Should Schools and Businesses Allow AI Tools in Assessments and Training?

Bottom line: Yes, but not everywhere, and not without redesigning the assessment first, because banning AI tools from training and exams mostly punishes the honest people while doing nothing to stop the ones determined to cheat. The real question isn't "should we allow AI," it's "does this assessment test something AI can't fake," and most current assessments fail that test long before anyone opens a chatbot.

What's happening in classrooms and offices right now

The policy debate is running years behind the actual behaviour. A survey by the Higher Education Policy Institute found that around 88 percent of UK undergraduates now use generative AI tools to help with their assessments, a huge jump from the small minority who admitted to it when the same question was asked the year before. That's not a niche habit anymore. That's the default way most students under 25 approach a written assignment.

Then there's a study out of the University of Reading that should worry every exam board in the country. Researchers submitted AI-generated answers into real undergraduate exams without the markers knowing, alongside genuine student work. Ninety-four percent of the AI-written answers went undetected. On average, they scored higher than the real students, by roughly half a grade boundary. If trained examiners can't reliably spot AI work sitting inside a live exam, the entire premise of banning it and hoping detection catches the cheats is dead on arrival.

The exam boards know this. The UK's Joint Council for Qualifications updated its guidance to say AI-generated content that isn't declared as such counts as malpractice, the same category as plagiarism. That's a sensible rule on paper. In practice, nobody can enforce it consistently, because there's no reliable way to prove a piece of writing came from a machine rather than a person who writes in short, plain sentences (which, incidentally, describes half the humans I know, myself included on a good day).

A training room story that changed how I think about this

I ran a two-day workshop early last year for the sales team of a UK manufacturing firm, about forty reps, mostly field-based, mostly not what you'd call digital natives. The final exercise asked each rep to write a cold outreach message to a fictional prospect, timed, twenty minutes, then we'd read a few out loud.

Nobody had told them AI was off-limits, because it hadn't occurred to me that it needed saying. Within about four minutes, I noticed maybe half the room typing into their phones in a way that had nothing to do with typing a message from scratch. A few weren't even hiding it. When I asked, most said they'd used ChatGPT to get a first draft, then edited it.

Here's the part that mattered. The messages that were obviously just pasted straight from the tool were flat, generic, forgettable, the kind of thing that gets deleted in half a second. But two reps used AI to get a rough draft down fast, then rewrote roughly 80 percent of it in their own voice, adding specific detail about the prospect's actual business. Those two produced the best messages in the room, by a distance. The tool wasn't the problem or the solution. What they did with the output was the whole story.

That workshop changed how I design every assessment I build now, whether it's for a client's sales team or for my own training programmes. I stopped asking "did you use AI" and started asking "what judgment did you apply after you used it." That's a completely different question, and it's the one that separates people worth hiring from people who aren't.

Training and assessment are not the same problem

This is where most workplace policies get muddled, and it's worth separating the two because they need opposite rules.

Training is about building a skill. If someone uses an AI tool to help them understand a new compliance framework faster, or to generate practice scenarios, or to summarise a dense policy document into something they'll remember, that's the tool doing exactly what it should. You want people to learn faster. Blocking AI during training because it feels like cheating is like banning calculators from a maths class and calling it rigour.

Assessment is about certifying that a skill exists. This is where AI use has to be either fully open (the task is designed assuming AI is used, and the judgment shown around it is what's graded) or fully closed (no devices, no tools, a supervised environment). The mess happens in the grey middle, where a company or school says "don't use AI" for an unsupervised, take-home assessment and then acts shocked when nobody follows the rule. You can't ban something you can't police. Pick open or closed, and mean it.

A four-step framework for deciding when to allow it

I use this with clients who come to me confused about how to write an AI policy for their induction programme or their internal certifications. It works for schools too, with small adjustments.

• Step one: name what you're testing. Recall of facts? Written fluency? Problem-solving under pressure? Judgment when the situation is messy? Write it down in one sentence before you write a single exam question.
• Step two: sort assessments into open-book or closed-book, and say so out loud. If AI is allowed, tell people plainly, in the instructions, not buried in a policy document nobody reads. If it isn't, make the setting closed, supervised, no phones, no laptops.
• Step three: build in a disclosure step, not a detection step. Rather than trying to catch AI use after the fact (which the Reading study shows you'll mostly fail at), ask people to submit their prompt history or a short note on how they used the tool alongside their answer. This shifts the culture from hiding to explaining, which is a far more honest habit to build in anyone, student or employee.
• Step four: redesign around 40/30/30. I now split every assessment I build into roughly 40 percent knowledge checks (fine for AI, low stakes, quick to redo), 30 percent applied scenarios specific to that person's actual context (harder for AI to fake convincingly, because it needs real detail only the person has), and 30 percent live, observed performance (a role play, a presentation, a real conversation, no tool can sit that exam for you).

Why most AI bans in training don't work

There's a part of this that businesses in particular don't like to hear. Plenty of the same companies writing strict "no AI in assessments" policies for their staff are using AI, extensively, to write the training material those staff are being tested on. The induction deck was drafted with Copilot. The onboarding quiz questions came out of ChatGPT. The slide design ran through Canva's Magic Design, the same AI feature that Melanie Perkins built her company's edge around, and there's a good lesson in how she scaled that tool that's worth reading if you're choosing which platforms to build training content in. Nobody flags any of that as a problem, because it's management doing it rather than the trainee. That's not a policy on AI. That's a policy on who's allowed to use it.

The uncomfortable truth underneath all of this is that a lot of assessments and training modules were badly designed long before AI turned up. They tested recall and format, things a language model happens to be brilliant at, because that's precisely what these models are trained to reproduce. The panic about AI "ruining" assessments is often really a panic about how thin the assessment was to begin with. If a chatbot can pass your compliance quiz in ninety seconds, that quiz was never testing competence. It was testing whether someone could find the right paragraph in the handbook.

I've written before about the wider question of whether AI tools should be allowed in schools and workplaces at all, and that piece covers the broader cultural argument. This one is narrower and sharper on purpose, because "should we allow it" and "how do we test people fairly once everyone already has it" are two different questions, and only one of them has a practical answer you can act on this quarter.

What good policy looks like in a school

Schools have it slightly harder because the stakes (GCSEs, A-levels, university offers) are higher and the supervision is patchier, especially for coursework done at home. The most workable approach I've seen from forward-thinking schools does three things: it moves a larger share of the final grade to supervised, in-class work; it treats AI as a legitimate research and drafting tool for homework, with mandatory disclosure of how it was used; and it teaches AI literacy directly, as its own short unit, rather than pretending the tool doesn't exist and hoping students won't notice. Some schools are also leaning into project-based and game-based assessment, where performance in a structured, interactive task is harder to fake than a written essay. There's a decent parallel in how geography puzzle games improve learning and engagement, because the same principle applies more broadly: active, structured tasks reveal understanding in a way a static essay question simply can't.

What good policy looks like in a

I have written more around this on the site: How Many AI Tools Exist, and How to Pick One Without Losing Your Mind, Should AI Tools Be Allowed in Schools and Workplaces?, 20 Free and paid webinar software tools for 2026.

Sources worth reading

• Anthropic's own documentation
• the NIST AI Risk Management Framework

Bottom line: AI tools should be allowed but with clear rules about when and how they're used. The goal isn't to ban AI, it's to test skills that still matter once AI is in the room, like judgment, editing, and knowing what to ask for. Schools and businesses that build this in now will be ahead of those that just say no.

For contributors covering this area, read guest post guidelines for AI writers.

Related: writing for us on culture and art.

Frequently asked questions

Should schools let students use AI during exams?

It depends on what the exam is meant to measure. If the goal is testing recall or basic understanding, AI use should probably be restricted. If the goal is testing how someone applies knowledge, solves problems, or produces finished work, allowing AI with disclosure makes more sense than pretending it doesn't exist outside the classroom.

How can businesses use AI in training without weakening skill-building?

Businesses can let employees use AI for drafts, research, or first passes, then assess the thinking behind the final output. Ask people to explain their choices, catch errors, or improve on what the AI produced. That way training still builds judgment, not just familiarity with a tool.

What are the risks of banning AI outright?

Banning AI tends to push its use underground rather than stop it. Learners and employees end up using it anyway, just without guidance on doing so responsibly. A ban also means people arrive in the workplace unprepared for tools they'll be expected to use on day one.

What policies work best for allowing AI in assessments?

Clear disclosure rules work better than blanket permission or blanket bans. Specify which tasks allow AI assistance, require people to note when and how they used it, and design some assessments to happen without AI at all so core skills still get checked directly.