Responsible AI Disclosure

1. Why We Publish This

Regulators have started citing companies whose AI-generated procedures entered the compliance system without adequate human review. The lesson from those cases is not that AI is dangerous. It is that a document which merely looks like a safety procedure is not the same thing as a safety procedure, and that no regulator, court, or injured worker will accept "the AI wrote it" as an answer.

We sell an AI-native safety platform. That is precisely why we are direct about this: AI makes safety teams faster, and it exposes organizations that let speed replace judgment. This page explains the line we draw in our own product and the line we urge every customer to draw in their program.

2. What AI Does in ECSafety

Aura AI™ drafts, ranks, flags, and explains. It pre-fills forms, triages incidents, drafts investigation reports, identifies hazards in photos, surfaces ranked recommendations, and runs gap analysis against your program and the regulations that apply to you. Every answer is grounded in your own safety data and Safety Docs, with citations you can check.

What Aura does not do is decide. Safety-critical outputs are drafts until a person with the authority and competence to approve them does so. That is a design decision, not a limitation we apologize for.

3. A Warning on AI & Rule-Setting

The highest-risk use of AI in safety is rule-setting: asking a model to write the policies, procedures, permits, rescue plans, and training that people will rely on when it matters. Treat every AI-drafted rule as a starting point that has not yet earned the right to govern work.

The dangerous failure is rarely an obvious hallucination. It is the confident draft that is mostly right and quietly missing the part that matters: the energy source a lockout procedure does not isolate, the rescue plan a fall protection policy never mentions, the trigger height that differs in your state, the verification step that proves a machine is actually de-energized before someone reaches in. A polished document with a gap like that is worse than no document, because it creates the appearance of a control where none exists.

• An AI-drafted procedure is not compliant until a competent person has checked it against the applicable standard, the manufacturer's instructions, and the way the work is actually done.
• Editing for grammar and adding a logo is not review. Review means checking the legal requirement, the hazard controls, and the field reality.
• Generic content produces generic protection. A procedure that does not name your equipment, your sites, and your crews is awareness material, not a rule.
• If a document would matter after an injury, an inspection, or a claim, it must not go live on AI output alone.

4. Jurisdiction Matters

"Write a compliant policy" is not a meaningful instruction, because compliance is always compliance with something specific. Federal OSHA, a state plan such as Cal/OSHA, the municipal requirements of the cities you build in, and consensus standards like ISO 45001 can each impose different obligations on the same task. A rule that is correct in Texas can be a violation in California.

This is why Aura is built to reason about the jurisdictions your sites actually operate in, and why our gap analysis checks your program against federal, state, and municipal requirements rather than a single generic rulebook. It is also why any AI-drafted document, from any tool, should state the jurisdiction it was written for, so the reviewer knows what to check it against.

5. Completion Is Not Competence

A completed training module proves someone progressed through content. It does not prove they can inspect a harness, lock out a machine, test an atmosphere, or stop unsafe work when they see it. AI makes training content cheap to produce, which makes this confusion easier to fall into at scale.

Verify competence where the work happens: demonstrations, supervisor observations, field checks, and scenario questions tied to your actual equipment and sites. ECSafety gives you the records and workflows for that verification, but the verification itself is human work, and it should stay that way.

6. You Own the Program

Safety law judges employers by the reasonable steps they took to protect people, not by how quickly their documents were produced. Responsibility for your safety program cannot be delegated to software, ours included. If an AI-drafted policy is wrong and work proceeds under it, the organization that adopted the policy owns the outcome.

We build ECSafety so that ownership is easy to exercise and easy to prove: named approvers, version history, and an audit trail that shows who reviewed what, when, and what changed. AI accelerates the work. Accountability stays with people.

7. The Guardrails We Build In

• Grounded, cited answers: Aura answers from your safety data, your Safety Docs, and the regulatory sources that apply to you, with citations. It is not a general chatbot with an opinion.
• Human approval on safety-critical output: drafted procedures, incident reports, and corrective actions route to a person for review before they take effect.
• Version control and audit trail: every document keeps its history: who approved it, when it changed, which version was in effect on a given day, and who received it.
• Review prompts, not silent automation: where an output is high-consequence, the product says so and asks for a competent reviewer instead of pretending certainty.
• Continuous regulatory learning, human-checked: new OSHA, state, and municipal guidance is folded into Aura as it publishes, and our team reviews how it is applied.

The data-protection side of these commitments, including what we do and do not train on, lives in the AI Policy.

8. Controls We Recommend in Your Program

Whether or not you use ECSafety, we recommend the same handful of controls for any AI that touches your safety system:

• Inventory your AI use. Know where AI is drafting safety content in your organization, including inside vendor tools, before an inspector tells you.
• Classify output by consequence. A hydration poster and a confined space rescue plan are not the same risk. Scale the scrutiny to the consequence of being wrong.
• Name the jurisdiction. Every compliance-related draft should state the regime it was written for.
• Require competent review. The reviewer must understand the hazard, the law, and the operation. For high-consequence procedures, that often means maintenance, engineering, or outside specialists, not just the person who wrote the prompt.
• Record the sources checked. Keep a trail showing critical content was verified against the standard, the manufacturer's instructions, and your own incident history.
• Test rules against the work. Before rollout, ask whether a crew could use the document to do the job safely on your site. If not, it is not ready.
• Audit on a cadence. Laws, equipment, and crews change. Re-review AI-assisted content periodically, not just when something goes wrong.

9. The Practical Standard

One sentence covers nearly every case: if a document would matter after an injury, an inspection, an enforcement action, a claim, or an investigation, it does not go live without competent human review. That is the standard we design for, and the standard we recommend you hold every AI tool to, including ours.

10. Further Reading

• EHS Today, "When AI Writes the Safety Policy, Who Owns the Mistake?" (July 2026), on the regulatory and due-diligence risks of unreviewed AI-generated safety content.
• OSHA training requirements and resources: the employer remains responsible for compliance with applicable standards.
• CCOHS on due diligence in occupational health and safety.

11. Contact

Questions about how AI is used in ECSafety, or about setting up the review controls above in your account, go to info@ecsafety.ai. We would rather answer a hard question before rollout than after an incident.