One formula, 280,000 grades
In spring 2020 England cancelled its A-level exams because of Covid. Teachers estimated a grade for every entry. Ofqual, the exams regulator, then ran a statistical model. For larger classes it kept each teacher’s ranking of their students and fitted it to the school’s results in earlier years. For the smallest classes it used the teachers’ grades, and it blended them in for small ones, which favoured schools with small classes.
Quick check
Guess first: what made the 2020 model most dangerous?
Results came out on 13 August. Of 718,000 entries, 39.1% were lower than the teacher’s estimate: 35.6% by one grade, 3.3% by two, 0.2% by three. That is about 280,000 entries, and university places rested on them.
On 17 August the government abandoned the adjusted grades and gave students their teachers’ estimates. Run the case through four questions.
Four questions, thirty seconds
1. If this is wrong, can it be undone? A wrong sentence in a draft is fixed by editing. A lost university place in August is hard to get back by September. A sent message, a payment, a public post and a dose are one-way doors. They deserve a pause that two-way doors do not.
2. Who bears the cost, and did they choose it? If it lands on you, you may accept the risk. The students chose nothing. They could not see the model and had no say in it. When the person harmed did not choose the tool, the bar rises sharply.
3. How many times will this happen? One wrong grade is an incident. The same rule applied to 280,000 entries is a policy. A human marker with a bias affects the scripts they mark. A model with the same bias affects everyone, identically, at once.
4. Is there a route to appeal, and does anyone answer it? On 15 August Ofqual published its criteria for appeals based on mock exams and suspended them within hours. A system with a working appeal can be wrong and still recoverable.
Three tiers
Most uses land in one of three tiers.
Low. Reversible, affects only you, happens once. Drafting, brainstorming, explaining a concept. Verification: skim, and save your attention for the other tiers.
Medium. Reversible with effort, or it affects colleagues, or it repeats. Emails to clients, reports someone will act on, code going into a system. Verification: check every name, number and claim, and know what the answer should be before you read the output.
High. Irreversible, or it affects someone who did not choose it, or it runs at scale, or there is no appeal. Anything medical, legal or financial. Anything that decides about a person. Verification: a primary source you open yourself, a second reader, and a record of what was checked.
Some uses sit outside the tiers, because the answer is not to use the tool: a crisis conversation, or a decision you are licensed to make in person.
Decide on a calm afternoon
Risk judged in the moment follows convenience. At six on a Thursday evening everything looks low risk. Write the rule down while you are calm: anything that goes to a client is checked against the source, and anything about a person gets a second reader.
An organisation needs the same thing on one page: which tasks may use which tools, what each tier requires, and whom to tell when something goes wrong. A policy that only says “use AI responsibly” hands the whole question back to the person least able to answer it at 6pm.
Check your understanding
Two uses: drafting a blog post you will edit before publishing, and writing a standard rejection sent automatically, from an address that takes no replies, to 3,000 job applicants. Why does the second need far more care?
Pick the one you would defend. Your answer stays private.
Your place and first choices stay on this browser. Classroom credit is recorded separately.
Stuck on something?Ask Rukmini, your AI tutor