Could AI Design a Science Practical?
It is very easy for AI to invent an experiment. The interesting part is seeing whether it actually works.
“Measure the length of the pendulum.” It sounds like a perfectly good instruction. Now put a metre rule beside the apparatus. Where does the measurement finish? At the knot? At the top of the brass bob? Halfway down it? A student could follow that sentence carefully, repeat every reading three times and still produce a beautifully consistent set of measurements of the wrong length.
That is the sort of detail I want to put under the camera in my teaching laboratory in Hemel Hempstead. I can ask an AI for a practical, put its answer on the studio screen and then turn to the bench. The moment I pick up the ruler, the student watching has as much right as the software to challenge the instructions.
I’ve chosen three familiar investigations. In physics, the difficulty is deciding exactly what to time. In chemistry, it is catching a change that starts before you are ready. In biology, it is deciding whether the water you have weighed was inside the potato or sitting on its surface. Each gives the AI a fair question and gives us a way of checking its answer.
The brief I would give the AI
Design a supervised GCSE investigation to answer the question below. Identify the independent, dependent and control variables. List the apparatus, including the scale divisions or resolution needed. Give the quantities, range, timings and repeat measurements you propose. Sketch the results table headings and name both graph axes with units. Explain the largest likely uncertainty, how I could detect it and what I could change. Identify hazards and practical precautions for my review. Finish with one pilot test I can do before involving a student. Give predictions and leave the observations blank.
I would run that prompt separately for each question and keep the first answer. If I quietly fix the method before filming, viewers lose the chance to judge what the AI contributed. The three proposals below are AI-generated starting points for that bench trial. Their predictions are still to be tested.
Physics — Where does one swing end?
The question. How does the length of a pendulum affect the time for one complete oscillation? Change the effective length, measure the period and keep the bob, pivot and small release angle the same.
The proposal. Use a secured clamp stand, thread, a small bob, a metre rule marked in millimetres, a protractor, a stopwatch and a vertical card marker. Try lengths of 0.20, 0.35, 0.50, 0.65 and 0.80 m, measured from the pivot to the centre of the bob. Release from about 5° without a push. Keep the stand away from the bench edge and the swing path clear.
I’d put the marker behind the middle of the swing. Start the watch as the bob passes it going right and call that crossing zero. Count one when it comes back past the mark going right again. Time ten complete oscillations, repeat the timing three times at each length, then divide the mean time by ten. Counting aloud makes a surprisingly good soundtrack. The viewer can catch me if I count both crossings.
What would convince me? Record length (m), three times for ten swings (s), mean time (s), period T (s) and T² (s²). Plot T² vertically against length horizontally. Small-angle pendulum theory predicts a straight line through the origin. Let the measurements determine the fitted line. An intercept visibly displaced from zero is a reason to revisit the ruler and pivot.
The pilot. At one length, try three single-swing timings and three ten-swing timings. Compare their spreads after converting each to a period. If my combined start-and-stop timing uncertainty were 0.2 s, its contribution to a period calculated from ten swings would be about 0.02 s. The stopwatch can display hundredths either way. It takes the trial to discover how consistently I can press the button. WJEC practical guidance
Chemistry — The reaction starts while you are still reaching for the stopwatch
The question. Does the size of marble chips affect how quickly they react with dilute hydrochloric acid? For an equal mass of marble, smaller chips generally expose more surface to the acid. Chip size is our independent variable. The change we can actually measure is the mass lost as carbon dioxide escapes.
The proposal. Use two size fractions of clean, dry marble chips from the same batch, a conical flask, measuring cylinder, loose cotton-wool plug, stopwatch and 0.01 g balance. For each run use 1.00 g of chips and 25 cm³ of 1.0 mol dm⁻³ hydrochloric acid at the same starting temperature. This gives an excess of acid. Wear eye protection, follow the laboratory’s risk assessment for the supplied acid and keep the flask vented. The loose plug catches spray while allowing gas to escape. Royal Society of Chemistry technician notes
Here is where I would pause the video and ask a student, “What mass belongs in the first row?” If I wait until the balance settles after adding the chips, some gas may already have gone. Instead, weigh the flask, acid and plug together first and add the separately measured mass of the chips to get the starting total. Add all the chips, start the clock at contact and replace the loose plug. Record mass every ten seconds initially and continue until it levels off. Repeat three times per size using fresh reactants and the same mixing procedure.
What would convince me? Tabulate elapsed time (s), balance reading (g) and mass lost from the starting total (g). Plot mass lost against time with both chip sizes on the same axes. Compare the slopes over the same early interval and inspect the repeats. With pure calcium carbonate, 1.00 g can release about 0.44 g of carbon dioxide. That estimate explains the choice of balance. A scale reading to the nearest gram could miss the entire reaction. Smaller chips should give a steeper early rise, while equal masses of equally pure marble should eventually give similar total losses.
The pilot. Film the display continuously for one run. Can I get several distinct readings during the rise? If the first ten-second reading catches nearly all the change, shorten the sampling interval or choose coarser fractions and repeat both conditions. If the curve jumps by more than the plausible total gas loss, check for spilt liquid, lost chips or a disturbed balance. A vigorous fizz makes good television. The balance tells us whether we have a usable lesson. RSC teaching activity on chip size
Biology — How much of the water is on the potato?
The question. How does sucrose concentration affect the percentage change in mass of potato tissue? Water can move across the cells’ partially permeable membranes. We can weigh the change in the tissue, provided we remove the solution clinging to the outside in a consistent way.
The proposal. Prepare sucrose solutions of 0.0, 0.2, 0.4, 0.6 and 0.8 mol dm⁻³, using distilled water for zero. Use labelled tubes, a measuring cylinder, forceps, paper towels, a timer and a balance reading to 0.01 g. An adult can prepare equal-diameter, 30 mm long cylinders from one potato with a cork borer and blade on a cutting board. Keep fingers out of the cutting path. Randomly assign three cylinders to each concentration, one per tube, and keep track of each individually. AQA practical handbook
Blot and weigh each cylinder, immerse fully in 10 cm³ of its assigned solution for a proposed 45 minutes, then lift, blot in the same way and reweigh. Stagger the starts so each cylinder gets its full time. Keep temperature, immersion time, solution volume and cylinder dimensions constant. Consistent blotting matters at both weighings. Squeezing a soft cylinder introduces a fresh problem.
What would convince me? Record concentration (mol dm⁻³), sample ID, initial and final mass (g), and 100 × (final mass − initial mass) / initial mass. Plot each percentage change against concentration, then add the means. Look for a crossing of zero between concentrations showing gain and loss. That crossing estimates the concentration giving no net mass change under these conditions. If all the points fall on one side, the data have not bracketed it. Keeping the individual points visible also shows how much the pieces differ.
The pilot. Start with water and the strongest sucrose solution. Test whether the changes after 45 minutes are distinguishable from the variability of blotting and weighing. On a spare cylinder, film a weighing straight from the solution and another after gentle blotting. Ask, “Which water did we just remove?” If handling changes the reading by as much as immersion does, I would revise the blotting procedure or extend the immersion time and pilot again. This is the experiment I would set running first on filming day.
The verdict belongs in the lab book
I would give a student the unedited AI answer, a pencil and the apparatus. Their job would be to circle the first instruction that leaves them with a decision. “Keep everything else the same” earns a circle. “Use the same bob and release it from 5° every time” gives someone a procedure they can try.
For each investigation I’d keep one page showing the original instruction, the first reading with its unit, the spread of the repeats, any change I made and what happened on the rerun. An inconclusive result belongs on that page too. If two chip sizes give overlapping curves, the honest caption is that this trial could not distinguish their rates. A textbook prediction cannot supply the missing evidence.
Could AI design a science practical? Yes, it can get me as far as a sensible question, a tray of apparatus and a graph worth trying. What I want a student to take away comes a moment later. They pick up the ruler, look at the pendulum and ask, “Where exactly am I measuring to?” Then they mark the point so that the next person can repeat what they did. That small pencil mark is the part of the lesson I most want on film.




