The widening gap
There is a widening gap between what the technology world promises and what is actually being built on factory floors. You can see it at every industry conference: the keynote is full of ethical AI principles, responsible AI frameworks, and human-centred design. The expo hall is full of vendor booths selling black-box models that no one on the customer side can audit.
Industry 5.0, as a concept, is supposed to be the human-centric, sustainable, resilient evolution of Industry 4.0. It is supposed to put workers and ethics at the centre of the technology stack. In practice, what most vendors are shipping under the Industry 5.0 banner is the same Industry 4.0 automation with a new sticker on it.
Why ethical AI is harder than it sounds
Ethical AI is not a checklist. It is not a section in your privacy policy. It is not a slide in your investor deck. Ethical AI is a property of a system, and it has to be designed in from the start. You cannot bolt it on after the model is trained.
The hard parts:
1. Explainability is a feature, not a metric
Most enterprise AI systems can tell you what they predicted. Very few can tell you why. For a manufacturing planner who is about to override a forecast that affects a $2M purchase order, "the model said so" is not an acceptable answer.
A genuinely ethical AI system in an operational context has to be able to show its work: what inputs it weighted, what historical patterns it matched, what confidence band it assigned, and what the alternative predictions were. Without that, the planner is not making a decision. They are placing a bet.
2. Audit trails are not optional
If you cannot answer the question "who changed what, when, and why," you do not have an ethical AI system. You have a liability.
This is not a hypothetical. Regulators in both India and the US are moving toward requiring audit trails for any automated decision that affects employment, credit, or operational outcomes. The DPDP Act in India and the CCPA/CPRA in California are the leading edge. More is coming.
An audit trail means: every override, every approval, every model run, every input change captured with actor and timestamp. Not in a log file that someone has to grep through. In a queryable ledger that an auditor can pull a report from in five minutes.
3. Human-in-loop is not a marketing claim
A lot of vendors claim human-in-loop. Very few actually implement it. Human-in-loop means the human has a real choice at the decision point, with enough information and enough time to make a different decision than the one the AI proposed.
If the AI proposes an action and the human has 30 seconds to approve or reject it before the system moves on, that is not human-in-loop. That is rubber-stamping with extra steps.
Genuine human-in-loop means the AI proposes, the human reviews, the human can ask the AI to explain, the human can reject, and the rejection is captured with a reason. Anything less is theatre.
The conference keynote versus the factory floor
Here is the test. Next time you are at an industry conference and a vendor is talking about ethical AI, ask three questions:
-
"Can your system show me, for any decision it made in the last 30 days, what inputs it weighted and why?" If the answer is no, walk away. The system is not auditable.
-
"Can your system show me, for any override a human made, who made it, when, and what reason they gave?" If the answer is no, walk away. The system is not compliant-ready.
-
"When your AI proposes an action, how long does the human have to review it, and what information do they see?" If the answer is "they approve in one click" or "they see a confidence score," walk away. The system is rubber-stamping, not human-in-loop.
Most vendors will not be able to answer these questions cleanly. That is the gap. That is why ethical AI is harder than it sounds. It is not a principle. It is a property of the system architecture, and most systems were not architected with it in mind.
What a genuinely ethical operational AI system looks like
For manufacturing and supply chain specifically, an ethical AI system has four properties:
-
Explainable predictions. Every forecast, every exception, every recommendation comes with the inputs, the model family used, the confidence band, and the historical pattern matched.
-
Full audit trail. Every create, update, delete, and approve action is logged with actor, timestamp, resource ID, and event type. The log is queryable and exportable for compliance review.
-
Genuine human-in-loop. Every write action the AI proposes requires human approval. The human sees the proposal, the rationale, the alternatives, and can reject with a reason. No silent automation.
-
No training on customer data. The AI does not learn from your data to improve someone else's product. Your data stays in your workspace. Period.
These are not features. They are the minimum bar for calling an operational AI system ethical. Anything less is marketing.
The regulatory direction
India's DPDP Act 2023 is now in force. It requires consent, purpose limitation, data minimisation, and grievance redressal. For AI systems that process personal data, it also requires the ability to explain decisions and provide recourse.
The EU AI Act, passed in 2024, classifies AI systems by risk level. Operational AI in manufacturing is likely to fall into the "limited risk" or "high risk" categories depending on use case, which means transparency, logging, and human oversight requirements.
The US is moving more slowly, but the CCPA/CPRA in California, the Colorado Privacy Act, and the Virginia CDPA all require similar transparency and audit capabilities for automated decision-making.
The direction is clear: audit trails, explainability, and human oversight are becoming legal requirements, not best practices. The vendors who are shipping systems without these properties are shipping products with a shelf life.
What to do about it
If you are evaluating an operational AI system for your manufacturing or supply chain operation, run the three-question test above. If the vendor cannot answer, they are not ready for the regulatory environment that is arriving.
If you are building an operational AI system, build the audit trail, the explainability, and the human-in-loop in from day one. Retrofitting them is 10x more expensive than designing them in, and in some jurisdictions it will soon be legally required.
Ethical AI is not a principle. It is an architecture. Build it that way.
This article is part of prompt/ed, the Tensor Analytics education hub. Written for manufacturing practitioners, not AI enthusiasts.