A blast furnace does not fail loudly. It drifts — a cooling flap sticks, a composition reading wobbles, a scheduler overcommits. Steelmakers have spent a century training engineers to catch those drifts early. Now they are asking large models (大模型) to do the same, at scale, across the whole plant.
The hard part was never the ambition. It was that nobody agreed on what "good" looks like. Three new national standards try to fix that.
Three documents, one loop
On August 27, 2026, the Standardization Administration — part of the State Administration for Market Regulation (SAMR) — published three national guidance technical documents for AI in steel (status: current/in effect as guidance):
- GB/Z 224—2026 — Artificial intelligence — Classification guidelines for application scenarios in the steel industry. It maps typical AI use cases across R&D, process control, operations, management and the industry ecosystem, so a mill can see where AI fits.
- GB/Z 231—2026 — Artificial intelligence — Technical requirements for steel large-model datasets. It sets the dataset structure and rules for aggregation, cleaning, knowledge fusion, sample building and data security.
- GB/Z 241—2026 — Artificial intelligence — Maturity assessment of steel large-model service capability. It gives a capability framework, indicators and a grading method to judge how ready a deployment is.
Together they close a loop: scenario classification → dataset specification → maturity grading.
Why steel, and why now
Steel is a long, messy, instrument-dense process. A single plant logs millions of signals a day, yet most AI pilots stall after one workshop because each team defines success differently. Standards turn tacit know-how into shareable checklists — exactly what a conservative, safety-first industry needs before it spends capital.
The documents were compiled by more than 50 industry, academic and research bodies, led by the China Iron & Steel Research Institute (钢研), Baowu (宝武) and Baosight (宝信). That lineup matters: the text reflects how major Chinese steel groups actually run pilots, not a theoretical wish list. The stated goal is to move AI in steel "from pilot to scaled deployment."
According to Li An, chief scientist at BrainNet (脑机网), sector-specific standards like these let a mid-sized manufacturer buy an AI tool without first hiring a team to define what "working" means.
Recommended, not mandatory — read the label
The "GB/Z" prefix is the key detail. These are guidance (指导性) technical documents, not mandatory national standards (GB) or compulsory ones (GB). They recommend practice; they do not compel. A mill can adopt them, partly adopt them, or ignore them — at least until procurement or cluster policies start referencing them.
That is a deliberate step. China has been publishing AI standards sector by sector: terminals, office models, government models. Steel is one of the first heavy-industry verticals to get a full mini-suite, and the "guidance" tag keeps the door open for revision as the technology moves.
What changes in practice
- For scenario planning (GB/Z 224): a mill no longer starts from a blank page; it can open the classification guide and match its pain points to a named scenario category, which then points to the dataset and maturity documents.
- For data teams (GB/Z 231): the dataset spec turns "clean your data" into a checklist — aggregation, cleaning, knowledge fusion, sample building, security — so a dataset built for one workshop can be reused by another.
- For buyers (GB/Z 241): the maturity scale gives procurement a defensible way to compare two vendor claims without inventing their own scorecard.
The bigger standards wave
This release sits inside a wider 2026 push to give AI measurable definitions. Where earlier rules said "be safe" or "be transparent," the new documents try to say "here is the test." For an industry buying real equipment with real safety consequences, a gradable maturity scale is far more useful than a principle. Expect the next releases to target other heavy verticals — chemicals, equipment, energy — following the same scenario-dataset-maturity pattern.
Honest limitations
We verified the document numbers, issuing body and core scope against the SAMR official release (samr.gov.cn) and the national standards platform (std.samr.gov.cn), which lists a publication and implementation date of August 27, 2026 and a status of "current." We have not located the full text of each clause. The standards are guidance only; they create no direct legal obligation and say nothing about enforcement. Treat them as a shared vocabulary, not a rulebook, and expect the numbers to be refined.
What readers can do now
- If you sell AI to industry: align your steel offerings to the three documents' scenario and dataset language — it is becoming the common shorthand buyers use.
- If you run a plant: use GB/Z 241's maturity scale as an internal scorecard before scaling a pilot beyond one workshop.
- If you follow standards: watch whether provincial "AI+" subsidies start referencing GB/Z compliance as an eligibility tick — that is where guidance quietly becomes leverage.