Why the work stays at the torch
Welding is judgment plus hand skill, repeated in places a machine cannot easily reach. A welder lays out and fits the joint, tacks it, strikes an arc, then reads the puddle and adjusts heat, angle and travel speed as the metal moves. Parts warp. Gaps open. Drawings and reality disagree. Those corrections happen second by second, with gloved hands on the work.
Robotic welding is not new, and it is genuinely good at one thing: the same seam, in the same fixture, thousands of times. That is why the robotics tier for this job reads as fixed automation. A cell welds what is brought to it. It does not climb scaffolding, crawl into a tank, weld overhead on a pipe rack, or cut apart a cracked bucket on a job site and build it back. Repair work, field work and one-off fabrication stay with people because the setup cost of automating a single weld is higher than the weld.
The market reflects that. About 416,210 welders, cutters, solderers and brazers were employed in the United States, with median pay of $53,750 a year (BLS, 2025), and BLS projects employment to grow 2.4% between 2025 and 2035. That is a trade with steady demand, not one falling off a cliff. The pressure shows up in a different place: shops that move high-volume production into cells hire fewer people for repetitive bench welding, which trims some of the simplest entry-level work.
What AI does, what it assists with, and what people keep
The automated slice of this job is thin. Our coverage score, which measures the share of task time AI can handle today, sits at 4 out of 100, and the share of task time where AI does the work on its own is 0%. What sits there is desk-side, not arc-side: logging weld parameters and producing paperwork around a job. You can read how that figure is built on the coverage method page.
The assisted share is 1%. This is where vision systems and software add real value: seam finding and parameter suggestions inside a robotic cell, and first-pass screening of weld images for porosity, undercut or cracking before a qualified inspector signs anything off. The tool narrows where to look. A person still calls it.
Everything else is yours. The share of task time that still needs a human is 99%. Fit-up and tacking on parts that do not match the print, position welding in tight or overhead spots, grinding and finishing to a spec, and diagnosing why a weld keeps failing all sit here. So does the safety judgment around hot work, confined spaces and ventilation.
What has actually been tested
No one has published a head-to-head test of an AI system against a qualified welder across this job’s real task mix. Our evidence grade for quality parity is D, and a D grade means not measured, so we publish no parity number for welding. Robotic cells have decades of production data behind them, but that data covers fixtured, repeatable seams, not the varied work most welders are paid for.
A test that would settle it is easy to describe and hard to run: a mixed set of joints, materials and positions, including field fit-up and a repair weld, completed by an automated system and by journeyman welders, with every weld inspected to code by an independent third party. Until something like that exists, treat confident claims in either direction with care. The quality parity method page explains how we grade evidence, and the full approach is set out in our methodology.
When this could change
Most likely after 2046 (8 in 10 of our scenarios). The replacement-year method page explains what that window measures and how it is modeled.
Two things could pull it earlier. Cheaper collaborative welding arms with vision-based seam finding are reaching small shops that could never justify a full cell, and programming them is moving from code to demonstration. A tight supply of experienced welders also pushes fabricators to automate the runs they can, rather than leave them unwelded.
Two things hold it back. Almost all of this job is physical, so progress depends on robot hardware that can move to the work and handle unfixtured parts, not on better language models. And welds that carry load are governed by codes and inspection: a machine-made weld on a pressure vessel or a structural joint still needs a qualified person to certify the procedure and the result. Our guide to humanoid robots and physical jobs covers why that gap is wider than demos suggest.
How to stay needed in the trade
Lean into the parts of the job a cell cannot take to the site. Field fit-up and repair welding, position work in awkward or confined spots, and weld inspection and code compliance are the three worth building a career around. Short runs, odd materials and broken equipment keep coming, and they do not arrive in a fixture.
Two skills raise your floor. First, certifications tied to codes and processes you can prove under test, which is what buyers and inspectors actually price. Second, robotic cell setup and programming: the shops buying arms need someone who can fixture a part, teach the path, tune parameters and tell a bad weld from an acceptable one. That person is usually a welder, not a programmer.
What to do: ask your employer who sets up and troubleshoots their welding cells, and put your name forward for that work before the robot arrives.
If you want to see how close work compares, look at welding, soldering, and brazing machine setters, operators, and tenders, structural metal fabricators and fitters and machinists. You can also put any two of them side by side on our compare page.
For the wider picture, see the metal workers and plastic workers family, the manufacturing sector page, our list of jobs that mostly need a person, and the guide on AI and trades careers. Every scored job is searchable in the rankings.