Table of Contents
Formal Verification Engineer Career: Checking AI's Proofs
The formal verification engineer career grew in 2026 because AI writes proofs faster than humans can check them. Here is the path, the pay, and where to start.
“It could be right. It could be nonsense. Who’s going to be able to check this?” That was Thomas Bloom, who maintains the Erdős Problems database, talking in mid-2026 about AI-generated mathematical proofs. It’s also the best one-sentence job description for the field that answers him. A formal verification engineer career is built on a single idea: instead of trusting a proof because a smart person read it, you rewrite it in a language where a computer refuses to accept an invalid step. The job existed in chip design and aerospace for thirty years. In 2026 it became the bottleneck for AI, and demand jumped.
Key Takeaways
- Formal verification is proving that a system does what its specification says, mechanically, with a proof checker rather than human review. Tools: Lean, Coq, Isabelle, TLA+, Dafny, and SMT solvers like Z3.
- The 2026 driver: AI produced the disproof of an 80-year-old Erdős conjecture in May, was correct on roughly 50% of runs, and by August was generating results faster than mathematicians could vet them.
- BLS has no “formal verification engineer” code. The closest official figures (May 2025): software developers $135,980 median, +10% to 2035; software QA analysts and testers $104,300, +6%; computer and information research scientists $140,300, +22%.
- Where the jobs actually are: chip and hardware companies, cloud infrastructure teams, aerospace and rail safety, cryptography and blockchain, and now AI labs.
- What a 13-year-old can do this year, free: finish Lean’s Natural Number Game in a weekend, then prove that the sum of two even numbers is even. That is genuinely the first rung.
What formal verification is, in one example
A specification says: this function returns the largest element of a list. A test checks the function on a few lists. A formal proof shows it returns the largest element for every possible list, including the empty one, the one with duplicates, and the one with a billion entries. Tests sample. Proofs quantify.
The tooling makes this practical. In Lean, you state a theorem and then write a sequence of steps; the kernel checks each one and rejects anything that doesn’t follow. There is no arguing with it, no “clearly,” no “the rest is routine.” Lean’s mathlib is a community-built library of formalized mathematics, open source and enormous. Coq and Isabelle do similar work with different ergonomics. TLA+, created by Leslie Lamport, verifies distributed system designs rather than code, and cloud teams have used it for years to find bugs in protocols before anyone writes the protocol.
The honest limitation, which matters for a kid considering this: a formal proof proves the code matches the specification. If the specification is wrong, the proof is worthless. Verification engineers spend a surprising amount of their time arguing about what the specification should say, which is a human problem no proof checker solves.
Why 2026 changed the demand curve
Three events stacked up.
On May 20, 2026, OpenAI announced that a model had disproved the Erdős unit distance conjecture from 1946. Nine mathematicians published remarks confirming it. But the model succeeded on roughly 50% of runs, and OpenAI did not publish failure rates. Human review does not scale to those odds.
By August 3, Quanta Magazine reported that OpenAI had announced ten more advances on August 1, Google DeepMind had resolved four problems and rediscovered nine forgotten ones, and mathematicians were now sitting on unvetted 100- to 200-page machine-written papers. Some teams began using proof-certification systems to triage.
Third, the research literature caught up. A 2026 arXiv paper, “From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier”, with contributors including Terence Tao, argued that LLMs can advance open problems specifically when paired with formal verification systems, and reported using formal proof systems throughout to ensure rigor. That is the thesis of an entire career in one paper: the model proposes, the checker disposes.
Meanwhile the older industries kept hiring. Formal methods have been standard in processor verification since the 1990s, and safety-critical avionics and rail software use them because certification standards require evidence, not testimony.
Skill to course: building a formal verification engineer career
| Skill you need | Where to learn it | Rough level | Why it matters on the job |
|---|---|---|---|
| Reading and writing proofs | High school proof-based geometry, then university discrete mathematics | Age 15+ | Everything else is notation on top of this |
| Propositional and predicate logic | Any intro logic course; MIT OpenCourseWare “Mathematics for Computer Science” | Undergraduate year 1 | Specifications are written in logic |
| A proof assistant | Lean’s free Natural Number Game, then Theorem Proving in Lean 4 | Age 13+ for the game | This is the daily tool |
| Type theory | Software Foundations (Coq, free online, Penn) | Undergraduate year 3+ | Explains why proof assistants work at all |
| Program semantics | A programming-languages course; Dafny tutorials | Undergraduate year 3 | Verifying code, not just math |
| Model checking | TLA+ video course by Leslie Lamport, free | Undergraduate or on the job | Distributed systems, protocols |
| SMT solvers | Z3 tutorials; constraint-solving assignments | Undergraduate or on the job | The automation layer under everything |
| Hardware verification | A digital design course plus SystemVerilog assertions | Undergraduate year 3+ | Chip companies are the largest employer |
Notice the ages. A motivated 13-year-old can start row three today. Rows one and two are the actual prerequisites for everything else, and they are just mathematics.
Pay and demand, labelled honestly
There is no BLS category for this role, so nobody has a defensible median salary for it specifically. What we have:
| Closest official category | Median pay (May 2025) | Projected 2025–2035 | Jobs (2025) |
|---|---|---|---|
| Software developers | $135,980 | +10% | ~1,700,000 |
| Software QA analysts and testers | $104,300 | +6% | ~187,600 |
| Computer and information research scientists | $140,300 | +22% | ~38,600 |
Formal verification engineers are counted inside the developer and research-scientist buckets, and industry reporting puts specialist verification salaries at or above the developer median, often well above in chip design. Treat any specific number you read as an estimate from job postings, not a measurement.
The demand picture is unusual: very small headcount, very low supply, high consequence. There are only a few thousand people worldwide who can formalize serious mathematics in Lean. That imbalance is why a teenager with three years of Lean contributions on GitHub gets interviews that a generic computer science graduate does not.
What a 10–15-year-old can do this year
Play the Natural Number Game
It’s free, browser-based, and teaches you to prove basic arithmetic facts in Lean by clicking and typing. Most motivated 13-year-olds finish the first levels in an afternoon. The moment your kid proves that addition is commutative and the system accepts it, they have done formal verification. Our parent-level explainer on how proof assistants turn math into code covers what’s happening under the hood.
Prove something about their own code
Take a tiny function they wrote, maybe “find the biggest number in a list,” and ask them to write down what it should do for every input, including the empty list. Most kids discover a bug in their own code during this exercise. That’s the lesson: specification first.
Break a specification on purpose
Give them a rule like “the password must contain a number” and ask them to find an input that satisfies the rule and is still terrible. This trains the exact instinct verification engineers use all day: the spec is the enemy, not the code.
Follow one mathlib pull request
The Lean community reviews contributions publicly on GitHub. Reading one discussion thread shows a teenager what professional mathematical rigor looks like as a conversation, which no textbook conveys.
What not to do
Don’t start with type theory or category theory because someone online said it’s the foundation. It is, and it’s also the wrong entry point. Start with the game, get the feeling of a machine rejecting a step, and let curiosity pull them toward theory. And don’t skip ordinary proof-writing practice: a kid who can’t write a clear paragraph proof will not write a clear Lean proof either. Our piece on how mathematicians verify an AI proof is a good read-together.
What to Watch For Over the Next 3 Months
- Week 4: One completed Lean proof, however small, and your kid can explain why “the computer wouldn’t accept it” is information rather than an insult.
- Month 2 red flags: They want the proof assistant to just tell them the answer. They get frustrated by error messages rather than curious. They skip the specification step. The third is the real warning sign, because specification is the job.
- Month 3 self-check: Ask them to state, in plain English, one property of a program they’ve written that should be true for every possible input. If they can state it precisely, they’re doing the hardest part of formal verification already.
Frequently Asked Questions
Is this a real job or an academic niche?
Both, and it has been for decades. Processor verification teams at chip companies, cloud infrastructure teams verifying distributed protocols, avionics and rail software suppliers meeting certification requirements, and cryptography implementations all employ formal methods engineers. The 2026 change is that AI labs joined the list.
Does my kid need a PhD?
For research positions at labs, usually. For industrial verification, a bachelor’s in computer science with strong logic coursework plus demonstrable proof-assistant work is often enough, and a master’s is common. The portfolio matters more here than in most fields because the work is publicly inspectable.
Won’t AI just do the verification too?
Partly, and that’s already happening: models are being used to generate candidate formal proofs which the checker then validates. Notice the structure. The checker is not a language model; it’s a small, trusted kernel. Someone has to write specifications, design the proof architecture, and decide what “correct” means. That’s the durable part of the job.
What if my kid is strong in math but doesn’t like programming?
This field sits exactly on that boundary, and many people enter from the mathematics side. Lean feels more like mathematics than like software engineering, especially in the mathlib world. It’s one of the few careers where “I like proofs more than apps” is an advantage.
How is this different from software testing?
Testing samples inputs and looks for failures; verification proves a property holds for all inputs. Testing finds bugs, verification establishes absence of a class of bugs, within the limits of the specification. Both are employed, and QA analyst roles are growing more slowly (+6%) than developer roles (+10%), which tells you where the market values depth.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- OpenAI. (2026). “Model disproves discrete geometry conjecture.” May 20, 2026. https://openai.com/index/model-disproves-discrete-geometry-conjecture/
- Alon, N., Bloom, T. F., Gowers, W. T., et al. (2026). “Remarks on the disproof of the unit distance conjecture.” arXiv 2605.20695. https://arxiv.org/pdf/2605.20695
- Quanta Magazine. (2026). “Why the legendary Erdős problems are falling to AI.” August 3, 2026. https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/
- Jiang, E., Liang, X., Zhang, Y., et al. (2026). “From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier.” arXiv 2607.07779. https://arxiv.org/pdf/2607.07779
- Bureau of Labor Statistics. (2026). “Software Developers, Quality Assurance Analysts, and Testers.” Occupational Outlook Handbook, May 2025 data. https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm
- Bureau of Labor Statistics. (2026). “Computer and Information Research Scientists.” May 2025 data. https://www.bls.gov/ooh/computer-and-information-technology/computer-and-information-research-scientists.htm
- Lean FRO. Lean theorem prover documentation and
mathlib. https://lean-lang.org/ - Lamport, L. TLA+ home page and video course. https://lamport.azurewebsites.net/tla/tla.html
