Every year, hundreds of millions of steel boxes move through the world’s seaports, and until recently a customs officer’s best tool for deciding which ones to open was experience and paperwork. That is changing fast. AI container scanning — software that reads X-ray images and gate camera footage automatically — is now doing the first pass of inspection at ports from Cameroon to Rotterdam, flagging the containers worth opening and waving the rest through.
The shift matters because the volumes are enormous. According to UNCTAD’s Review of Maritime Transport 2025, ports worldwide handled 920 million TEUs (twenty-foot equivalent units) in 2024, a 6.9% jump on the previous year. No customs service on earth has the staff to physically open even a meaningful fraction of that.
What is AI container scanning?
AI container scanning is the use of machine learning models — usually computer vision — to interpret images of shipping containers and automatically flag anomalies, contraband, damage, or mismatches with the declared cargo.
It sits on top of hardware customs services already own. Most large ports run non-intrusive inspection (NII) systems: high-energy X-ray or gamma-ray scanners that produce a grainy grey image of a container’s contents without anyone breaking the seal. Historically a trained officer studied each image on screen. The AI layer now looks first.
Two distinct jobs are involved, and they are often confused:
- Security and revenue screening — does the cargo match the declaration, and is anything illegal hidden inside?
- Condition and identity checks — which container is this, and is it damaged?
Both are being automated, but by different systems and usually by different vendors.
How do customs officers use AI to scan shipping containers?
A container passes through a drive-through or rail-mounted X-ray scanner, which produces a density image in seconds. An AI model trained on millions of previous scans compares that image against the declared HS commodity code, highlights regions that look inconsistent, and assigns a risk score. The officer reviews only the flagged images.
The Port of Douala in Cameroon is one of the better-documented examples. Writing in WCO News (Issue 2, 2026), the World Customs Organization describes how Douala moved from targeted risk-based checks in 2016 to near-systematic scanning of both imports and exports. Four high-capacity scanners each process up to 100 containers per hour, and the images feed a centralised platform called D-TECT that normalises data from several different scanner manufacturers.
The results reported are specific: in 2025, roughly 90% of declarations requiring correction in the port’s Littoral 1 customs sector were identified through scanner imagery analysis rather than physical inspection. Average declaration processing time for scanned containers was 1.2 days.
The WCO notes the system now applies AI to image interpretation, automated anomaly detection, and predictive modelling against HS codes — and connects scanner output to the terminal operating system and the national single window, so a flagged container can be held before the truck even arrives.
What can AI actually detect inside a container?
Current systems reliably detect cargo that does not match its declaration — the wrong commodity, the wrong quantity, or a container declared empty that is not. Detecting a specific concealed object, such as a weapon buried in a legitimate load, remains much harder and is where false positives concentrate.
Smiths Detection’s iCMORE family is a good illustration of how narrowly these models are scoped in practice. Rather than one algorithm that finds “anything suspicious,” iCMORE ships as separate trained detectors — including iCMORE Non-Empty Container, which does nothing but identify containers falsely declared empty, a common trick for avoiding duty and moving stolen goods. Each detector plugs into X-ray systems the operator already runs.
That narrowness is deliberate. A model asked to find one well-defined thing can be tuned to a usable false-positive rate; a model asked to find everything cannot.
Imaging hardware is improving alongside the software. The EU-funded MULTISCAN 3D project, documented by the European Commission’s CORDIS service, developed static 3D X-ray scanning that completes a container scan in three to four minutes. Three-dimensional data helps classification considerably, because shape and density together are far more distinctive than a flat silhouette.
Beyond security: AI inspection at the terminal gate
The second use of AI container scanning has nothing to do with contraband and everything to do with money — specifically, who pays when a container turns up dented.
At the terminal gate, arrays of cameras photograph every visible face of a container as the truck rolls through, without stopping. Optical character recognition reads the BIC and ISO codes stencilled on the box, and a computer vision model classifies visible damage: holes, punctures, patched panels, oxidation, dents. The result is a timestamped photographic record of the container’s condition at the exact moment it entered or left the terminal.
This is a live commercial market, not a pilot. In January 2026, WorldCargo News reported that Finnish terminal automation supplier VISY released ADDS 2.0, a new version of its Automatic Damage Detection System that inspects every container movement and integrates with the company’s existing OCR and gate operating system products.
The payoff is straightforward: damage disputes that used to be resolved by argument are now resolved by photograph, and gate transactions get shorter because nobody is filling in a paper condition report.
Does AI container scanning actually speed up trade?
Yes — the clearest gains are in throughput and clearance time rather than detection accuracy. Automating the first-pass review means officers spend their time on flagged containers instead of all containers, and gate automation removes a manual stop from every truck visit. The Douala data showing 1.2-day average clearance for scanned containers is the sort of result operators are chasing.
It is worth being precise about where the speed comes from. AI is not scanning faster than the hardware allows; the scanner still takes the time it takes. What changes is the queue behind it — the human review bottleneck largely disappears. That same “machine triages, human decides” pattern is showing up across logistics, from cutting cargo ship turnaround time in ports to robots picking and packing orders in warehouses.
Where AI container scanning still falls short
Three limitations come up repeatedly, and none are close to solved.
Some containers cannot be scanned at all. The WCO points out that flat racks, open-top containers, tank containers, and outsized shipments generally cannot be presented to a scanner and still require physical inspection by customs officers.
Training data is unbalanced. Genuine smuggling attempts are rare relative to legitimate cargo, so models see very few real positive examples. This is the classic rare-event problem, and it is why detectors for narrow, common infractions perform far better than general-purpose threat finders.
Adversaries adapt. Unlike a machine predicting equipment failure, container screening faces opponents who actively study what the scanner misses. A model trained on last year’s concealment methods degrades as methods change, which makes retraining an operating cost rather than a one-off project.
What this means for shippers and importers
For anyone moving freight, the practical consequence is that declaration accuracy now matters more than it used to. A mismatch between the paperwork and the X-ray image is exactly what these systems are best at catching, and it is caught automatically rather than by chance.
Two habits are worth adopting:
- Get HS codes right. Predictive models compare the image against the declared commodity code. A sloppy or optimistic classification is now a reliable way to get flagged.
- Photograph containers yourself at handover. Terminals increasingly hold their own timestamped damage evidence. If you do not have equivalent records, you will lose disputes you might previously have split.
The broader trend is clear enough: the container that used to move on trust and paperwork now moves under an image that a machine has already read.
Frequently Asked Questions
What percentage of shipping containers are scanned?
It varies enormously by port and country. Historically only a small share of containers were scanned or physically inspected, but several ports have moved toward near-systematic scanning — the WCO describes the Port of Douala as achieving nearly complete scanning coverage of imports and exports, excluding container types that cannot physically be scanned.
Does AI open shipping containers?
No. AI container scanning is non-intrusive: it analyses X-ray, gamma-ray, or camera images without the container being opened. The decision to physically open and unstuff a container is still made by a human customs officer, usually because the AI flagged an anomaly.
Can AI detect drugs or weapons in a container?
Sometimes, but this is the hardest case. Systems perform best on well-defined, relatively common anomalies — falsely declared empty containers, commodity mismatches, undeclared quantities. Detecting a small concealed object inside a legitimate dense load remains difficult and generates significant false positives.
What is automatic damage detection at a port gate?
It is a camera and computer vision system that photographs every face of a container as a truck passes through the terminal gate, reads the container’s BIC and ISO codes with OCR, and automatically classifies visible damage such as dents, holes, patches, and corrosion. The output is a timestamped condition record used to settle damage claims.
Which companies build AI container scanning systems?
Smiths Detection (the iCMORE algorithm family) and Rapiscan Systems are among the established names in X-ray-based cargo screening. On the gate automation and damage inspection side, suppliers include VISY, Camco, and AllRead. Customs authorities typically layer these on top of scanner hardware they already operate.
