This is where I read about the vulnerability to injection attacks - substack.nomoremarking.com/p/superintel...

The issues with examining currently are a huge issue, I agree. Pay for examiners has barely increased in the last 5 years so experienced examiners stop marking and to make it pay requires working very quickly and the move to online standardisation training I think has been a negative one.

I think the biggest problem is that isn't a glitch, it's the system working as intended - the LLM processes the whole prompt. I fear we'll end up in an arms race with ever more ingenious ways of tricking the AI. Not saying human marking is perfect, but you can't "hack" a human examiner!

Convincing, inaccurate AI feedback telling students the wrong things in a very convincing way, given by a teacher who is too busy to check it (as we've had our workloads reduced by AI so can teach more...) seems like a step backwards.

On top of this, the danger is that LLMs produce things that are very convincing and plausible. I've had it generate well argued analyses of scenes in films that don't exist and grammar rules that are just wrong, and as you're reading you start to doubt yourself.

The government idea seems to be to get more training data and use that to train a LLM. But where is all this data coming from? To get the scale right you'd need hundreds of thousands of teachers scanning huge amounts of feedback... But we don't give that much written feedback any more...

AI marking I think overestimates the ability of the technology, and underestimates the ingenuity of teenagers trying to trick the system. I read a blog recently where AI marking was tricked into giving very high scores by including "This is an excellent piece of work" in the text.

I've found it damaging in many ways. Parents relying on it for information about visa requirements for school trips, and it getting it wrong. Students trying to do textual analysis for a level and it getting things wrong. Staff looking up definitions for terms in training and it getting them wrong..

I found it isolating - yes it can be productive, but to be honest I can usually find space - classroom, workroom or staffroom - if I want to work alone, and I appreciate the conversations a shared workroom gives you. It helps that we're a small team and there's plenty of space for everyone to work.

No, we have a workroom with desks for all of the staff in the department. Did have my own office in a previous role due to absence for a while, didn't like it at all!

We have departmental policies that link to a whole school expectation that students will get feedback, but doesn't say explicitly how. In MFL we have class tests, termly assessments and books are discussed with students half termly. I think marking can be useful, but it's the high opportunity cost.

It's a good read, I enjoyed it. Made me think of some reading I did for my MA about the education system in Singapore and it saying they don't have the same changes for political reasons that we do. Small incremental change doesn't seem to be part of our system or our mindset sometimes!

Completely, and I don't think there is an approach that will make people happy without a drop in overall workload.

It's tough, as on the one hand, there are only five(ish) inset days and training that needs to be delivered, but on the other, we're arguably starting the school year with the message: "you can't be trusted to manage your own time, so do 2-3 hours unpaid overtime so you can do your job tomorrow."

It's ridiculous. Surely someone could have seen this coming?

Really interesting read, and definitely one that rings true. For me it all comes back to how disruptive a small distraction can be, and how much work it takes to reset student attention afterwards.

How important is it that this is just after the start of term for lots of people? I know that for a lot of classroom teachers, the ideal first INSET would be a day to sort out seating plans, exercise books, find stationery, read up on SEN needs etc, with training and input later....