A retiree named Frank Delgado opened a letter from his city's parking authority. His residential permit renewal was denied. The letter gave no reason, just a case number and a line that said "your application did not meet eligibility criteria." Frank called. The clerk could not explain it either, because the eligibility check now ran through an automated address-verification system, and nobody at the counter knew how it worked. Frank spent three weeks, two trips downtown, and one call to a council member's office to learn the truth: the system had mismatched his address because his street had been renamed in 2019. A two-minute fix took three weeks because no one would, or could, tell a citizen why the machine said no.

Imagine the same thing happening to you. A government agency makes a decision about you using an AI system. You do not know it. You do not understand how the decision was made. You have no way to appeal it. This is a violation of due process. It is also a violation of the fundamental right to know how government is affecting you. This lesson is about the obligation to tell citizens when AI is involved in decisions affecting them, and about meeting that obligation in plain language, before a Frank ends up at a council meeting.

Why transparency is not a nicety

Transparency is foundational to democratic government. Citizens have a right to understand how government works, how decisions are made, and how those decisions might affect them. When government uses AI, that right does not disappear. If anything it becomes more important, because the decision now arrives through a mechanism that is harder to see into than a clerk with a rulebook. In the private sector a company can keep its algorithm secret as a trade advantage. Government cannot, for reasons baked into how public power works.

AI opacity. These systems are often hard to understand, so people need a clear explanation rather than access to a model. Power. Government decisions are binding, and citizens need to understand the reasoning behind something they cannot refuse. Trust. If citizens do not understand how decisions are made, they do not trust them. Legitimacy. Democratic governance depends on an informed citizenry, and you cannot be informed if government hides how it works.

Consent of the governed. People accept government decisions partly because they can see how they are made. A decision no one can explain feels arbitrary, and arbitrary government loses trust fast. Due process. You cannot appeal a decision you do not understand. Frank could not contest a reason he was never given, so transparency is the precondition for every other protection a citizen has. It is increasingly required. Federal guidance to agencies, including the White House Office of Management and Budget's direction on agency use of AI, pushes agencies to inventory their AI, notify the public, and explain AI that affects people's rights and benefits.

The transparency spectrum: from none to full

Transparency is not a switch. It is a dial, and different decisions call for different settings. Think of five levels, from least to most open.

1. Secret. Citizens do not even know AI is involved. Frank's parking authority sat here. This is the level that breeds lawsuits and front-page stories.
2. Notice. Citizens are told that an automated system is used. "This application is screened with an automated tool."
3. Explanation. Citizens are told the main factors. "Your renewal was denied because our records did not match your address of record."
4. Individual reasons plus appeal. Citizens get the specific reason for their outcome and a clear way to contest it, to a human.
5. Full openness. The agency publishes how the system works, its accuracy, and its testing, so journalists, researchers, and advocates can scrutinize it.

The right setting depends on stakes. A chatbot that helps you find a form needs Level 2. A system that denies benefits, housing, or a permit needs at least Level 4, and the agency as a whole should aim for Level 5 across its AI portfolio. The test is simple: could the person affected find out that AI was involved, and understand enough to challenge the result? If not, you are not transparent, no matter what your website says.

Public notice that people actually read

Many agencies think they have done notice because there is a paragraph buried in a privacy policy. That is notice in the legal-fiction sense, not the human sense. When government uses an AI system to make a decision affecting someone, they should provide notice, and the notice should carry six things, each in a sentence a person can read once and understand.

• That AI was involved. "An AI system was used in making this determination."
• What the AI did. "The system reviewed your application and scored your likelihood of success if accepted into the program."
• How it worked, if explainable. "The system considered factors including education level, employment history, and prior program participation."
• What data it used. "The system reviewed information from your application and our employment records."
• The outcome. "Based on this analysis, your application was approved or denied."
• Next steps. "You can appeal this decision and request human review."

Notice should be written in plain language rather than technical jargon, accessible to people with different education levels, available in multiple languages if relevant, and easy to understand. Timing matters as much as content: notice should be provided at the same time as the decision, or shortly before if possible, and always before the decision takes effect. A notice that arrives after benefits stop is a record of what happened, not a chance to prevent it.

Notice also belongs at the point of decision, not only on a policy page. Frank's denial letter was the perfect place to say "this decision used automated address verification; if it looks wrong, call this number for a human review." That single sentence would have saved three weeks. The four questions to ask of any notice you write are whether it says the system exists, what it does in one sentence a non-expert understands, what kinds of information it considers, and what the reader should do next.

Explanation: what it has to cover

Notice tells someone that AI was involved. Explanation tells them why the decision came out the way it did. Four things belong in one. Why the decision was made: not just "the AI said so," but the reasoning. What factors were important: "your employment history was weighted more heavily than education because it is more predictive of success." What you could have done differently: "if you had included more details about your work experience, the outcome might have been different." Uncertainty: "this determination is based on patterns in historical data and might not apply to your specific circumstances."

An explanation should be specific to the individual, tailored to their situation, honest about limitations, and non-technical. That last quality is where most attempts die. Compare a poor explanation, "the AI model determined your score was 0.42," which is incomprehensible to the person affected, with one that names the factors, gives the person's own numbers, states the threshold, and points at a concrete way to do better next time. The second costs more to write once and saves every subsequent phone call.

Here is the better version, with one deliberate omission. "We evaluated your application based on factors we know predict success in this program. You scored well on employment history (9/10) and program relevance (8/10), but lower on educational background (5/10). Your overall score was below the acceptance threshold (7.5/10). You could strengthen a future application by pursuing additional training."

The omission is deliberate and worth naming. The source for this example also stated a single overall score, and that figure does not follow from the three component scores it lists, so it has been left out and the components kept. Publishing components alongside a total that cannot be reconstructed from them is its own kind of transparency failure. The applicant who takes your explanation seriously enough to check the arithmetic is exactly the applicant you least want to lose, and an unexplainable total teaches them that the whole notice is decoration.

Plain-language explanations: the real skill

The hardest part of transparency is not deciding to be open. It is writing an explanation a person can actually use. Compare two versions of the same denial reason.

A good plain-language explanation does three things: it says what happened, it says why in human terms, and it tells the person their next move. Aim for the reading level of a local newspaper, define any term you cannot avoid, and always end with a path to a human. The failure to watch for is an explanation that is grammatically fine and informationally empty, because it will pass an internal review and then confuse every single person who receives it.

Appeal and human review rights

Notice and explanation are important, and they are not enough. People need recourse. An agency that tells someone precisely why it took something from them, and offers no way to challenge it, has produced a very clear account of an unaccountable decision. Five things belong in the recourse you offer, and each of them should appear on the notice itself rather than somewhere a determined person could eventually find.

• Right to appeal. "You can appeal this decision."
• Right to human review. "You can request that a human reviewer look at your case."
• How to exercise these rights. A clear, simple process, named on the notice.
• Timeline for response. "Your appeal will be reviewed within 10 business days."
• Authority of the reviewer. "The human reviewer can override the AI recommendation."

The human review process itself should be genuine rather than rubber-stamping, accessible rather than unreasonably difficult, timely rather than months of waiting, and competent, meaning the reviewer understands both the system and the decision. Those four qualities are what turn the sentence "you may request human review" from a promise into a protection. Offer the sentence without the four qualities behind it and you have told a citizen to spend their time on a process that was never going to change anything.

A worked case: a denied job training application

Your agency uses an AI system to determine eligibility for a job training program. A citizen applies. The system recommends denial. Here is what full transparency looks like when it is written out, and it is worth noticing how ordinary the language is. Nothing here requires a lawyer to draft or a data scientist to approve. It requires someone to decide that the applicant is entitled to know what happened to them.

The notice. "Thank you for applying to our job training program. Your application was reviewed by an AI system as part of our eligibility screening process. The system reviewed your application and employment history. Based on this review, the system recommended that your application not be approved at this time." That is four sentences, and it moves the applicant off Level 1 of the spectrum by itself.

The explanation. "The system scored your application on factors we know predict success in this program. Work experience: 7 out of 10, meaning your current experience is good but you have gaps. Educational background: 5 out of 10, because this program typically serves people with a further qualification. Program fit: 6 out of 10, because your interests align with the program in some areas but not all. Overall score: 6 out of 10, against an acceptance threshold of 6.5 out of 10." Those three components average exactly to the stated overall, which is what an applicant checking your arithmetic should find.

The appeal. "You have the right to appeal this decision. If you believe the AI assessment was incorrect or did not account for important factors in your situation, you can request human review. Contact the named officer by the date given to request an appeal. The human reviewer will look at your full application and can override the AI recommendation." Notice, explanation, and appeal together. That is transparency, and it fits on one page.

A usable artifact: the citizen transparency checklist

Before any AI system that touches the public goes live, walk it through this checklist. If you cannot check a box, you have a transparency gap to close first.

• Inventory. Is this system listed in our public AI inventory, with a plain description of what it does?
• Notice at the point of decision. Does the letter, screen, or form where the decision lands tell the person an automated system was involved?
• Individual reason. Does each affected person get the specific reason for their own outcome, in plain language?
• Reading level. Has the explanation been tested against a plain-language standard, with jargon removed or defined?
• Language access. Is the notice available in other languages where the population we serve needs them?
• Human path. Is there a clear, staffed way to ask questions and request human review, named on the notice itself?
• Accessibility. Does the notice meet accessibility standards, such as Section 508, so people using screen readers and people with limited English can understand it?
• Public-facing documentation. For higher-stakes systems, have we published how it works and how well it performs?

Run Frank's parking system through this list and it fails every box but the first. After his complaint, the authority added one sentence to its denial letters and a phone line to a trained clerk. Address-mismatch disputes that used to take weeks now resolve in a single call, and complaints to council members about the parking system dropped to near zero. That is a real improvement and it is not a clean bill of health. Adding the sentence closed the notice gap. The individual-reason, accessibility, and published-documentation boxes were still unchecked, which means the next Frank with a different problem may be back where the first one started.

Anti-patterns to watch for

• No notice at all. An agency uses an AI system to make decisions but does not tell people. Risk: citizens do not know they are being affected by AI, so they cannot understand or appeal decisions.
• Technical jargon instead of plain language. Notice is provided but uses terms only engineers understand. Risk: citizens do not understand it, so the notice is useless while looking like compliance.
• No appeal process. An agency provides notice and explanation but no way to appeal. Risk: if the decision is wrong, the person has no recourse and the clarity of the explanation makes that worse, not better.
• Fake explanation. An agency provides an explanation that does not explain anything, such as "our algorithm determined your ineligibility based on input parameters." Risk: citizens are confused rather than informed, and the agency believes it has discharged a duty it has not.
• Notice buried in a privacy policy. A paragraph on a policy page is treated as having told the public. Risk: nobody affected ever reads it, and the point of decision, where it would have helped, stays silent.
• Treating one sentence as the whole duty. A single line added to a letter closes the notice gap and is then cited as proof the system is transparent. Risk: individual reasons, language access, accessibility, and published documentation quietly stay unaddressed.
• Numbers that do not add up. Publishing component scores alongside a total the components cannot produce. Risk: the reader who checks concludes the whole explanation is decorative.

Practice prompts

• If your agency uses an AI system to make decisions affecting citizens, do you provide notice that AI is involved? Find the actual words and read them aloud to someone outside your field.
• If a citizen asked "why was my application denied by the AI system?" what would you tell them? Write the answer in full sentences, then remove every term you would have to define.
• What would good transparency look like for an AI system in your field, at each of the five levels of the spectrum?
• Take one decision letter your office sends and rewrite it so it carries all six notice elements without getting longer.

Reflection

Find an AI-based government decision you are aware of, in your agency or anywhere, and evaluate its notice and explanation. Do citizens get notice that AI is involved? Is the explanation in plain language? Can citizens appeal, and is the process clear and accessible? If any answer is "no" or "unclear," you have identified a transparency gap. Write down which of the eight checklist boxes it fails, who owns each of those boxes, and what the smallest version of a fix would be. Then ask yourself what it would take to get that one sentence onto the actual letter.

Glossary

• Notice. Informing someone that a decision has been made and explaining that AI was involved.
• Explanation. Describing how and why an AI system reached a particular decision.
• Plain language. Writing that is clear and understandable to people without technical background.
• Appeal. A formal process to contest a decision.
• Human review. Evaluation of an AI decision by a trained human who can override the AI's recommendation.
• Transparency spectrum. The range of openness from secret, through notice and explanation, to individual reasons with appeal and full published documentation.

Related lessons

• Algorithmic Fairness in Government explains why an unexplained automated decision is a due-process failure and not merely a communication problem.
• The Human in the Loop covers what has to be true for the human review you promise on a notice to be genuine rather than a rubber stamp.
• When Government AI Goes Wrong shows what happens when opacity and a missing appeal path meet a system operating at scale.
• Public Reporting and Algorithmic Transparency goes further into the published documentation that Level 5 of the spectrum asks for.

Closing

Transparency is not a nice-to-have. It is a fundamental right and a requirement of democratic governance. When your agency uses AI, tell people. Explain it clearly. Give them a way to contest it, staffed by someone who can actually change the outcome. That is how you maintain legitimacy and trust, and it is also, in the plainest operational terms, how you stop spending three weeks of staff time on a problem one sentence would have prevented. When citizens understand how decisions are made, they are more likely to accept those decisions even when they disagree with them.

Key Takeaways

• Transparency is the precondition for every other right. A citizen cannot appeal, correct, or trust a decision they cannot understand, so openness is not a courtesy, it is the foundation of due process.
• Notice is essential and has six parts. That AI was involved, what it did, how it worked, what data it used, the outcome, and next steps, delivered before the decision takes effect.
• Treat transparency as a dial, set by stakes. A helper chatbot needs only notice; a system that denies benefits, housing, or permits needs individual reasons plus a real path to appeal.
• Notice belongs at the point of decision. A paragraph in a privacy policy is legal fiction; one clear sentence on the actual letter or screen is real notice.
• Plain language is the hard part and the whole point. Say what happened, why in human terms, and what the person can do next, at a local-newspaper reading level, in the languages your public reads.
• Explanation must be honest about uncertainty. Say that the determination rests on patterns in historical data and might not fit this person's circumstances.
• Appeal and human review rights are essential. The review must be genuine, accessible, timely, and competent, or the offer of it is worse than no offer.
• Run the transparency checklist before launch. Inventory, point-of-decision notice, individual reasons, plain language, language access, a human path, accessibility, and published documentation separate real openness from a reassuring webpage.

Frequently Asked Questions

We describe our AI use on our privacy page. Is that notice?

Not in any sense that helps the person affected. That is notice as legal fiction. Real notice appears at the point of decision, on the letter, screen, or form where the outcome lands, and it says an automated system was involved, what it does, what information it considers, and who to contact. Frank's denial letter was the place, and one sentence there would have replaced three weeks of his time.

Our model is not explainable. Does that excuse us from explaining?

No. The duty is to explain the decision to the person, not to publish the model. Name the factors that mattered, give the person their own values on those factors, state the threshold, say honestly that the determination rests on patterns in historical data and might not fit their circumstances, and tell them how to reach a human. If you genuinely cannot say any of that, you have learned something important about whether the system belongs in a decision that affects rights.

How much should we publish about how the system works?

Match it to stakes. Level 2 notice suits a chatbot that helps someone find a form. A system that denies benefits, housing, or a permit needs at least individual reasons plus a real appeal, and the agency as a whole should be working toward publishing how its systems work, their accuracy, and their testing, so journalists, researchers, and advocates can scrutinise them.

Does telling people about the AI just invite gaming and appeals?

More appeals from people with a legitimate complaint is the system working. Frank's case took three weeks and reached a council member precisely because there was no cheap way to raise it. Once the authority named a phone line, the same disputes resolved in a single call. Opacity does not reduce the volume of complaints; it raises the cost of every one of them and moves them somewhere more expensive.

We added the disclosure sentence. Are we transparent now?

You closed one gap. Check the rest of the list honestly: does each person get the specific reason for their own outcome, has the wording been tested against a plain-language standard, is it available in the languages your public reads, does it meet accessibility standards such as Section 508, and is there a staffed human path named on the notice itself. A single sentence is a real improvement and it is not the whole duty.