AI in NDT Defect Detection: What It Actually Does (and Doesn't Replace)

Machine learning is showing up in inspection workflows — mostly as a faster first pass over PAUT C-scans and digital radiographs, not as a replacement for the certified inspector who signs the report. Here is what the technology can realistically do today, where it falls short, and how the compliance framework around it actually works.

How AI Actually Helps NDT Inspection Today

The methods that generate structured, image-based digital data are where machine-learning tools have found real traction — phased-array ultrasonic testing (PAUT), where the output is a C-scan, B-scan, or S-scan image, and radiographic testing (RT), where digital detector arrays and computed radiography produce a digital image file instead of exposed film. In practice, a trained model scans through a batch of C-scan images or radiographs and flags the regions most likely to contain an indication — a crack, porosity, lack of fusion, slag, or a geometric anomaly worth a second look — so a reviewer's attention gets directed to the highest-priority frames first. On high-volume jobs that kind of triage can meaningfully cut review time. It is not the same as autonomous defect detection or automated code-acceptance disposition, both of which remain firmly in human territory.

What AI Does Not Replace — the ASNT Level III Sign-Off

SNT-TC-1A and ASME Section V Article 1 require examinations to be performed and dispositioned by personnel qualified under an employer's Written Practice, with an NDT Level III holding final technical authority for procedure interpretation, personnel qualification, and disposition of results. That requirement doesn't reference software — it references a certified, accountable person. A model, however well it performs on a benchmark data set, doesn't hold a certification and can't be held accountable under SNT-TC-1A for a disposition decision. The realistic, industry-converging workflow is human-in-the-loop: AI flags candidate indications, a qualified inspector reviews every flagged region (and spot-checks unflagged regions), and the Level III retains sign-off authority on the final report.

Realistic Limitations of AI-Assisted Defect Detection

False positives and false negatives: every detection model trades reviewer time (false positives) against missed real indications (false negatives), and there is no single published industry-wide accuracy figure that holds across geometries, materials, and vendors. Domain shift: a model trained on one weld geometry, material grade, or scanner make often performs noticeably worse on a different one it wasn't trained on, and field conditions routinely fall outside a model's training distribution. No settled code-acceptance framework: standards bodies are still working through how AI-assisted acquisition and interpretation fit within existing personnel- and procedure-qualification requirements. Explainability: a Level III's disposition traces to a documented procedure and a named, accountable person; a model's "why did it flag this" answer is often much harder to defend in an audit or client dispute.

Where the Technology Is Heading

Standards committees within ASME Section V and various API and industry working groups are actively discussing how personnel-capability and procedure-qualification language should account for AI-assisted acquisition and interpretation. In parallel, the broader shift toward structured digital data capture — digital radiographs instead of film, encoded PAUT data instead of hand sketches — is what makes AI-assisted triage possible in the first place. The realistic near-term picture: AI as a productivity tool that helps a qualified inspector get through more data faster, not a tool that removes the qualified inspector from the process.

How This Connects to Atlantis NDT

Atlantis NDT doesn't market an in-house AI defect-detection product. Where Atlantis NDT has real, shipped capability that connects to this topic is the infrastructure any AI-assisted or fully manual inspection workflow still needs: the Atlantis NDT ERP document-control module — revision-locked procedures, role-based access to the current approved revision only, Level III/QA Manager approval workflow, and report numbering that permanently links every issued report to the procedure revision in effect at inspection time. On the certification side that AI doesn't replace, see Atlantis NDT training and the ASNT Level I/II/III certification guide. Inspection findings, however they were first flagged, still need a home in the asset's history — see the Atlantis NDT Digital Twin platform for how indications and Level III sign-off chains get consolidated over time. Procedures also still need to be written and qualified in the first place — see NDT technical procedure development.