AI Testing Gap Threatens Safety of Plastic Plants

As federal officials signal a push for faster nationwide adoption of artificial intelligence—including President Trump’s recent announcement of a plan to establish an “AI Corps”—a question has emerged in the manufacturing sector: how well can companies track the performance of AI systems once they transition from testing to daily production use?

One voice raising these questions isBotGaugeChief Executive Officer Pramin Pradeep,The company provides AI-driven quality assurance tools for engineering teams. Pradeep believes that many industrial AI deployments are validated once before release and rarely re-evaluated, even as underlying models, sensors, and vendor software continue to evolve.

Pradeep says the blind spot is that companies test AI like traditional software—testing it once before launch and never again. “AI systems do not remain static. Models drift, sensor data shifts, vendors push updates, and agents take paths that no one has scripted.”

It is worth noting that BotGauge’s products are built around continuous AI monitoring, so Pradeep’s assessment reflects the viewpoint of a company with a commercial interest in this issue, rather than an independent audit of industry practices. That said, the broader phenomenon he describes—AI models experiencing “drift” when real-world conditions differ from their training data—is a well-recognized problem in machine learning research and is not unique to his claims.

Why plastic manufacturing illustrates this concern

Pradeep points out that plastic processing is an example of an industry operating within particularly narrow tolerances. He states that melt temperature, pressure, cooling time, and resin moisture content all fall within tight ranges, leaving little room for error. He believes that AI systems optimizing these variables can perform reliably most of the time, while occasionally producing defective outputs without triggering warnings.

For illustrative purposes, Pradeep describes a hypothetical scenario in which a factory deploys artificial intelligence to adjust injection molding parameters. The system passes acceptance testing under a specific set of conditions. Several months later, a change in resin supplier introduces slightly different moisture content, causing the AI’s recommendations to deviate from safe operating ranges—despite no changes to the underlying code. This example is intended to illustrate general risks rather than cite a documented incident at a specified facility, and it has not undergone independent verification.

Transition from one-time testing to continuous monitoring

Pradeep contrasts this risk with traditional software testing practices, which typically assume that once a system is verified, it will remain consistent over time. He argues that this assumption breaks down in the era of AI explosion, as outputs and behaviors may change due to shifts in underlying data and models.

He recommends that manufacturers consider continuous testing rather than one-off testing, including how to record and track decisions, how the system handles incomplete or anomalous sensor data, how to monitor the impact of model or vendor updates, and whether there is a reliable mechanism for human operators to override AI-driven decisions.

Pradeep points out that several existing frameworks involve relevant principles, including the National Institute of Standards and Technology's AI Risk Management Framework, the ISO/IEC 42001 standard, and the EU Machinery Regulation. The extent to which these frameworks specifically endorse any particular testing methods (including Pradeep's suggestions) varies, and they can be interpreted differently by various stakeholders.

The Role of Human Expertise

Pradeep also emphasized that automated testing is not intended to replace human judgment. He stated that domain experts, such as process engineers, remain crucial for identifying when AI outputs are problematic in ways that automated systems may fail to detect on their own.

He stated that a process engineer would recognize a setpoint as suspicious long before it actually violated specifications. He described a model in which automated systems handle the scale of continuous testing, while human experts make the final judgments.

It is an open question for the industry.

The aforementioned claims have not been independently verified against specific event data. This account does not include perspectives from plastic manufacturers, process engineers, competitive testing suppliers, or regulatory agencies, who might offer alternative views on the actual extent of this testing gap and how the company is currently addressing it.

As policy discussions on artificial intelligence deployment continue at the federal level, including proposals such as an “AI Corps,” the broader issue of how to monitor industrial AI systems post-deployment remains unresolved and actively debated within the manufacturing sector.