Gains:
- Being able to distinguish where artificial intelligence saves time in the physics workflow (calculation, simulation, analysis, derivation) and where unit, order and accuracy decisions are left to humans, depending on the level of risk.
- Ability to apply a discipline that controls each output by defining its unit, ensuring its level, and verifying it with an independent tool.
- Understanding why unit/order errors, made-up constants and false references are risks that must be taken into account in this profession from the very beginning.
Physics is the art of describing nature with numbers and equations; But the cruelest rule of this art is this: a unit error, a sign mistake or a shift of order (a result being 10, 100 or 1000 times larger/smaller) invalidates the entire result. Using centimeters instead of meters in an calculation could cause a satellite to miss its orbit; misspelling one exponent is enough to disprove the bearing calculation of a bridge. That's why using artificial intelligence (AI) in physics requires discipline. In this unit, you will learn where the tools we call large language model (LLM - a type of artificial intelligence that produces text statistically, with the logic of "the next most likely word") actually save time in physical work, where they are dangerous, and how to verify each output.
First, a critical concept: hallucination is when AI produces information that is not actually true, with full confidence, as if it were true. In physics this is; It comes in the form of a made-up constant value, the name of a non-existent equation, an incorrect unit, or a reference to a non-existent article. The critical point is this: LLM is not a calculator or a physics engine. It is a text generator that statistically imitates the texts it sees in the training data. He produces correct physics most of the time because he has seen many correct physics texts; but the difference between "true most of the time" and "true all the time" is everything in physics.
Where does artificial intelligence work in physics and where does it not?
Think of AI as a quick draft, idea, and coding partner: valuable but essential to validation. The following distinction is the backbone of this module.
Quest
Contribution of AI
Responsibility of the physicist
numerical calculation
Writes Python/NumPy code
Run the code and confirm the result with the unit
simulation
Establishes a differential equation solution framework
Verify by physical conservation laws
Experimental data analysis
Curve fitting code suggests statistics
Confirming the model's compliance with physics
Model/formula derivation
Provides derivation outline and intermediate step
Checking each step and boundary condition
visualization
Generates graphics code
See the accuracy of axis, unit and scale
Literature/writing
Makes summary, draft, language corrections
Confirming every reference and fact at the source
Rule of thumb: Use AI to navigate, code, and edit the account, not the account itself; Always verify the numerical result with a deterministic tool (executed Python code, dimensional analysis, manual check). "Deterministic" here means "always gives the same and precise result to the same input"; An executed NumPy code is like this, it is not a number that LLM gives off the top of his head.
Three deadly traps of AI in physics
1. Unit and order errors. AI can silently switch between joules and electronvolts, kelvins and centigrade, radians and degrees. The result looks numerically "reasonable" but is physically wrong.
2. Made-up constants and equations. The AI can "fill in" a material constant or a cross-sectional area it doesn't remember with a number that seems reasonable. It can give a name to a law that does not exist.
3. False attributions. AI can produce a completely made-up article in a real magazine, with real-sounding author names. We will cover this in depth in unit 9.
Tip: Approach every numerical AI output with three questions: (1) What is its unit? (2) Is the order (to the power of 10) physically reasonable? (3) How do I check this independently? These three questions catch the vast majority of physics errors in the bud.
Step by step: safe AI workflow in physics
1. Define the problem with its units. Instead of "Find the speed of an object", "If an object with a mass of 2 kg accelerates for 3 s with a constant force of 5 N, what is its final speed in m/s?" say. Write clearly what is given and the desired unit.
2. Request step-by-step solution and unit tracking. Ask the AI to show how the unit is moved at each intermediate step. Units are the most powerful tool for catching errors.
3. Verify by running the code. Confirm the symbolic or numeric result with an executable Python code. This is the central principle of this module.
4. Make a physical countercheck. Is energy conserved? Does the result give the expected behavior in the limit case (when the mass goes to zero and the speed is very small)? Does the dimensional analysis hold?
5. Confirm constants and references. Speed of light, Planck's constant, density of a material—verify it all from a reliable reference (CODATA, textbook, datasheet).
three mini cases
Case 1 — Unit trap. An engineering student asked AI about the energy of a photon and got the answer E = 3.2 × 10⁻¹⁹ J (correct). But in the next step, AI wrote it down as "about 2 eV". The student checked the J conversion of 1 eV = 1.602 × 10⁻¹⁹: the correct result was about 2 eV, consistent. But in another question YZ wrote 3.2 × 10⁻¹⁹ J as “3.2 eV”; When the student did the conversion manually, he saw that it should be 2 eV and caught the error. Time lost: 3 minutes.
Case 2 — Fitting fixed. A teacher asked the AI for the viscosity of water at a certain temperature. YZ gave a number that seemed reasonable but did not cite the source. When the teacher compared the value to an engineering manual, he found that the number given by the AI was about twice the actual value. Without verification, a laboratory calculation would turn out to be completely wrong.
Case 3 — Confirmation gained. A researcher asked the AI for a Python code that numerically solves the equation of motion of a pendulum. The code worked, but the energy steadily increased over time — physically impossible. The researcher realized that this was due to the error of the numerical method (simple Euler step), asked the AI for a method that maintained energy conservation (e.g. smaller step or a more stable solver) and fixed it in 10 minutes. If it were not for the energy conservation check, the incorrect result would be considered correct.
Four copyable templates
1) Verifiable solution with unit tracking:
Your role: physics assistant. Solve the following problem STEP BY STEP: [problem]. Write all the givens and the desired result in SI units. Show how the unit is moved at each step. Confirm the result with an executable Python (NumPy) code. Finally, evaluate in a sentence whether the order is physically plausible.
2) Fixed/data confirmation:
Write the value of [physical constant or material property] and its SI unit. Cite the source (e.g. CODATA, manual). If you don't know the exact value, say "verify this value from a reference" and DO NOT FIT. Also state your uncertainty (if any).
3) Checking my own solution/code:
Below is my solution/code. Check each step and its unit. If there is an ERROR, show in which step and why; If it is true, say "true". Especially check unit consistency and order. Write a new solution; Just check mine. My solution: [here]
4) Physical sanity check request:
Give 3 independent physical checksums (don't do the calculation) for the following result: dimension/unit analysis, boundary state behavior, conservation law check. For each, write in one sentence what I should check in this result.
Weak prompt / Strong prompt
Weak: "Calculate how long a ball stays in the air."
Result: Missing data, unclear unit, single line answer; cannot be verified, errors cannot be caught.
Strong: "For a ball thrown from the ground with a vertical speed of 20 m/s, calculate the time it stays in the air STEP BY STEP by taking g = 9.81 m/s², ignoring air resistance. Show the units at each step. Confirm the result with a Python code and write down what the time will be if the speed doubles as the limit case."
The result: Auditable steps, unit tracking, executable verification, and physical counterchecking.
Common mistakes
- Neglecting the unit. Considering a number without units is the most common and expensive mistake in physics. Every number has a unit.
- Not checking the rank. What is the result to the power of 10? The mass of a human is 70 kg, not 70 g or 70 tons. Rank checking takes seconds, saves hours.
- Not validating constants. A made-up constant is the most insidious hallucination because it seems plausible.
- Trusting the code without running it. Even if the code written by the AI "looks right", it should not be considered definitive until it is run and the result is checked.
- Not performing physical checks. Are energy, momentum, charge conserved? Are the boundary states as expected? Without these checks, the result remains incomplete.
Caution: Always verify a result before exporting it to a report, a student, or a design. In physics, an incorrect value grows exponentially in the chain of measurement, design, and security. Unverified AI output is never a substitute for confirmation from a competent physicist.
In summary
AI is a powerful drafting, coding and editing tool in physics; but it is not a calculator, physics engine or reliable reference. The biggest risks are unit/order errors, made-up constants, and spurious references. So the basic discipline is clear: define the problem with its units, solve it step by step and with unit tracking, verify by running the code, do physical counterchecking, verify constants and attributions. We will deepen this validation mindset in each unit throughout this module.
Application task
Choose a physics problem of medium difficulty from your field (a kinematics, energy, or electricity question). Solve it to AI with the 1st template, unit tracking. Then verify the result in an independent way (running Python code, dimensional analysis, or manually). Make at least one order check and one boundary condition check. Write down your findings in 5-6 sentences: where did the AI work, where was validation needed, was there a unit/order error?
checklist
- [ ] I clearly defined the problem with the given items and the desired unit.
- [ ] I asked AI for step by step solution and unit tracking.
- [ ] I verified the result with executable code or manually.
- [ ] I checked that the rank is physically reasonable.
- [ ] I have made at least one physical check (conservation, limit state).
- [ ] I have confirmed the constants I use from a reliable source.
- [ ] I have not passed on any unverified results.