Description: California's ADMT rules let a business avoid the consumer opt-out right only if the appeal is handled by someone who can actually revisit the decision. A template appeal process doesn't solve the review problem. It preserves the evidence that no one reviewed the person the first time or the second.
Date: 2026-05-31
Canonical: https://proofofreview.ai/record/the-human-appeal-exemption-is-not-a-way-around-the-review

# The Human Appeal Exemption Is Not a Way Around the Review

A candidate is rejected by an automated hiring screen and sends in an appeal. They attach a certificate the model appears to have missed, a short note explaining a gap in their work history, and a reference from a former manager. Two days later, the response comes back: the application has been reviewed and the decision stands. Inside the platform, the reviewer opened the appeal queue, checked that the original score was below threshold, confirmed there was no processing error, and sent the standard denial language. The appeal record is neat. It has a case number, a timestamp, and a named employee attached to it. What it doesn't show is anyone asking whether the added information should have changed the result.

[Section 7154](/glossary#section-7154) is the reason that workflow exists. California's ADMT rules give consumers a right to opt out of automated decision-making for significant decisions, but the rule makes room for a substitute: the business can provide an appeal to a human reviewer instead. That sounds like the cheaper path. An opt-out program has to be disclosed before the automated process runs, supported through at least two request methods, and backed by a non-automated alternative for the person who invokes it. An appeal inbox looks smaller and later. It catches only the people who object after the decision has already landed.

The text of the rule closes that escape. Under Section 7154(b)(1), the designated reviewer has to analyze the system's output and the other information relevant to changing the decision, including whatever the consumer submits. The reviewer has to know how to interpret the output, and has to be able to change the result based on that analysis. Put next to [Section 7001(e)(1)](/glossary#section-7001e1), the standard for [human involvement](/glossary#human-involvement), it is the same structure with a different filing cabinet. The person handling the appeal has to do the thing the original reviewer would have had to do for the system not to count as [ADMT](/glossary#admt) in the first place.

That is where the ordinary appeal process goes wrong. It treats the challenge as a quality-control ticket: did the model run, did the score map to the right policy, did the system send the right notice. Those are useful questions, but they are not the ones the regulation asks. The candidate's certificate, explanation, and reference are not bugs in the pipeline. They are new material bearing on the decision. A reviewer who never weighs them is not performing the analysis Section 7154 requires, even if the workflow records the file as reviewed and closed.

The danger is not simply that the exemption fails. It is that the failed appeal creates a cleaner record of the failure than the original decision did. A template response shows its own limits. It says the prior result was affirmed, but not what was considered, what mattered, or what could have moved the file. If the first pass was a recruiter clearing low-score candidates without opening their materials, the later correspondence does not cure the missing judgment. It adds a second named employee to the same absence. What was supposed to be a safeguard becomes a timestamped [rubber-stamp](/glossary#rubber-stamping).

The [UnitedHealth case](/record/unitedhealth-nh-predict-and-what-a-90-percent-reversal-rate-proves) shows the same structure in reverse. NaviHealth's nH Predict model cut off Medicare Advantage coverage when a patient's actual recovery ran past the model's population prediction. More than 90% of those terminations were reversed once a person later read the clinical record. The appeal did not prove that the system worked because a correction path existed. It showed that the first process and a real clinical review were producing opposite answers on the same patients. The later review was evidence about the earlier one.

That is why the appeal substitute is narrower than it looks. It does not let a business run the automated process first and add a human name later. It lets the business defer the consumer's protection only if the later review can actually reach the person, the facts, and the model output at the same time. The reviewer has to see the individual's file, not only the score. The consumer's submission has to be capable of changing the answer, not merely being attached to a closed result. There has to be some account of what was weighed and why. [Dwell time](/glossary#dwell-time) is not a safe harbor, but it is still arithmetic: ninety seconds on an appeal containing new material is enough time to confirm a prior answer and too little time to remake one. The same arithmetic that exposed Cigna's physicians clearing batches at [1.2 seconds per denial](/record/cigna-pxdx-and-the-1-2-second-review) applies when the queue is appeals instead of claims.

The appeal path is not weaker than the opt-out path. It is different in timing, not in substance. A business that chooses it has not avoided building a human decision process; it has promised to produce one when the consumer asks. The record that matters is not the existence of a form, a case number, or a response letter. It is whether the person reading the appeal had enough in front of them, and enough authority, to make the decision come out differently.