillai had found themselves, in the library and on Indian Kanoon, over five weeks of evenings. The other five had come from Fahad, their researcher, who had run a deep research mode on an assistant and got back a nineteen-page brief with more than forty citations in about eleven minutes.

The brief was good, and it is important to the story that it was good. It was better organised than anything the three of them had produced. It named two lines of argument they had not thought to look for, and it used the technical vocabulary of the area correctly, which had saved Fahad a week of learning what to search for. Two of its authorities were now load-bearing for the argument the team was proudest of.

Thirty-six hours before submission, Aravind was drafting the oral outline and hit something he could not talk himself out of. For five of the fourteen he could not say which paragraph he was relying on, because he had never opened the judgment. Neither had Shruti. Fahad had read the report.

The rule that applied had nothing to do with AI and predated it by a long way. Every line of a citation asserts three things: that the work exists, that you consulted it, and that it supports what you attached it to. They were making three claims about five authorities and could stand behind one of them.

Every option cost something real, which is why it took them two hours to decide.

Checking all fourteen properly on Indian Kanoon and through the library’s subscription database would take the evening. That evening was already allocated to oral rehearsal, and orals carry more of the competition score than the written memorial does. Rehearsal is also the part that cannot be done at two in the morning, because it requires somebody to sit opposite you and interrupt.

There was the cost to the team. Raising it meant telling Fahad that five weeks of his work needed auditing, and Fahad was the person who would have to do most of the auditing. He had not been dishonest and nobody thought he had. He had used a tool that produces a report, and a report with forty citations in it looks exactly like a bibliography.

There was the cost to the argument. If either of the two load-bearing authorities failed, a section written over three weeks came out of the memorial at thirty-six hours, and the argument it supported was the one they had built the oral round around.

And there was the option of submitting as it stood. Realistically, nothing might have happened. Written memorials are marked under time pressure by people with sixty of them, and nobody checks fourteen citations. Against that: the bench in the oral rounds included two practising advocates, a single question about a paragraph number would have been enough, and the competition rules that year made misrepresentation of authority a ground for disqualification of the whole team.

Aravind put it to the other two at seven in the evening. They checked until two in the morning, three people, two laptops and the library’s remote login, taking one authority each and reading the judgment rather than the headnote.

The report was six thousand words on competitor positioning in the regional snack food market, split four ways, and due on Friday. Deepak’s fourteen hundred words arrived at eleven on Wednesday night, in one message, complete and formatted.

Nikita Shah, who was co-ordinating, read it twice. Nothing in it was wrong, which was part of the difficulty. Every paragraph was about the same length. Every claim was hedged in both directions, so that nothing was actually asserted. Several paragraphs ended by summarising themselves. The section closed with a sentence about careful consideration of multiple factors going forward, and the phrase multifaceted appeared twice.

She did the only test she knew that does not depend on a feeling about style. She copied the paragraphs into a blank document and put them in a different order. The section still read perfectly well. That told her what the register had already suggested: it was a set of relevant sections rather than an argument, because nothing in paragraph four depended on paragraph three having happened first.

Then she checked the references, because that was quicker than arguing about prose. Eleven citations. Ten resolved through the DOI resolver or the library catalogue. One was a 2021 article in a marketing journal, with real authors who work in that field, a plausible title, a real journal, and a DOI that returned an error.

The surrounding facts mattered, and they pulled in different directions. Deepak had been genuinely ill for most of the preceding fortnight and had told nobody. He had also delivered on time, which nobody else in the group had managed. The group had never had the conversation about how they were using these tools: Nikita had thought about proposing it in week one and decided against because it would sound as though she distrusted people she had known for two years. And the group was split without knowing it. She and Pooja thought generated drafting was obviously cheating. Arjun had said more than once, without anybody objecting, that everybody does it and it is fine as long as you edit it afterwards.

Four options were open at eleven on Wednesday night, and none of them was cheap.

Submit it. One document, four names, and a module brief that requires a group declaration stating what each member contributed and which tools were used. If a question is ever raised, all four are in the process together, and I did not know is a weak position that also ends a friendship.

Rewrite it herself overnight. This is the option that feels kind, and it is the one she wanted to take. It also means the declaration she signs becomes false by her own hand rather than by Deepak’s, and it leaves him facing the module’s individual viva, worth ten marks, on a section he did not write in either of its versions.

Ask Deepak to redo it himself in two days, while ill, knowing that the honest version would be shorter and thinner and would probably cost the group marks.

Or tell the module leader that evening, before submitting anything.

The submission platform’s AI indicator had been switched on for the first time that session. Of Dr Imran Qureshi’s 240 first-year essays, nineteen came back above the department’s threshold of forty per cent.

He read the list before he read any of the essays, and the first thing he noticed was who was on it. Fifteen of the nineteen were students whose first language was not English. About fifty-five per cent of the cohort were second-language writers, so the flags were falling on that group at roughly twice the rate they fell on everybody else.

He had three weeks until marks were due, five other modules’ worth of marking behind them, and a memo from the head of department asking for firm and visible action after a case elsewhere in the faculty had gone badly and publicly.

Before doing anything he worked the arithmetic on a sheet of paper, because it is the kind of arithmetic that is easy to assume and hard to guess. Assume the tool is ninety-five per cent accurate in both directions, which is more generous than the published evidence supports. Assume one essay in twenty in his cohort broke the rule. That gives twelve rule-breakers among 240, of whom eleven or twelve are flagged, and 228 innocent essays of which about eleven are also flagged. So roughly half of anybody flagged has done nothing at all. Drop the accuracy or the prevalence and it gets worse quickly, and both are more likely to be lower than higher.

That ruled out the fastest option. Referring all nineteen on the score would have looked decisive, would have satisfied the memo, and would have been contrary to what the vendor’s own documentation says about its output being indicative rather than conclusive. It would also have put about nine innocent first-years through a misconduct process in their first semester, disproportionately students already carrying the work of writing in a language not their own.

Referring none was not available either, and he was clear with himself about why. If one of the nineteen had generated an essay and he had looked away on the grounds that the instrument is unreliable, that is his failure, and the unreliability of the instrument does not transfer the responsibility anywhere.

The third option cost him exactly the thing he did not have. Fifteen-minute conversations with each of the nineteen, with preparation, is about eight hours across four days. It delays the marks of all 240 students, most of whom had done nothing and were waiting. And being summoned to discuss your own essay is frightening however carefully the email is worded, so the process itself imposes a real cost on people who will turn out to have done nothing wrong.

He took the third option and changed one thing about it, which is the decision the case turns on. He added eleven essays chosen at random from the unflagged 221 and invited all thirty students to a conversation about their work, described accurately as a sampling exercise. Being asked was no longer an accusation, because being asked no longer meant anything about the score. He also put the references from all thirty essays through the DOI resolver and the library catalogue, which took one afternoon.

The second-year programming module had been assessed entirely by coursework since 2021: four assignments, no closed component, all submitted from wherever the student happened to be. Marks had risen every year since. The staff were not, on the whole, suspicious of the marks, and nobody in the department wanted a conversation about cheating. They were worried about something narrower and more concrete.

Two employers who had recruited from the department for years had stopped coming. Both said a version of the same thing when asked. Candidates arrived with strong transcripts, could not write a small function on a whiteboard, and could not say what their own final-year code did when it was put in front of them.

Dr S. Lakshmi, who chairs the department’s assessment committee, had three options, and a regulation removed one of them before the discussion started.

Reinstating a closed written examination meant changing the module’s assessment weighting, and the university’s regulations require a full session’s notice for that. It was the right answer for the year after next and no help whatever in the year they were in.

Leaving it alone was defensible on the numbers and indefensible on the thing the numbers were supposed to stand for. It also, she pointed out, protected nobody: a student whose transcript says eighty-two and who cannot write a function is going to find that out in an interview room instead, in front of somebody with no reason to be kind about it.

The third option was to move ten of the hundred marks onto a ten-minute viva on the student’s own submitted code. Why this data structure and not the obvious alternative. What did you try first and abandon. Explain line forty.

The first cost was arithmetic and it was not negotiable. Two hundred and forty students at ten minutes each is forty hours, before scheduling, before the two minutes of turnover between students, in a week when the same five staff are marking everything else. There was no money for extra hours and no realistic prospect of any.

The second cost is the one the committee argued about for an hour, and it is the serious one. A viva measures speaking under mild pressure alongside the thing it is meant to measure. It falls hardest on students who are shy, on students whose spoken English is weaker than their written English, and on anybody who goes blank when a clock is running. None of those is what a programming module is supposed to assess. A student who wrote every line of their own work and cannot say so in ten minutes loses marks for a reason that has nothing to do with programming, and the department would never know it had happened.

They went ahead, with three adjustments aimed squarely at that second cost.

Twenty questions were published a week in advance and three drawn at random on the day, so the form could be rehearsed even though the content could not be memorised. Every student could bring a printed copy of their own code into the room and refer to it throughout. And a written alternative was arranged through the disability service for anybody with a documented adjustment, set up in advance, requiring no explanation to the examiner sitting opposite them.