No evidence can establish that AI will never fully replace humans, because "never" cannot be tested. What can be checked is where current rules require a person to stay involved, and those rules are narrow and specific. California's automated decisionmaking regulations count a decision as human-involved only if the reviewer understands the tool's output, reviews it, and has authority to change the decision, and Quebec's private-sector privacy law gives people affected by purely automated decisions a way to submit observations to a staff member who can review them.
Key takeaways
- California's approved regulations define automated decisionmaking technology (ADMT) as technology that replaces or substantially replaces human decisionmaking, and ADMT duties apply to significant decisions from January 1, 2027.
- Under the California text, a human reviewer must know how to interpret the output, review and analyze it, and have authority to change the decision.
- Quebec's Act respecting the protection of personal information in the private sector requires notice of a decision based exclusively on automated processing and a chance to present observations to a staff member able to review it.
- Canada's federal Directive on Automated Decision-Making covers both fully and partly automated decisions, including cases with a human in the loop.
- These rules regulate how automated decisions are made, and none of them says that people must keep any particular job.
Why "never" is the wrong test
Claims that AI will never replace people, or that it surely will, are predictions about the future. The labor statistics published by agencies such as the U.S. Bureau of Labor Statistics describe employment by occupation and projected change, not limits of what machines can do. Law offers a different kind of evidence, because it states in writing when a human must be present. Reading those texts shows that the requirement attaches to a decision about a person, not to an occupation, and that it comes with conditions.
California: what counts as a human decision-maker
The California Privacy Protection Agency's regulations under the California Consumer Privacy Act were approved by the state's Office of Administrative Law on September 22, 2025, and took effect January 1, 2026. The ADMT article requires a business that uses ADMT for a significant decision to comply no later than January 1, 2027. A significant decision is one that results in the provision or denial of financial or lending services, housing, education enrollment or opportunities, employment or independent contracting opportunities or compensation, or health care services.
The definition of ADMT turns on the human role. Technology "substantially replaces" human decisionmaking when its output is used to make a decision without human involvement, and involvement requires a reviewer who knows how to interpret and use the output, reviews and analyzes it with any other relevant information, and has authority to make or change the decision. A rubber stamp does not meet that test.
The regulations also allow an opt-out exception when a business offers an appeal to a human reviewer with authority to overturn the decision. Separate exceptions for some hiring, admission, work-allocation, and compensation uses apply where the tool works for its purpose and does not unlawfully discriminate. So the text does not always require a person in the loop, and it describes the conditions under which a person is not required.
Quebec: notice and a reviewer for exclusively automated decisions
Section 12.1 of Quebec's private-sector privacy Act applies when an organization makes a decision about a person based exclusively on automated processing of personal information. According to the Commission d'accès à l'information, the organization must inform the person no later than when it notifies them of the decision, and must give them the opportunity to present observations to a member of its staff able to review the decision. The trigger is the word "exclusively," so a process in which a person genuinely weighs the output sits outside this duty. The Commission's own guidance page does not give a date for this obligation, and this article does not state one.
Canada's federal directive
The Treasury Board's Directive on Automated Decision-Making applies to departments listed under the Financial Administration Act, described in the scope guide as approximately 97 federal institutions. The guide says the directive applies to full and partial automation, including cases where there is a human in the loop. Having a person in the process therefore does not remove a federal decision from the directive's assessment requirements. The Armed Forces are excluded. The directive's tiered requirements in its appendices were not accessible for review when this article was prepared, so they are not described here.
The three rules side by side
| Rule | Trigger | What the person must be able to do | Status as of October 24, 2026 |
|---|---|---|---|
| California ADMT regulations | ADMT used for a significant decision about a consumer | Review the output, understand it, and have authority to change the decision | Approved; ADMT compliance due January 1, 2027 |
| Quebec Act, section 12.1 | Decision based exclusively on automated processing | Receive observations and review the decision | In force |
| Federal Directive on Automated Decision-Making | Full or partial automation of administrative decisions | Not described here | Scope guide page dated September 15, 2026 |
What these texts do not say
None of them protects a job category, sets a share of work for people, or forecasts what AI will do. They assign duties to organizations that deploy automated decisions about individuals, and each leaves out decisions outside its scope, such as tasks that do not affect a person's finances, housing, schooling, work, or care. This page is general information and not legal advice; applying any of these rules to a particular situation takes a lawyer's review of the current text.
FAQ
Does any law say AI must never make a decision alone?
Not in the three texts reviewed. They require notice, review opportunities, or assessments, and California lists circumstances in which an opt-out or appeal is not required.
What is a human in the loop under California's text?
A reviewer who knows how to interpret and use the technology's output, reviews and analyzes it along with other relevant information, and has authority to make or change the decision.
Does the Quebec rule apply to employers using AI screening?
It applies to decisions based exclusively on automated processing of personal information by private-sector organizations covered by the Act. Whether a particular screening process meets that description depends on how much a person weighs the result.
Are these rules the same across Canada and the United States?
No. California's rules are state regulations, Quebec's is a provincial statute, and the Directive is federal government policy for federal institutions. Other provinces and states have different or no equivalent rules.
Related reading
- Will AI replace most jobs
- Scientists and AI: tasks versus jobs
- Whether AI makes people lazy
- Predictive analytics and data-driven decisions
More in AI and Society.
Last verified: October 2026
Sources
- Text of Regulations (CCPA Updates, Cyber, Risk, ADMT, and Insurance Regulations), California Privacy Protection Agency, approved September 22, 2025 (source page)
- Principaux changements de la Loi 25, Commission d'accès à l'information du Québec (source page)
- Guide on the Scope of the Directive on Automated Decision-Making, Treasury Board of Canada Secretariat, page dated September 15, 2026
- CCPA Updates page, California Privacy Protection Agency, approval and effective dates