Safety professionals are fine with AI drafting a report. They are not fine with AI deciding who gets disciplined or what caused an incident with nobody checking. Nearly half of organizations have no AI ethics policy to draw that line. You don't need a long one. A single page that says what AI may draft, what a named person must sign, and what AI never decides covers most of it.
ASSP's 2026 survey of nearly 1,000 safety professionals, Trust, Ethics and Governance, asked what worries people about AI in EHS. (Our overview of the full survey covers adoption and results.) The answers were specific. Concern was lowest for routine work and highest where AI output turns into action on a person.
What safety professionals are worried about
Each of these drew concern from 80% or more of respondents, with roughly half saying a great deal of concern:
- Less human judgment, ownership or accountability (85% concerned).
- Decisions based on inaccurate information, hallucinations or out-of-context judgments (85%).
- Poor or incomplete data, or misclassification, leading to unfair actions (85%).
- Data privacy and security (83%).
- Automatically generated discipline, interventions or risk-cause assignments with little or no human validation (80%, and 48% a great deal).
Experience softens the worry but doesn't remove it. About half of non-adopters express a great deal of concern about AI ethics in EHS, against about 3 in 10 at the most advanced organizations. Even the people furthest along want a person in the loop when the stakes are a worker's safety or record.
Governance hasn't kept up
Across all organizations in the survey:
- 49% have no AI ethics policy or guidelines.
- 55% have no AI policy for EHS at all.
- 40% have no AI governance measure of any kind.
Policies do grow with experience: 62% of fully integrated organizations have AI ethics guidelines, against 20% of those still in discovery. But that means most teams are using AI now, under rules they haven't written yet. Half of respondents also said they want training on validating AI output, which is the skill a policy depends on.
A one-page AI policy for an EHS team
Here is an outline you can adapt. It is written for a safety consulting firm, but an in-house team can swap "client" for "site leadership." It isn't legal advice; have your own counsel look at anything that touches discipline or personal data.
- What AI may do. Draft reports, observations and summaries from field notes and photos. Draft training content. Search and summarize past records. Suggest categories, trends and follow-up items.
- What a named, qualified person must do. Review and approve every report before it reaches a client. Approve any corrective action, its owner and its due date. Accept or reject evidence that a hazard was fixed. Approve training before learners see it. The approver's name and the date go on the record.
- What AI never decides. Discipline. Fitness for duty. Root cause of a serious incident. Whether a stop-work condition exists. AI output can inform these; a person makes them and owns them.
- Data. List the data you feed AI tools (photos, voice notes, incident records) and what you won't (medical, biometric, anything you would not show the worker). Confirm in writing that your vendors don't train their models on your data.
- Checking output. Anything AI drafts is checked against the source: the photo, the note, the standard cited. Citations to regulations are verified before they go out.
- Telling people. Clients know which parts of your deliverables are AI-assisted and that a qualified person reviews them. Workers know what is captured about them and why.
- Ownership. One person owns this policy and reviews it every six months, or when you add a tool.
For any AI-assisted record, ask: could I name the person who approved it, and would they defend it in a deposition? If not, the workflow is missing a step.
Build the stop into the workflow
A policy only holds if the software makes the review step the default rather than an extra click people skip under deadline. Look for three things in any tool you adopt:
- Drafts are drafts. Nothing AI produces goes to a client or a worker until a person confirms it.
- The source stays attached. The reviewer can see the photo or note that each line came from, so checking takes seconds rather than a rewrite.
- The approval is recorded. Who approved what, and when, is part of the record, not an email somewhere.
Where Safely Track fits
Safely Track is built this way. AI fills in the report from your voice notes and photos, and you review, edit and confirm it before it goes to the client. Corrective actions are closed by a person who accepts or rejects the proof of fix, not by the software. In the course builder, every slide links back to the deck it came from, and nothing reaches a learner until you publish.