Case file
- What happened: In October 2017 Kobe Steel admitted employees had falsified inspection data on strength and other quality properties of aluminium, copper and steel products sold to hundreds of companies, including major automakers and Boeing.
- Scale: An independent investigation found misconduct spanning decades and multiple divisions, in some areas reaching back to the 1970s; regulators ultimately identified no safety defects.
- Root cause: A normalised practice of giving customers the expected number rather than the measured one, sustained by delivery pressure, weak internal controls and no escalation path for bad news.
- The bill: JIS licences revoked, leadership resigned, a ¥100 million fine (about €800,000) and suspended sentences for former executives in 2021, on top of tens of billions of yen in losses.
Here is an uncomfortable observation from two decades of audits: the supplier that should frighten you is not the one with messy records. It is the one whose records are too clean. Kobe Steel held JIS certification and passed customer audits for decades while employees rewrote strength data on aluminium, copper and steel shipped to hundreds of customers. The instruments were calibrated. The certificates were current. The system deciding which number got written down was rotten, and no audit sampled it.
The situation
Kobe Steel was a foundation stone of Japan's manufacturing reputation. A century-old materials group spanning steel, aluminium and copper, it supplied extrusions, strip and forgings to around 500 companies – Toyota, Honda and Boeing among them – with JIS industrial-standard certification on the products and customer supplier audits layered on top. Everything rested on one assumption: a materials certificate from a Japanese giant is fact. For decades, it wasn't.
How it unfolded
A whistle-blower tip triggered an internal investigation, and on 8 October 2017 Kobe Steel admitted that inspection data on aluminium had been falsified – strength and elongation figures rewritten so lots met customer specifications. Within weeks the admission widened to copper and steel. A third-party committee reported in 2018 that the practice, in some areas, stretched back to the 1970s. Government ministries examined affected aircraft and vehicles and found no safety defects: the metal itself was usable; the paper was fiction. JIS licences were revoked, the president and other executives resigned, prosecutors indicted the company and former executives in 2020, and the Tokyo District Court closed the case in 2021 with a fine and suspended sentences.
Root-cause anatomy
The technique was mundane. When a lot missed specification, inspectors re-ran the test, re-sampled, or retyped the value until the certificate matched what the customer expected. Results travelled from instrument to report by hand. No read-only raw-data trail, no timestamped export, no independent review of laboratory records. The measurement system itself was fine – gauges calibrated, methods valid. The layer between instrument and certificate was built for editing, and for forty years it was used.
Organisationally, the pressure had no counterweight. Customer specifications and delivery dates were absolute, and sales pressure flowed straight into the laboratory. Rewriting a number became accommodating the customer – a craft handed down from older inspectors to younger ones and normalised over four decades. The one question that mattered – why does this plant always pass? – never travelled upward. Falsification here was not a hidden crime. It was an unremarkable Tuesday.
Where the quality system failed
Audit failed by design, not through negligence. A VDA 6.3 process audit or an ISO 9001 surveillance walks the process, interviews operators, samples the records the system produces. When falsification is normalised, the operator describes the falsified workflow with a straight face. The records reconcile with the procedure. The auditor certifies a consistent fiction. Sampling plans test whether a system exists and is followed; they assume the record is honest. No clause in any standard requires comparing a certificate with the instrument's own memory.
An audit can prove you own a ruler; it cannot prove you reported what the ruler said.
The other gates stayed silent as well. The PFMEA carried no failure mode for "inspector rewrites data", because detection rankings assume the detection system acts in good faith. The CAPA gate never fired – out-of-spec results were corrected out of existence before they became data. The 8D process existed only in counterfactual: containment happened at the wrong layer, by concealing the failure, so root cause never began. And on the customer side, hundreds of buyers, aerospace primes included, never independently re-tested incoming material properties. A certificate chain is transferred trust, not verified truth.
What would have caught it
None of this needed exotic tools. Blind retests of retained samples, sent to an independent laboratory and set against the shipped certificate, would have surfaced the practice inside a year – one mismatch is a finding, a pattern is a scandal. So would raw-data forensics: read-only, timestamped instrument exports compared with reported values. Hand-typed numbers hugging the lower specification limit leave a signature. So does a plant that never breaches a limit and never wobbles in capability; real processes drift. Separate the duties: whoever absorbs the delivery pressure must not hold the pencil that writes the certificate. And build an escalation route where reporting an out-of-spec lot earns credit rather than punishment, then run unannounced audits that pull instrument logs and retained-sample stores instead of the prepared binder.
My take
I was nowhere near Kobe's plants, but I have spent two decades on both sides of this table: building a greenfield QA organisation for a 900-plus-person automotive plant at SNOP, running supplier quality at Witte Automotive, auditing heavy industry as a consultant. The case matches my scars. Data integrity is not a laboratory problem; it is a governance problem with a laboratory address. My biggest defect-cost cut – 70% at SNOP – came from QRQC and A3 rooms where bad news arrived early and was rewarded, so problems surfaced at hour two rather than year twenty. That habit is also how you hold customer escalations at zero. My work in ethical hacking sharpens the point: a quality certificate and a log file are the same class of artefact, worthless the moment silent rewriting becomes possible. At every supplier audit since, my first request has been the instrument export, never the binder. The binder tells you what they want you to know. The instrument tells you what happened.
What this means on your floor
- Compare instrument raw exports against shipped certificates, unannounced, several times a year.
- Retain samples and retest them blind at an independent lab; keep the original data sealed.
- Treat statistically flawless results as a warning – real processes drift, reported ones don't.
- Make the honest number cheaper than the falsified one, or you have scheduled the next scandal.
Kobe Steel's instruments told the truth for forty years; nobody built a system that could hear them. Certificates sample documents and audits sample processes, and both rest on the one thing neither can measure: what an organisation does when the number comes back wrong at 15:40 on a Friday with a truck already loading. That moment is governed or it isn't. Governance, not the certificate, decides whether the measured number or the expected one reaches the customer – and by the time a court is involved, the answer was written decades earlier, one quiet edit at a time.