Chapter 15: Communicating AI Decisions to Stakeholders

Part III: Transparency and Explainability

Opening: A Question Without an Answer

Maria Gonzalez was fifty-three years old when her oncologist sat down across from her and said words she had not anticipated. The cancer treatment she had been receiving was producing diminishing returns, and the hospital's AI-powered clinical decision support system had concluded — based on her tumor genetics, treatment history, comorbidities, and a vast pattern-matched database of comparable patients — that continuing aggressive treatment was unlikely to improve her prognosis. The system's recommendation: transition to palliative care.

Her doctor, a capable and compassionate physician, explained this as clearly as he could. But Maria had questions her doctor could not answer. What data had the system used? Were there patients like her who had defied the recommendation? What was the system's error rate? Had it ever been wrong about people with her particular cancer type? Could she ask for the recommendation to be reviewed by another physician — or did the hospital treat the AI's output as a second opinion that superseded his judgment? Could she see the factors the system had weighted most heavily?

The doctor did not know. The hospital's IT department had purchased the clinical decision support tool from a vendor; the vendor's model was proprietary; and neither the physician nor the hospital's administration had access to the underlying logic. The system had produced a recommendation. The recommendation had been communicated. But Maria had not been given information — not really. She had been given a verdict without a rationale, a conclusion without evidence, an outcome without recourse.

This is the central challenge of this chapter: AI systems are increasingly making or informing decisions that affect people profoundly, but the humans those decisions affect often cannot understand what happened, why it happened, or what they can do about it. The communication gap between what AI systems produce and what affected humans need is not merely a technical problem. It is an organizational problem, a legal problem, a design problem, and — most fundamentally — an ethical problem.

Learning Objectives

By the end of this chapter, students will be able to:

1. Identify the three levels of stakeholders affected by AI decisions and explain the distinct communication needs of each group.

2. Describe the key legal and regulatory requirements for AI communication in the United States and European Union, and articulate the gap between legal compliance and meaningful communication.

3. Analyze the specific challenges of communicating AI outputs to professionals who use AI tools, including the risks of automation bias and automation complacency.

4. Design meaningful explanation frameworks for individuals affected by AI decisions, including counterfactual explanations and plain-language communication standards.

5. Evaluate community-level communication practices for publicly deployed AI systems and apply the participatory governance model.

6. Explain how to communicate uncertainty, confidence levels, and system limitations in ways that are accurate and accessible to non-technical audiences.

7. Apply organizational design principles for building AI communication infrastructure, including feedback loops and escalation processes.

8. Distinguish between legitimate persuasion and manipulation in AI communication, and identify the ethical obligations around AI disclosure.

Section 15.1: The Communication Gap in AI Decision-Making

There is a profound asymmetry at the heart of modern AI-assisted decision-making. On one side sits the AI system: a statistical model trained on historical data, capable of processing vast quantities of information and producing outputs — scores, classifications, recommendations — with apparent precision. On the other side sits the human being whose life the system's output affects: a patient, a job applicant, a loan seeker, a person whose benefits have been cut, a resident of a neighborhood under AI-powered surveillance. Between these two stands often nothing more than a number, a flag, a yes or no — with no bridge to understanding.

This is the communication gap, and it is one of the defining ethical challenges of the AI era.

Three Levels of Affected Parties

The people affected by AI decisions do not form a single homogeneous group. They exist at three distinct levels, each with different relationships to the AI system and different communication needs.

The professional intermediary is the person who uses the AI tool in the course of their work — the physician reviewing a diagnostic recommendation, the loan officer evaluating an AI-generated credit score, the judge considering a recidivism risk assessment, the HR professional screening AI-ranked resumes. These individuals have technical competence in their domain; they may have limited AI literacy; they have professional and legal responsibilities for the decisions they make; and they often have a misplaced trust in the AI's outputs based on the system's apparent sophistication. The doctor does not question the AI's recommendation because the AI seems to know things the doctor does not. The loan officer approves the application the AI approved and declines the one the AI declined without scrutinizing why. The professional intermediary is both the AI system's primary user and one of the most important failure points when AI communication breaks down.

The directly affected individual is the person whose life is changed by the AI-informed decision. The patient who is denied a treatment. The applicant who is turned down for a mortgage. The job seeker who never gets a call back. The benefits recipient whose aid is cut. These individuals typically have no direct relationship with the AI system; they experience only its outputs through the decisions made by professionals or institutions. They often do not know an AI was involved in their case at all. And when they do know, they frequently cannot access the information they would need to understand, challenge, or seek redress for the decision.

The affected community is the population that experiences AI deployment at the collective level, not just the individual level. The neighborhood whose residents are targeted by predictive policing algorithms they did not choose and cannot review. The demographic group systematically disadvantaged by a hiring algorithm deployed across an industry. The low-income community whose residents are offered worse loan terms because an AI has learned to associate their zip code with higher default risk. Community-level AI effects are often invisible because they manifest as patterns across many individual decisions rather than any single, identifiable harm. The community-level communication problem requires different tools — public accountability, aggregate data disclosure, and participatory governance — rather than individual explanation.

The Mismatch Between Technical Outputs and Human Needs

AI systems produce outputs in the language of statistics and machine learning: confidence scores, classification probabilities, feature importance rankings, attention weights. These outputs are meaningful to data scientists but carry no inherent communicative value to the humans they affect. A 73% confidence score on a cancer recurrence prediction is not information a patient can use. It does not tell her what 73% means relative to the base rate for her cancer type. It does not tell her what the system got wrong for the 27% of cases where it was less confident. It does not tell her whether a 73% confidence score at her hospital means the same thing as it would at another hospital using a different model.

The mismatch is not merely a translation problem — a matter of converting technical jargon into plain language. It is a structural problem that arises from the fundamental difference between what AI systems do and what human communication requires. AI systems optimize for predictive accuracy. Human communication requires clarity, accuracy, actionability, and dignity.

Clarity means the communication must be understandable to its recipient with their existing background knowledge. This is different for a radiologist reviewing an AI diagnostic tool than for a patient receiving a treatment recommendation. The same information requires different framing, vocabulary, and level of detail.

Accuracy means the communication must faithfully represent what the system actually did and what its outputs actually mean — including its limitations, its error rates, and the uncertainty inherent in its predictions. Accuracy in AI communication is complicated by the fact that vendors and deployers routinely overstate their systems' capabilities, which means the humans communicating AI outputs to affected parties often do not themselves possess accurate information.

Actionability means the communication must give its recipient something they can do with it. Telling a loan applicant they were denied because of their "credit profile" gives them nothing they can act on. Telling them specifically which factors in their credit history drove the decision — and, ideally, what they could change to improve their outcome — gives them something to work with.

Dignity means the communication must treat its recipient as a person whose interests matter, not merely as a data point whose case has been processed. The difference between a cold, bureaucratic automated notice and a communication that acknowledges the stakes for the person receiving it is not trivial. It reflects a fundamental choice about whether the humans AI affects are subjects of moral concern or merely objects in a data pipeline.

These four requirements — clarity, accuracy, actionability, dignity — form the foundation for evaluating any AI communication. The chapters that follow examine how different types of stakeholders require different applications of these principles, and what organizational and regulatory systems are needed to deliver on them.

Section 15.2: Legal and Regulatory Communication Requirements

Law has not kept pace with AI-powered decision-making, but a patchwork of legal requirements has emerged that creates minimum floors for AI communication. Understanding these requirements — and, critically, understanding their limitations — is essential for any organization deploying AI in consequential domains.

ECOA Adverse Action Notices

The Equal Credit Opportunity Act (ECOA) and its implementing regulation, Regulation B, have required adverse action notices in credit decisions since 1976. When a lender denies a credit application, reduces a credit limit, or takes other adverse action, it must notify the applicant with the principal reasons for the decision. The Consumer Financial Protection Bureau (CFPB) provides a sample notice form that lists common adverse action codes — high balance on existing accounts, too many recent credit inquiries, insufficient credit history — and requires lenders to check the boxes that best describe the reason for their decision.

The adverse action notice regime worked adequately when credit decisions were made by underwriters applying rule-based criteria. It is strained to the breaking point by modern machine learning-based credit scoring. When a neural network processes thousands of variables to produce a credit score, identifying the "principal reasons" for its decision is not straightforward. The model doesn't select reasons from a predetermined list; it produces a numerical output that is the aggregate effect of thousands of weighted inputs. Post-hoc explanation methods — the techniques discussed in Chapter 14 — can approximate which features most influenced a particular prediction, but these approximations are themselves uncertain and may not faithfully represent the model's actual reasoning process.

The CFPB has acknowledged this challenge in its guidance on machine learning models in credit, but has not yet resolved the fundamental tension between ML-based decision-making and the disclosure-based accountability model of ECOA. Lenders using ML models must still produce adverse action notices, but the notices they produce often bear at best a probabilistic relationship to the actual factors that drove the model's output. A borrower told that their application was denied because of "insufficient payment history" when the model was actually most influenced by a pattern in their spending data that the system associates with default — but that the lender cannot describe in plain language — has received a notice that is technically compliant but substantively misleading.

GDPR Article 22

The European Union's General Data Protection Regulation (GDPR), in force since May 2018, contains what is arguably the world's most significant regulatory requirement for AI communication. Article 22 states that individuals have the right not to be subject to a decision based "solely on automated processing" that produces "legal or similarly significant effects." Where automated processing does produce such effects, individuals have the right to human review, the right to express their point of view, and the right to "meaningful information about the logic involved."

The scope of Article 22 is narrower than it might appear. It applies only to decisions made "solely" by automated means — a requirement that creates significant room for evasion through token human involvement. It covers "legal or similarly significant effects," which the GDPR does not define precisely, though regulatory guidance has indicated this includes credit decisions, employment decisions, insurance rating, and benefits determinations. The right to "meaningful information about the logic involved" falls short of a full right to explanation — a question extensively debated in the academic literature and examined in detail in Chapter 17.

GDPR enforcement of Article 22 has been uneven. Supervisory authorities in the EU have issued relatively few sanctions specifically for Article 22 violations, though the provision has been invoked in enforcement actions concerning automated decision-making in credit, employment, and social scoring contexts. The Dutch Data Protection Authority fined the Dutch Tax Authority 2.75 million euros in 2022 for using an automated fraud detection system (the toeslagenaffaire childcare benefit scandal) that used nationality as a discriminatory factor and failed to provide adequate explanation or human oversight — though the case involved multiple violations beyond Article 22.

EU AI Act

The EU AI Act, enacted in 2024 and phasing into effect through 2027, layers additional transparency and communication requirements on top of GDPR for AI systems in scope. For "high-risk" AI systems — defined to include AI used in healthcare, education, employment, critical infrastructure, law enforcement, and several other domains — providers must supply technical documentation and instructions for use that enable deployers to understand the system's capabilities and limitations. Deployers of high-risk AI systems must provide information to the people subject to those systems' outputs, at minimum indicating that a decision was made with AI involvement.

The AI Act also prohibits certain AI practices entirely on the basis that their opacity is itself harmful — AI systems designed to manipulate users subliminally or exploit psychological vulnerabilities are banned, reflecting the principle that AI communication must not be designed to circumvent conscious understanding.

FDA Requirements for AI Medical Devices

The Food and Drug Administration regulates AI-based software as a medical device (SaMD) under a framework that has evolved rapidly as clinical AI has proliferated. The FDA requires manufacturers of AI medical devices to provide labeling that communicates the device's intended use, its performance metrics, the population it was validated on, and known limitations. The FDA's 2021 Action Plan for AI/ML-Based Software as a Medical Device emphasized the importance of transparency to users and proposed a "predetermined change control plan" that would require manufacturers to communicate significant algorithmic updates.

For clinical decision support tools, the FDA's guidance distinguishes between tools that "inform" clinical judgment (lower regulatory scrutiny) and tools that "drive" clinical decisions (higher scrutiny, more robust disclosure requirements). In practice, this distinction is difficult to maintain: a tool labeled as advisory may function as directive if clinicians consistently follow its recommendations or lack the expertise to critically evaluate them.

NYC Local Law 144

New York City's Local Law 144, effective July 2023, requires employers in New York City who use automated employment decision tools (AEDTs) to conduct annual bias audits of those tools and to publish the results on their websites. Employers must also notify candidates when an AEDT has been used in an employment decision. Local Law 144 represents one of the first employment-specific AI disclosure requirements in the United States, though its implementation has faced enforcement challenges and questions about audit quality.

The Gap Between Compliance and Meaningful Communication

The consistent theme across these regulatory frameworks is that legal compliance and meaningful communication are not the same thing. A credit adverse action notice can check the required boxes without giving the applicant information they can actually use. A GDPR-compliant "meaningful information about logic" disclosure can be technically accurate without being comprehensible to the person who receives it. An FDA-labeled AI medical device can include performance metrics in its technical documentation without that information ever reaching the patient whose diagnosis it influenced.

The gap between compliance and communication is not accidental. It reflects the interests of deployers who benefit from AI opacity, the limitations of regulators who lack technical expertise, and the structural reality that disclosure requirements were designed before modern machine learning existed. Closing this gap requires not just legal compliance but a genuine organizational commitment to communication as a value — and that is the subject of the sections that follow.

Section 15.3: Explaining AI to Professionals

The professionals who use AI tools — physicians, judges, loan officers, HR professionals, social workers, parole officers, radiologists — are not passive conduits for AI outputs. They are supposed to exercise judgment, apply professional expertise, and take responsibility for the decisions they make. In practice, however, AI systems subtly and systematically undermine this independence in ways that create serious ethical risks.

What Professionals Need From AI Communication

To use an AI tool responsibly, a professional needs more than the system's output. They need to understand the output's confidence level — not just whether the system recommended Treatment A over Treatment B, but how confident it was, and what that confidence level means in practical terms. They need to know which features drove the recommendation — not necessarily all the technical details, but enough to assess whether the factors the model weighted most heavily are the factors they would expect to be important. They need to know the system's known limitations and failure modes — the circumstances under which the model has been shown to perform poorly, the populations on which it was not validated, the conditions that fall outside its training data. And they need to know the system's error rates in contexts comparable to their own — what the false positive and false negative rates are, what a false positive or false negative looks like for their patients or clients, and what the consequences of those errors are.

This information should be available to professionals not as a dense technical appendix they can theoretically access, but as an integrated part of their workflow. The radiologist reviewing an AI-flagged image should see not just "suspicious finding detected" but "system has 88% sensitivity and 72% specificity in this finding type; the positive predictive value in your patient population is approximately 34%." This requires vendors to provide appropriately contextualized information and organizations to train professionals to interpret and use it.

Automation Bias: The Risk of Over-Trust

Automation bias is the tendency of humans working with automated systems to over-weight the system's recommendations, particularly when they are under time pressure, cognitively loaded, or uncertain. It is one of the most well-documented human-AI interaction phenomena, observed in aviation, medicine, air traffic control, military operations, and many other domains.

The dynamics of automation bias in AI-assisted professional decision-making are predictable. A radiologist reviewing dozens of AI-flagged and unflagged images per shift will gradually learn — whether they realize it or not — to look more carefully at flagged images and less carefully at unflagged ones. Over time, the AI's judgment substitutes for the radiologist's independent review rather than augmenting it. The result is that unflagged images the AI missed receive less scrutiny than they would have before the AI was deployed. This is the automation bias paradox: a system meant to improve accuracy can degrade it by reducing the vigilance of the human reviewer.

For AI communication to counteract automation bias, it must do more than present the AI's recommendation. It must actively support independent professional judgment. This means presenting AI outputs in ways that do not foreclose alternative interpretations, that highlight uncertainty rather than projecting confidence, and that explicitly prompt the professional to consider factors the system may not have weighted appropriately. Some research suggests that presenting AI confidence levels as ranges rather than point estimates reduces automation bias; that emphasizing the system's limitations before presenting its recommendation helps professionals maintain appropriate skepticism; and that requiring professionals to commit to a judgment before seeing the AI's recommendation reduces the anchoring effect.

Automation Complacency: The Risk of Atrophied Judgment

Related to but distinct from automation bias is automation complacency — the gradual reduction of professional vigilance and skill that occurs when professionals consistently rely on automated systems to perform tasks they could perform themselves. Pilots who rely on autopilot lose manual flying skills. Radiologists who rely on AI diagnostic tools may lose pattern recognition ability in the areas where the AI is strongest. This skill atrophy matters not only because it reduces performance in edge cases but because it creates professional dependence on AI systems that may fail, be unavailable, or be deployed in contexts for which they were not designed.

The communication implication is that professionals need not just explanation of AI outputs but ongoing professional development that maintains their independent expertise. Training programs that integrate AI tools should include explicit components on maintaining critical judgment and professional skill, not just on how to use the AI tool effectively.

What Meaningful Human Review Looks Like

Both GDPR and the EU AI Act require "meaningful human review" in contexts where automated decisions have significant effects. This standard is easy to state and difficult to operationalize. Human review is not meaningful if the human reviewer has no information that would allow them to identify errors in the AI's output. It is not meaningful if time pressure or organizational incentives make actual review impossible. It is not meaningful if the human reviewer lacks the expertise to evaluate the AI's recommendation. And it is not meaningful if organizational culture treats overruling the AI as exceptional and requiring justification while accepting its recommendations as default.

Meaningful human review requires: professionals who have sufficient information about the AI system's outputs, limitations, and error modes; professionals who have sufficient expertise to evaluate those outputs independently; workflow designs that provide time and support for genuine review; organizational cultures that expect and support professional judgment; and accountability mechanisms that track and audit review decisions. The last point is underappreciated: if no one monitors whether human review is actually occurring — if there is no audit of the rate at which human reviewers agree with the AI, no tracking of cases where reviewers overrode the system, no investigation of outcomes in overridden versus accepted cases — then "meaningful human review" is a formality, not a function.

Case Examples

Consider two contrasting scenarios. In the first, a hospital deploys an AI sepsis detection tool that flags patients at elevated risk. Nurses receive the flag and are expected to conduct a clinical assessment; their assessment drives the care plan, and the AI flag is documented alongside the nursing assessment in the patient record. Nurses are trained in the tool's sensitivity and specificity; they know that the tool has a false positive rate of roughly 40%, meaning that for every 10 patients flagged, 4 are not actually developing sepsis. They use this information to calibrate their response: every flag receives assessment, but the flag does not determine the care plan. The human review is genuine.

In the second scenario, a financial institution deploys an AI credit scoring model. Loan officers review the AI's recommendations on a screen that shows the score, a color-coded risk rating (green/yellow/red), and a brief summary. The officer's override rate is monitored; too many overrides trigger a conversation with the supervisor. Loan officers have no access to information about the model's feature importance or error rates. They are trained to treat the AI's recommendation as the starting point and deviations as exceptions requiring justification. In this scenario, the nominal "human review" is not meaningful — it is performative compliance with a governance requirement while functional authority rests with the model.

The difference between these scenarios is not technical. It is organizational, cultural, and ethical. It reflects choices about how much genuine authority professionals are expected to exercise, how much information they are given to exercise it, and what accountability structures support or undermine independent judgment.

Section 15.4: Explaining AI to Affected Individuals

If communicating AI decisions to professionals is challenging, communicating them to the individuals those decisions affect is profoundly more so. The directly affected individual — the patient, the loan applicant, the job seeker, the benefits recipient — typically has no background in AI or statistics, no professional training in the relevant domain, and often significant emotional stakes in the outcome. They need information that is plain-language, specific, actionable, and that respects their dignity as persons whose interests matter.

What Individuals Need

Research on what individuals want to know about AI decisions affecting them has identified a consistent set of concerns. People want to know what information was used to make the decision — not a technical description of the model's inputs, but a comprehensible statement of what the system knew about them. They want to know the main reasons for the outcome — the factors that most influenced the decision, stated in terms they can understand and connect to their actual situation. They want to know what they can do — what would change the outcome, whether they can appeal, who they can talk to. And they want to know that someone is accountable — that there is a human being responsible for the process and accessible if the outcome seems wrong.

These are not exotic demands. They are the same things people have always wanted to know when institutions make consequential decisions about their lives. What is different about AI-informed decisions is that the mechanisms for producing this information have to be deliberately designed and built into the system — they do not emerge naturally from the way AI systems work.

The Dignity Dimension

There is an aspect of AI communication that efficiency-focused discussions often miss: the dignity of the person receiving the communication. A patient who is told, in impersonal language, that "the algorithm recommends against further treatment" is not being treated with dignity — even if the clinical recommendation itself is correct. A benefits recipient who receives a form letter citing "automated assessment" without any personal acknowledgment of their circumstances is not being treated with dignity. A job applicant who receives a generic rejection email generated by an HR AI without any individualized feedback is not being treated with dignity.

Dignity in AI communication does not require pretending that the AI did not influence the decision. It requires treating the person as a subject of concern — as someone whose situation matters to the organization making the decision — rather than as a data point that has been processed and disposed of. In practice, this means communicating in person or by phone for high-stakes decisions rather than by automated notification; offering the opportunity to discuss the decision with a human being; and acknowledging the person's circumstances, not just the outcome of the process.

Counterfactual Explanations for Individuals

Among the explanation approaches discussed in Chapter 14, counterfactual explanations are particularly well-suited to individual communication. A counterfactual explanation answers the question: "What would have been different about my situation that would have changed the outcome?" For a loan applicant, this might be: "If your debt-to-income ratio had been 10 percentage points lower, the decision would have been different." For a benefits recipient: "If your recorded hours of home care had been assessed at the higher rate used before the new assessment system, your benefit level would have remained the same."

Counterfactual explanations are actionable. They tell the person not just why the decision went the way it did, but what they could change — or, just as importantly, what about their circumstances the decision hinged on, which may reveal that the decision process was applying the wrong criteria or using incorrect information. A patient told that "the model weighted your age heavily" cannot do anything with that information. A patient told "the model's recommendation would have been different if your treatment had started earlier" has information she can use to question whether the decision model is appropriate for her situation.

The Literacy Challenge

Meaningful individual communication about AI decisions faces a significant literacy challenge. AI literacy varies enormously across the population. Some people have detailed understanding of how machine learning systems work; most do not. Even absent AI literacy, statistical literacy varies widely: many people do not have intuitive understanding of what a 70% probability means, or what a false positive rate of 15% implies for their individual case. And beyond statistical literacy, domain literacy — understanding what the AI's output means in the context of medicine, finance, or law — varies similarly.

This means that effective individual communication about AI decisions cannot be standardized at a single level. It must be adaptive: capable of being explained at multiple levels of complexity, with a default of plain-language communication designed for the least technically prepared recipient. The burden of literacy adaptation must fall on the organization, not on the individual. A benefits recipient should not be told they need to understand machine learning to challenge a decision that cut their benefits.

Language and Accessibility

Individual communication about AI decisions must reach people regardless of language, literacy level, or disability. For organizations serving linguistically diverse populations, this means providing AI explanations in the languages those populations speak — not just in English, with a note that translation services are available. For organizations serving people with disabilities, it means ensuring that AI communications are accessible via screen readers, in formats that can be converted to audio, and through channels accessible to people with mobility or cognitive limitations.

Accessibility is not merely a legal obligation under the Americans with Disabilities Act or comparable statutes. It is an ethical requirement: if the people most affected by AI decisions are systematically the least able to access information about those decisions, then the communication system compounds rather than mitigates the power asymmetry that AI creates.

What a Meaningful Adverse Action Notice Looks Like

The current legal minimum for a credit adverse action notice — a form letter with checked boxes indicating which of a handful of predetermined reasons applied to the decision — falls far short of what meaningful communication requires. Consider the contrast between two versions of the same communication.

Current legal minimum: "We regret that we are unable to approve your credit application. The principal reasons for this decision are: (1) Excessive obligations in relation to income. (2) Length of time accounts have been established."

Meaningful communication: "We carefully reviewed your credit application and were unable to approve it at this time. Based on our assessment, the two factors that most significantly affected our decision were your current debt load relative to your income, and the length of time your credit accounts have been open. Specifically, our review found that your monthly debt payments represent approximately 48% of your monthly income, and our criteria typically require this to be below 43%. Your credit accounts have been open for an average of two years, and our criteria typically look for an average of at least four years. If you were able to reduce your debt-to-income ratio — for example by paying down existing balances — or if you wait until your accounts have been open for a longer period, your application might be approved in the future. You have the right to request a free copy of the credit report we used to evaluate your application. If you believe any information in that report is inaccurate, you have the right to dispute it. If you have questions about this decision, you can speak with a member of our credit team at [phone number]."

The second version is not a legal requirement — it goes well beyond what ECOA demands. But it is what meaningful communication looks like. It gives the applicant specific, actionable information. It connects the decision to their actual circumstances. It tells them what they can do. And it treats them as a person whose situation deserves a genuine explanation, not just a regulatory notice.

The Right to Challenge: What a Genuine Appeal Process Involves

Any meaningful individual communication about an AI-informed decision must include genuine recourse — not just a notification that a decision was made, but a path to challenge it. A genuine appeal process has several features. It is accessible — easy to find, clearly described, available without unreasonable cost or procedural burden. It involves actual human review — not re-running the same AI model and receiving the same output, but a human being with authority to override the system's recommendation examining the case de novo. It has a realistic timeline — not a six-month wait while the applicant's situation continues to be affected by the original decision. And it is transparent about its outcome — if the appeal fails, the person receives a genuine explanation of why.

Many organizations' appeal processes fail most of these tests. An appeal process that requires written submissions evaluated by the same department that made the original decision, takes three months to resolve, and responds with a form letter citing the same reasons as the original notice, is not a genuine appeal process. It is a process designed to absorb objections without addressing them.

Section 15.5: Communicating AI to Communities

Individual-level communication, however well designed, cannot address the community-level effects of AI deployment. When AI systems are used to make decisions that affect not just individuals but communities — through predictive policing, public benefits administration, smart city applications, or other public-facing uses — community-level communication is needed.

Predictive Policing: What Communities Deserve to Know

Predictive policing systems — algorithms that predict where crimes are likely to occur or which individuals are at elevated risk of criminal activity — have been deployed in dozens of US cities. These systems affect communities in ways that go far beyond any individual's interaction with law enforcement: they shape where police resources are concentrated, which neighborhoods receive increased surveillance, and how residents experience public safety.

Community members affected by predictive policing have a legitimate claim to know that such systems are being used, what data they use, what their error rates are, and how their outputs influence policing decisions. This claim has not been consistently honored. Many predictive policing deployments have been made without public disclosure, relying on the argument that releasing information about the system would allow criminals to circumvent its predictions. This argument — whatever its merit — reflects a choice to prioritize operational secrecy over democratic accountability, and it is a choice that communities should be able to contest.

The Santa Cruz, California ban on predictive policing (enacted 2020, the first in the nation) and the subsequent wave of similar ordinances in Oakland, Portland, and other cities represent one response to this accountability gap: communities deciding, through democratic processes, that they do not want this category of AI deployment in their jurisdiction. Whether or not one agrees with that outcome, the process — communities making informed collective choices about AI deployment — represents the right structure.

Public Benefits AI: What Applicants Should Know

AI systems used in public benefits administration — determining eligibility for Medicaid, SNAP, housing assistance, and other programs — affect some of the most vulnerable members of society. These individuals have the least power to challenge the systems that affect them, the fewest resources to navigate complex appeal processes, and the greatest stakes in accurate, fair decisions.

Public benefits AI communication requires, at minimum: disclosure that an automated system was used in the decision; a plain-language explanation of the factors that drove the decision; information about the applicant's right to appeal and how to do so; and access to a human being who can review the decision and has authority to override the automated system. The Arkansas Medicaid case discussed in the case study accompanying this chapter illustrates what happens when these requirements are not met.

Smart City AI: Transparency With Urban Residents

Smart city applications — AI systems that optimize traffic, energy use, public space management, and city services — represent a different category of community AI deployment. These systems typically do not make individual decisions with identifiable impacts; they shape the environment in which residents live. The transparency requirements for this category of AI are less developed than for decision-making AI, but the principle that communities have a right to know about the AI systems that shape their environment is gaining traction.

Emerging best practices for smart city AI transparency include: public inventories of AI systems deployed in the city, with plain-language descriptions of their purposes and functions; mechanisms for public input on proposed AI deployments; and regular public reporting on system performance and equity impacts.

The Participatory Model: Communities as Participants, Not Recipients

The most advanced framework for community-level AI communication moves beyond transparency-as-disclosure toward participatory governance. In this model, community members are not simply informed about AI systems after they are deployed; they are involved in decisions about whether to deploy them, what purposes they should serve, and how they should be designed and monitored.

New York City's Automated Decision Systems (ADS) Task Force, established by Local Law 49 of 2018, represents an early attempt at this model. The task force brought together city officials, technologists, community advocates, and subject matter experts to examine how the city uses automated decision systems and to recommend transparency and accountability practices. Its 2019 report identified dozens of AI systems used by city agencies and recommended reforms to disclosure, explanation, and community engagement practices. The task force's work has influenced subsequent policy, though implementation has been slower than advocates hoped.

The participatory model reflects a fundamental principle: the legitimacy of AI deployment in public contexts depends not just on its technical performance but on the democratic processes through which communities choose to adopt it. Technology that is imposed on communities — even if it performs well — lacks the democratic foundation that public institutions require.

Section 15.6: Communicating Uncertainty and Limitations

AI systems do not produce certainties. They produce probabilities, predictions, and recommendations based on patterns in historical data — outputs that carry inherent uncertainty and are bounded by the limitations of the data and methods from which they were derived. Communicating this uncertainty honestly is both an ethical obligation and a practical necessity for appropriate use.

The Overconfidence Problem

The primary obstacle to honest uncertainty communication is the incentive structure of the AI industry. Vendors sell AI systems by emphasizing their capabilities; admitting limitations reduces sales. Deployers adopt AI systems to improve efficiency and performance; acknowledging uncertainty undercuts the rationale for adoption. The result is a pervasive tendency to present AI systems as more capable and their outputs as more reliable than they actually are.

This overconfidence has real consequences. A clinical decision support tool whose confidence scores are presented without context as to their base rate performance will lead clinicians to over-act on high-confidence predictions (increasing unnecessary interventions) and under-act on low-confidence predictions (potentially missing real conditions). A fraud detection system whose false positive rate is not communicated to customer service representatives will lead to unnecessary account holds and customer harm. A hiring AI whose screening criteria have not been validated against actual job performance will be used as if it were more predictive than it is.

Confidence Intervals and Error Rates: Explaining Them Without Technical Background

Communicating confidence levels and error rates to non-technical audiences requires translation, not just disclosure. The following principles apply:

Use absolute frequencies rather than percentages where possible. Research consistently shows that people understand "3 out of 100 people in your situation would experience this outcome" better than "there is a 3% chance of this outcome." The frequency framing makes the probability more concrete and easier to connect to one's own situation.

Explain error rates in terms of consequences, not just numbers. Telling a clinician that the sepsis detection tool has a 40% false positive rate is less useful than telling them that for every 10 flags, approximately 4 patients will not actually be developing sepsis and will not benefit from the escalation protocol. The consequences of errors are what drive appropriate clinical decision-making, not the abstract rate.

Use visual representations where available. Frequency diagrams, icon arrays, and natural frequency displays have been shown to improve statistical understanding significantly compared to verbal or numerical descriptions alone. When AI outputs are displayed in interfaces used by professionals or consumers, visual representation of confidence and error should be part of the standard interface design.

Acknowledge what is not known. Uncertainty communication should include not just quantified uncertainty (confidence intervals, error rates) but qualitative acknowledgment of the limits of the model's applicability. "This tool has not been validated on patients with your combination of conditions" is a critical communication, even if the model's developers cannot quantify exactly how much less reliable its predictions are for such patients.

The Role of Disclaimers

Disclaimers — "this tool is for informational purposes only," "this output should not be used as the sole basis for decisions," "results may vary" — are ubiquitous in AI communication. They serve a dual function: they provide some genuine notification of limitations, and they provide legal cover for deployers when things go wrong.

The problem with disclaimers as a communication strategy is that they are rarely effective communication. Research on warning labels and disclaimers consistently shows that people discount them, particularly when the disclaimers are presented in small print, at the bottom of a document, or in standardized language they have seen before. A disclaimer that says "this is for informational purposes only" on an interface that presents the AI's output prominently and authoritatively does not neutralize the authority effect — it just adds a phrase that users learn to ignore.

Genuine uncertainty communication requires integrating limitation information into the primary communication, not appending it as a disclaimer. The confidence level should be next to the recommendation, not in a footnote. The known failure modes should be visible in the interface, not buried in technical documentation. The system's validation population should be disclosed to users at the point of use, not just in an appendix to the procurement contract.

Case: AI Diagnostic Confidence and Clinical Communication

Consider a radiology AI tool that processes chest X-rays and flags potential nodules for radiologist review. The tool reports a confidence score for each flag — say, 87% for a particular finding. What does this mean, and how should it be communicated?

The 87% confidence score does not mean there is an 87% chance the patient has the condition. It means the model's internal classification probability for this image was 87%, which is a different thing entirely. The probability that the patient actually has the indicated condition depends on the base rate of the condition in the patient population being screened — which might be very low. If the prevalence of this nodule type in the general screening population is 1%, then even an 87% confidence score translates to a positive predictive value of roughly 6.6% — meaning that of 100 images flagged with this confidence level, only 6-7 represent actual nodules.

The clinician who does not understand this is likely to act on the flag as if it were highly reliable. The patient who is told "the AI is 87% confident" may be significantly alarmed, not understanding that this still means a roughly 93% chance the finding is not what the AI suggested. Honest communication requires translating confidence scores into their actual clinical meaning for the patient population in question — a translation that requires population-level data and cannot be done from the model output alone.

Section 15.7: Organizational Processes for AI Communication

Individual skills and regulatory requirements are insufficient to produce good AI communication. Good AI communication requires organizational infrastructure: processes, systems, training, and culture that make it possible and expected for the people who use and deploy AI to communicate honestly and meaningfully with the people those systems affect.

Communication as a Design Requirement, Not an Afterthought

The most common organizational failure mode in AI communication is treating it as an afterthought — something to be addressed after the AI system is built, tested, and deployed. In this model, the AI team builds the model; the compliance team reviews its legal exposure; and someone — often without technical knowledge of the system — writes the disclosure notices and trains the frontline staff. The result is communication that is legally defensible but not genuinely informative.

Effective AI communication must be designed in from the beginning. During the problem formulation stage, before the model is built, the organization should be asking: Who will receive communications about this system's outputs? What do they need to know? What can we build into the model that will enable meaningful communication? These questions shape not just the communication strategy but the model design: systems built with interpretability as a design goal produce better communications than systems for which interpretability is retrofitted through post-hoc explanation methods.

User Testing of AI Explanations

AI explanations should be tested with representative samples of their intended recipients before deployment. This is not a radical idea — it is standard practice in communication and instructional design — but it is remarkably rare in AI deployment. Testing should ask: Do recipients understand the explanation? Do they understand what they can do with it? Do they feel their situation has been acknowledged? Are there systematic differences in comprehension across demographic groups, literacy levels, or languages?

User testing of explanations should inform not just the communication but the model. If testing reveals that no plain-language explanation of a model's outputs is comprehensible to its intended recipients, this is information about the model's suitability for its intended use — perhaps the model needs to be replaced with one whose outputs can be meaningfully communicated.

The Feedback Loop: Using Communication Failures

When AI communications fail — when individuals successfully challenge decisions, when errors are discovered through appeals, when community groups identify patterns of discriminatory outcomes — these failures contain important information about the AI system as well as the communication strategy. Organizations that treat appeals and challenges as mere administrative noise, to be resolved with minimum disruption, lose the opportunity to learn from them.

Effective organizations build feedback loops that connect communication outcomes back to model evaluation. An elevated rate of successful appeals may indicate that the model is making errors; it may also indicate that the explanation is misleading recipients about the basis for decisions, causing unnecessary challenges. Both possibilities are worth investigating. A pattern in which challenges disproportionately succeed for members of a particular demographic group may indicate disparate error rates — a fairness problem in the model itself.

Training Programs for Professionals

Professionals who use AI tools need training in how to communicate AI-informed decisions, not just in how to use the tools. This training should cover: how to explain to a patient, client, or applicant that an AI system was involved in the decision; how to answer questions about why the system produced the outcome it did; how to explain recourse options; and how to maintain appropriate professional independence rather than deferring entirely to the AI's recommendation.

Training should be mandatory, not optional, and should be refreshed when the AI system is updated. It should include scenarios drawn from actual cases the professional is likely to encounter, not just abstract principles. And it should be evaluated for effectiveness — through simulation, role play, or outcome monitoring — rather than treated as a compliance check box.

Escalation Paths

AI communication systems need clear escalation paths for cases where the standard explanation is insufficient or where the individual is challenging the decision. Who does the front-line professional escalate to? Who has authority to override the AI system? What documentation is required for an override? What is the timeline for resolution? These questions must be answered before the system is deployed, not improvised when an escalation actually occurs.

Escalation systems should be staffed appropriately. If a hospital deploys an AI system that affects clinical decisions, there should be medical professionals with authority and expertise to review challenged decisions, not just administrative staff who can explain the process but cannot change the outcome.

Documentation and Accountability

Communications about AI-informed decisions must be documented: what was communicated, when, to whom, and by what means. This documentation serves multiple functions. It provides accountability: if a decision is challenged, the organization can demonstrate what the affected individual was told and when. It enables internal audit: by reviewing communications systematically, organizations can identify patterns of inadequacy. And it provides data for improving future communications: which communications led to challenges, which explanations were accepted, which generated follow-up questions.

Documentation requirements should be proportionate to the stakes of the decision. Communications about high-stakes AI decisions — healthcare, credit, benefits, employment — warrant robust documentation. Communications about lower-stakes applications can be documented less intensively.

Section 15.8: The Ethics of AI Communication

AI communication raises ethical issues that go beyond getting the information right. They concern the fundamental integrity of the relationship between institutions deploying AI and the people those systems affect.

The Line Between Persuasion and Manipulation

Persuasion — presenting information in ways designed to influence beliefs and decisions — is a normal part of communication. But when persuasion techniques exploit cognitive biases, information asymmetries, or power imbalances to produce choices the recipient would not make if fully informed, it becomes manipulation.

AI communication can cross this line in subtle ways. A loan officer presenting an AI-recommended credit decision can frame it in ways that make challenge seem futile ("the system has been very accurate in our experience") when the actual accuracy data would support a more equivocal assessment. An AI system can be designed to present its outputs in formats that are technically accurate but practically misleading — a confidence score that sounds more precise than it is, a risk classification that sounds more reliable than it is. The design of interfaces through which AI outputs are communicated is itself a communication choice with ethical implications.

Disclosure of AI Involvement

A distinct ethical question concerns not how AI decisions are explained but whether AI involvement is disclosed at all. When someone believes they are speaking with a human customer service representative and is actually communicating with an AI, they are being deceived. When someone believes their job application is being reviewed by a human recruiter and it has in fact been screened by an AI, they may be denied information they would have sought if they had known.

The FTC has issued guidance indicating that failure to disclose material facts, including AI involvement in interactions with consumers, can constitute deceptive practice under Section 5 of the FTC Act. California's BOT Disclosure Law (SB-1001, effective 2019) requires disclosure of the automated nature of a bot when the bot is interacting with California residents in ways intended to deceive. The EU AI Act requires disclosure when AI systems interact with users — with exceptions for clearly artistic or creative contexts.

These requirements reflect a principle that has significant traction in both law and ethics: people have the right to know when they are interacting with or being assessed by an AI system. Honesty about AI's role in consequential decisions is not just good practice — it is ethically obligatory.

Honesty as a Design Principle

Ultimately, the ethics of AI communication rests on a commitment to honesty as a design principle, not just a compliance requirement. AI systems should not be designed to create the appearance of capabilities they lack. They should not produce explanations designed to appear more informative than they are. They should not be deployed with communications strategies that obscure their role, overstate their accuracy, or discourage challenge.

Organizations that treat AI communication as a risk management exercise — asking "what do we need to disclose to limit our legal exposure?" — will consistently fall short of what genuinely affected people need and deserve. Organizations that treat AI communication as a genuine service — asking "what do the people this system affects need to know, and how can we best provide it?" — will do better, both ethically and practically. Honest communication builds trust. Trust enables the human oversight that makes AI safer. Safer AI makes better decisions. The ethics of AI communication are not in tension with organizational interest; they are aligned with it.

Discussion Questions

1. A hospital has deployed an AI system that recommends whether patients should be discharged or continue receiving inpatient care. The system's recommendations are shown to nurses and attending physicians but not to patients. A patient advocacy group argues that patients have the right to know if an AI influenced the recommendation to discharge them. How should the hospital respond? What information should patients receive, and in what form?

2. A lending institution uses a machine learning model to make credit decisions. The model uses 400 variables and produces accurate predictions, but no individual variable explains more than a small fraction of any individual decision. The adverse action notices the institution sends describe the top three contributing factors to each decision — factors identified by a post-hoc SHAP analysis that the bank's own data scientists acknowledge may not perfectly represent the model's actual reasoning. Is this adequate disclosure? What would better communication require?

3. A city has deployed a predictive policing system that the police department argues reduces crime. Community organizations in affected neighborhoods object that they were not consulted before deployment, that the system perpetuates historical patterns of over-policing in their neighborhoods, and that they cannot get meaningful information about how the system works. The police department argues that disclosing operational details would compromise the system's effectiveness. How should this conflict be resolved?

4. A social media platform uses AI to moderate content and remove posts that violate its community standards. Users whose posts are removed receive an automated notice saying their post violated community standards. They can appeal, but appeals are reviewed by the same AI system. Is this adequate communication and recourse? What would a more ethical system look like?

5. An insurance company uses an AI model to set health insurance premiums. The model is more accurate than traditional actuarial methods at predicting health costs, but it uses factors — including shopping behavior, social media activity, and geolocation data — that customers do not know are being used. The company discloses only that it uses "a proprietary pricing model." Is this adequate disclosure? Does the accuracy of the model affect the ethical analysis?

6. A benefits agency's AI system incorrectly terminates a disabled person's benefits. The person receives a form letter citing "assessment update" as the reason. When they call to appeal, they are told the decision was made by an automated system and that a human review will take 90 days. What are the ethical failures in this scenario? What would a system that met minimum ethical standards look like?

7. You are the Chief AI Officer of a large employer that uses AI for hiring. A new hire discovers, through a journalist's story, that your company's hiring AI systematically downscored applicants who attended historically Black colleges and universities. You did not know about this pattern. How do you communicate with: the people wrongly screened out; current employees; the public; regulators?

This chapter continues with case studies on the Arkansas Medicaid welfare algorithm and AI communication in electronic health records. The following chapter examines transparency in AI marketing and advertising — a context where communication obligations intersect with commercial incentives in particularly fraught ways.