Blog · · 5 min read

Can you trust AI with engineering numbers? How to keep invented specs out of a hardware project

AI drafts hardware designs fast, but it can also invent a part number or a load rating that looks perfectly real. Where invented numbers creep in, and the checks that keep them out of drawings, BOMs and orders.

An AI assistant can read a 90-page datasheet in seconds, size a bracket, list the parts for a prototype and write the order notes. That speed is real. So is the other side: now and then it writes a number that nobody ever measured, a part number that does not exist, or a rating that belongs to a different variant. The problem is not that this happens. The problem is that a wrong number looks exactly like a right one.

In hardware, a wrong number does not stay on the screen. It becomes a hole that is too small, a motor that stalls, or a box of parts that do not fit. This article shows where invented numbers come from, which ones to watch, and the checks that keep them out of a project.

Why AI invents numbers

A language model writes the most plausible continuation of a text. When the answer is in a document it has open, plausible and true are usually the same thing. When it is not, the model still produces something plausible: a part number with the right pattern, a torque value in the right range, a standard with a believable title.

Numbers are where plausible and true drift apart most easily. A bearing's load rating, a sensor's range or a resistor's power rating look the same whether they were copied from a datasheet or made up. Nothing in the text itself tells you which one it was.

The usual suspects

In engineering work, the same kinds of errors come up again and again:

  • Part numbers that almost exist. Manufacturers encode package, temperature range or voltage in the suffix. A model can combine a real base number with a suffix that was never made, or pick the variant with the wrong package.
  • Values from the sibling part. The stall torque of the motor with a different gearbox, the load rating of the next bearing size, the input range of the 5 V version of a 3.3 V sensor.
  • Unit and scale errors. N instead of kN, millimetres instead of inches, mA instead of A. These are older than AI: NASA's Mars Climate Orbiter was lost in 1999 because one piece of ground software reported thruster impulse in pound-force seconds where the rest of the system expected newton-seconds.
  • Stale facts. A part that has been discontinued, a price from last year, a standard that was revised.
  • Sources that were never read. A summary of a paper or a datasheet based only on a search snippet, presented as if the whole document had been checked.

Rule one: every number needs a place you can open

The single most useful rule is simple: a number only enters the design if it comes with a source you can open. That can be a page in a datasheet, a clause in a standard, a measurement, or a calculation stored in a file. "The AI said so" is not a source.

When no source can be found, the number does not get guessed. It becomes an open question, listed where a person will see it. A short list of open questions is far less dangerous than a confident document with three invented values hidden inside.

It also helps to mark how well a source was read: a value from the full datasheet is stronger than one from a search preview, so the second kind gets a "snippet only" note.

Checks that catch the rest

Sources reduce invented numbers, but they do not catch everything. A few cheap checks catch most of what slips through:

  1. Recalculate instead of trusting. Key values such as stress, current, mass or runtime are computed in a small script from inputs that have sources. If an input changes, the result is computed again rather than edited by hand. We showed what such a check found in a stress check on a jumping robot.
  2. Measure from the model. Masses, clearances and hole sizes come from the CAD geometry, not from the text that describes it. If the description says 8 mm and the model says 6 mm, one of them is wrong, and that gets resolved before anything is ordered.
  3. Confirm part numbers where they are sold. A part number counts as real when a distributor or the manufacturer lists it, with the right package, in stock or with a lead time. This is also where a discontinued part shows up.
  4. Sanity bounds. A printed plastic bracket holding two tonnes, a coin cell running a motor for a week, a 0402 resistor dissipating a watt: values far outside the usual range get flagged, even when they come with a source.
  5. Two sources for what matters. Values that decide safety or cost, such as the load on a fastener or the rating of a power supply, need a second independent source or a test.

None of these checks needs special tools, only discipline, and repetitive discipline is what AI does well when told to.

What AI is genuinely good at here

The same systems that occasionally invent numbers are also very good at catching inconsistencies, as long as they have the documents:

  • Finding the one table in a long datasheet that holds the value you need.
  • Comparing three candidate parts on the same ten criteria.
  • Spotting that the drawing, the bill of materials and the quote request disagree about a material or a quantity.
  • Re-running every calculation after a design change and listing what moved.

Used this way, AI turns "I think it's fine" into a list of what was checked against which source.

Where a person still decides

Some decisions should stay with a person even when every number has a source: anything that can injure someone, claims about certification or compliance, and the final decision to spend money. AI can prepare these decisions, show the sources and assumptions, and say what is still open. It should not make them alone.

How makemachine.ai handles it

On makemachine.ai, each step of a project, from concept to CAD, simulation and sourcing, delivers its results together with the sources it used, linked where they can be opened. Values without a source are listed as open questions instead of being filled in. Calculations are stored as files that can be run again, and every result names its assumptions. The project owner sees the open points and decides. To see what happens to a cost estimate when the real prices arrive, read why the first real bill of materials doubled the estimate.

The short version

AI makes hardware work faster, and it can invent numbers that look real. Treat its output like a capable new engineer's first draft: every number needs a source you can open, key values are recalculated, part numbers are confirmed where they are sold, and anything without a source stays an open question. The person who signs off still decides.

Want to see how makemachine.ai keeps a hardware project traceable from idea to order? Request early access.

Watch hardware get built by AI. Follow public projects on makemachine.ai, read every result in full and get an email when there is something new.

Get started See the live projects

More from the blog

Open-hardware projects built with AI