Skip to content
AI.info

Responsible AI

Contestability, Appeal, Recourse, and Redress

Design notice, challenge, evidence access, human review, correction, recourse, compensation, and systemic learning.

By the end you can

Visual

The layers a contestability and remedy review must connect

Notice, evidence, review, interim protection, repair. A route that stops at any of the first four leaves the person informed and unhelped.

Two records mark the ends of that route in this lesson. An applicant was denied housing on a tenant-screening score. The property manager's answer was that it could not override the outcome. That is Louis v. SafeRent Solutions, filed 25 May 2022. In Michigan, the Talent Investment Agency went looking behind its own automated decisions. It reviewed 62,784 fraud-penalty cases nobody had appealed, and reversed 85 percent of the findings a computer program had made.

The first is a route that exists on paper and cannot be walked. The second is what piles up behind such a route, uncounted, until somebody counts it.

FigureProcess · 5 steps
  1. 1

    Notice

    Reveal the decision, AI role, consequence, time limit, and available rights.

  2. 2

    Evidence access

    Provide relevant data, reasons, policy, and source information.

  3. 3

    Challenge and review

    Offer accessible submission and independent human reconsideration.

  4. 4

    Recourse and interim protection

    Give feasible actions, alternatives, and pause where delay is harmful.

  5. 5

    Redress and learning

    Correct records and decisions, restore opportunity, compensate, and remediate systemic cause.

Comparison

Explanation, Recourse, or Redress?

Explanation lets a person understand. Recourse tells them what to do. Redress repairs what already happened. Only the third changes an outcome already suffered.

Redress needs authority behind it, and authority is something a court has either supplied or has not. Michigan's unemployment system disqualified claimants, garnished wages and seized tax refunds without notice or an opportunity to be heard. The system was the Michigan Integrated Data Automated System. A class action followed. The Michigan Supreme Court decided it on 26 July 2022, in Bauserman v Unemployment Insurance Agency, and held for the first time that money damages are available for a Michigan constitutional tort: “Although we have never specifically held that monetary damages are available to remedy constitutional torts, we now hold that they are.”

Read the three cards against that sentence. An explanation of MiDAS would have told claimants why their wages were taken. Recourse would have told them what to file next. What they needed was the third column. It took until 2022 for a court to say the third column existed at all.

FigureComparison · 3 columns

Explanation

Helps a person understand the decision.

  • Necessary for informed challenge
  • May be incomplete or unfaithful
  • Does not itself change the outcome
  • Should connect to rights and evidence

Recourse

Identifies feasible ways to obtain a better outcome.

  • Must respect causal and institutional constraints
  • Can become unfair burden shifting
  • May include alternative service routes
  • Should not require impossible self-transformation

Redress

Repairs a wrong or harmful outcome.

  • Can include correction, restoration, compensation
  • Requires authority and case finding
  • Should address similarly affected people
  • Feeds systemic remediation

Example

Louis v. SafeRent: a screening score the court found undisclosed

The tenant-screening scenario is not a scenario. It is a docket.

Louis v. SafeRent Solutions was filed on 25 May 2022 and reached the motions to dismiss on 26 July 2023. Judge Angel Kelley denied Metropolitan Management Group's motion in full. SafeRent's motion she granted in part and denied in part, dismissing only the Chapter 93A claims, Counts VII and VIII. The court held that “SafeRent is subject to the FHA”, because its score “determined who was qualified to occupy a housing unit”. Getting there, it credited the amended complaint's allegation that SafeRent “effectively controls the decision to approve or reject a rental application”.

The evidence asymmetry is not an inference drawn here. It is a finding in the opinion: “SafeRent does not disclose the weight assigned to any of the factors considered in the SafeRent Score, nor does it provide the specific sources of its data.”

  • Notice gap: The denial rests on a score whose weights the vendor “does not disclose” and whose data sources it does not provide. The applicant learns the outcome and nothing she could use to frame a challenge to it.
  • Evidence asymmetry: The court did not have to be persuaded that the vendor and the landlord hold the record. It wrote the asymmetry into its own recitation of the facts, and credited the allegation that SafeRent “effectively controls the decision to approve or reject a rental application”.
  • Burden shift: To show the score is wrong, an applicant argues against a model whose factor weights are undisclosed. The company holding those weights is the one the court found “subject to the FHA”, because the score “determined who was qualified to occupy a housing unit”.
  • Time sensitivity: Filed 25 May 2022. Motions to dismiss decided 26 July 2023. Settlement finally approved 20 November 2024. Housing opportunities disappear on a much shorter clock than that, and nothing in the route pauses them.
  • System learning: The systemic fix arrived through litigation, not through the appeal. The settlement approved on 20 November 2024 is worth up to $2,275,000. Under it SafeRent agreed to change its screening practices for at least five years, and to stop issuing accept/decline recommendations for housing-voucher applicants unless the scoring model is validated for fairness.

What contestability and remedy changes in practice

Contestability is the practical ability to question an AI-influenced decision. An appeal asks for the decision to be taken again. Recourse identifies feasible steps or alternative routes. Redress repairs harm, through correction, restoration, compensation, or some other remedy. A meaningful process gives timely notice, understandable grounds, access to the relevant evidence, and a low-burden way to submit a challenge. It then gives independent review, a pause where delay causes harm, a reasoned response, correction of the source data, and propagation of the lesson to similar cases.

Understandable grounds was once written down in plain terms, and the terms were a regulator's. The adverse action notice requirements of ECOA and Regulation B “apply equally to all credit decisions, regardless of the technology used to make them”. That is Consumer Financial Protection Circular 2022-03, released by the Consumer Financial Protection Bureau on 26 May 2022 and published in the Federal Register on 14 June 2022. It then closed the opacity defence: “A creditor cannot justify noncompliance with ECOA and Regulation B’s requirements based on the mere fact that the technology it employs to evaluate applications is too complicated or opaque to understand. A creditor’s lack of understanding of its own methods is therefore not a cognizable defense against liability for violating ECOA and Regulation B’s requirements.”

Now read the second half of the record. On 12 May 2025 the CFPB withdrew the circular. It went out as item 14 of a 39-entry list of other guidance. That list was one part of 67 documents withdrawn at once: 8 policy statements, 7 interpretive rules, 13 advisory opinions, 39 other guidance documents. A stated duty and a staffed, standing duty are different objects. The difference is visible in the Federal Register, three years apart.

Answering a demand for redress with an explanation leaves the harm exactly where it was.

Case

Articles 85 and 86, in a section headed “Remedies”

Contestability gets a legal hook of its own in the AI Act, and Article 86 does not stand alone on the page. The affected person's remedies sit in Chapter IX, Section 4, under a heading that says exactly that: “Remedies”. The regulation is (EU) 2024/1689, dated 13 June 2024 and published in the Official Journal on 12 July 2024.

Article 86 lets an affected person obtain clear and meaningful explanations of the AI system's role in a decision, for the high-risk uses listed in Annex III. Beside it, Article 85 supplies a complaint route: “Without prejudice to other administrative or judicial remedies, any natural or legal person having grounds to consider that there has been an infringement of the provisions of this Regulation may submit complaints to the relevant market surveillance authority.”

The dates matter as much as the wording. Under Article 113 both provisions applied only from 2 August 2026 — three weeks before this lesson's date. Article 6(1) and the obligations that go with it do not apply until 2 August 2027. A review conducted today is looking at rights that are weeks old, against classification duties still a year out.

Case

CK's enforceable order, and five years of nothing

GDPR Article 22(3) is older and narrower. It requires human intervention, a chance to be heard, and a right to contest. Neither provision creates a working remedy on its own.

A right with no case handler behind it is a page on a website. CK is that sentence with a court file behind it. She was refused a mobile contract worth EUR 10 a month on an automated creditworthiness score. In October 2019 she won a final Austrian ruling ordering disclosure of the logic. More than five years later she still had nothing: “The decision of 23 October 2019 has become final and is enforceable under Austrian law. However, CK’s application for enforcement of that decision, lodged by CK with the City Council of Vienna, which is the enforcing authority, was rejected on the ground that D & B had met, to the requisite standard, its obligation to provide information, even though that company had not provided any additional information after that decision was adopted.”

The Court of Justice decided her case on 27 February 2025, CK v Magistrat der Stadt Wien, against Dun & Bradstreet Austria. It set the standard the enforcing authority had failed to apply. “Meaningful information about the logic involved”, under Article 15(1)(h) of the GDPR, means the controller must explain the procedure and principles actually applied. Trade secrecy does not permit a blanket refusal. The allegedly protected material goes instead to the supervisory authority or the court to balance.

Note what it took to get there. A final order, a rejected enforcement application, and a reference to the Court of Justice of the European Union. For a contract worth EUR 10 a month.

Steps

From the notice to the systemic cause, with the deadlines attached

The five steps run from the notice a person receives to the systemic cause their appeal exposes. In some jurisdictions they are not abstractions at all. For tenant-screening companies operating as consumer reporting agencies, 15 U.S.C. § 1681i writes them as deadlines.

The reinvestigation is free, and it is clocked: “Subject to subsection (f) and except as provided in subsection (g), if the completeness or accuracy of any item of information contained in a consumer's file at a consumer reporting agency is disputed by the consumer and the consumer notifies the agency directly, or indirectly through a reseller, of such dispute, the agency shall, free of charge, conduct a reasonable reinvestigation to determine whether the disputed information is inaccurate and record the current status of the disputed information, or delete the item from the file in accordance with paragraph (5), before the end of the 30-day period beginning on the date on which the agency receives the notice of the dispute from the consumer or reseller.”

Thirty days, then. Not more than 15 additional days, if the consumer supplies relevant information inside that window. Written notice of the results not later than 5 business days after the reinvestigation is completed.

Step 5 has a statutory form too. Any consumer reporting agency compiling files on a nationwide basis must run an automated system. Through it, the furnishers who feed the agency report a reinvestigation that found information incomplete or inaccurate to the other nationwide agencies. That is propagation as a costed obligation rather than a good intention.

Step 3 is where the design usually breaks, and here a defendant says so in writing. The opinion records the manager's own answer to the rejected applicant: “Metropolitan explained to Louis that “the third-party service” it utilized “to screen all prospective tenants has denied [her] tenancy,” and it “cannot override the outcome of the Tenant Screening.””

A reviewer without authority to reverse is not a review step. It is a place the file stops.

FigureProcess · 5 steps
  1. 1. Give timely notice

    Explain the outcome, AI role, relevant grounds, deadline, and contact.

  2. 2. Provide evidence access

    Offer correctable data and sufficient information to frame a challenge.

  3. 3. Make review independent

    Use qualified reviewers with authority, time, and no penalty for reversal.

  4. 4. Protect against delay

    Pause or provide alternatives when lost opportunity is irreversible.

  5. 5. Repair and learn

    Correct records, restore benefits, compensate, find similar cases, and change the system.

Example

Walk the appeal as an outsider, then count who never appealed

Complete the appeal yourself, without insider knowledge, and time it. Then ask the harder question. How many people never reached the form at all?

Michigan's own agency answered that one with numbers. On 11 August 2017 the Talent Investment Agency announced the completion of a review covering October 2013 to August 2015: “TIA-UI has reviewed 62,784 cases for which people were assessed a fraud penalty and did not seek an appeal. Of those cases, 40,195 were originally resolved by way of computer program based on available information. As part of the review, 85 percent of these original fraud findings were reversed.” Among the 22,589 cases a human investigator had touched, 44 percent were reversed. Only 4,955 cases had ever gone through the appeals process. The agency was refunding more than $20.8 million.

  • Appeal walk-through: Complete the process as an affected person, without insider knowledge. At each step, check whether the person you reach can actually reverse the outcome. Metropolitan's written answer to Louis was that it “cannot override the outcome of the Tenant Screening.”
  • Burden inventory: Measure time, documents, language, cost, travel, fear, and access needs. Compare the result against a statutory yardstick that already exists: 30 days for a free reinvestigation, not more than 15 additional days, 5 business days for written notice of the results.
  • Case propagation: Define how one successful appeal triggers a search for similarly affected people, and size what that search will find. In Michigan the ratio was roughly thirteen unappealed penalties for every one appealed — 62,784 against 4,955 — and 85 percent of the machine-decided findings were wrong. A remedy route measured only by the appeals it receives is measuring the smaller number.
  • Recourse audit: Remove recommendations that are impossible, discriminatory, unsafe, or controlled by the institution. That includes any step asking the person to disprove a factor whose weight the vendor “does not disclose” and whose data sources it does not provide.

Key idea

Asked to disprove a score whose weights are undisclosed

A process is not contestable when the person must prove an error using information only the organization possesses. Six things decide whether it is usable: burden, timing, language, disability access, fear of retaliation, and whether independent help can be had. On the facts recorded in Louis v. SafeRent Solutions, the applicant cleared none of the six. The screening route still counted internally as a way to challenge the outcome.

Not every model decision can be explained completely. Not every adverse outcome is erroneous. The process should still provide a fair review of evidence, policy, and individual circumstances. What the record shows instead is three things stacked on one applicant. A score that “determined who was qualified to occupy a housing unit”. A company that “does not disclose the weight assigned to any of the factors considered in the SafeRent Score, nor does it provide the specific sources of its data”. A landlord answering that it “cannot override the outcome of the Tenant Screening.” Not one of the three was hers to move.

A route to challenge can exist on the control list and be unavailable on every dimension that decides usability.

What an affected person can actually reach

A remedy route is judged by what an affected person can actually reach, not by the policy that describes it. CK reached a final, enforceable order on 23 October 2019 and had no additional information more than five years later. Louis reached a landlord that could not override the vendor. The 62,784 Michigan claimants who never appealed reached nothing at all. Then an agency review reversed 85 percent of the computer-resolved findings among them and refunded more than $20.8 million.

Name the burden, timeliness, or case-propagation finding that would force the reviewer to redesign, restrict, remedy, or retire the system. If the only evidence you have is that a form exists, you have measured existence and nothing else.

Key takeaways