Elon Musk’s lawsuit against Colorado is being framed, in the usual way, as a clash between technology and regulation. But beneath the legal filings sits a more unsettling question—one that reaches far beyond any single state or any single company: what happens to democratic legitimacy when decisions are made by systems that cannot clearly justify themselves?
The issue isn’t simply whether AI can produce an answer. It’s whether society can interrogate the path that led to that answer. In a democracy, power is supposed to be contestable. Policies are debated, evidence is challenged, and officials are expected to explain themselves in ways that other people can evaluate. When AI systems influence outcomes—who gets access to services, how risks are assessed, which messages reach voters, how resources are allocated—the expectation of justification doesn’t disappear. It becomes harder, and therefore more politically charged.
Colorado’s case, as it has been discussed publicly, highlights a tension that has been growing for years: the gap between automated decision-making and human accountability. Even when an AI system is accurate, even when it performs well in tests, the democratic problem remains. If the reasoning is opaque—if the system cannot provide a meaningful explanation—then oversight becomes performative rather than substantive. People can be told what happened, but not why it happened in a way that allows them to challenge it.
That distinction matters because “trust” is not the same thing as “accountability.” Trust is a feeling; accountability is a structure. A democracy needs structures. It needs mechanisms that allow citizens, courts, journalists, and regulators to ask: What rule was applied? What evidence was used? What assumptions were built in? What trade-offs were made? And if the outcome is wrong—or unfair—how can it be corrected?
AI systems often struggle with these questions, not because they are malicious, but because many of the most powerful models are designed to optimize performance rather than to produce transparent reasoning. They can generate outputs that look like explanations while still failing to provide a faithful account of the internal process that produced the result. In other words, an AI can sometimes talk convincingly without actually revealing the causal chain that matters for accountability.
This is where the philosophical question becomes practical. If an AI can’t justify itself, can we still treat its outputs as legitimate inputs to governance? Or do we end up with a new kind of authority—one that is technically impressive but institutionally unanswerable?
To understand why this is more than a technical concern, it helps to consider what “justification” means in democratic life. Justification is not merely a narrative. It is a claim that can be tested. If a government agency denies benefits, it must cite reasons that can be reviewed. If a regulator imposes penalties, it must show the basis for its decision. If a court rules, it must explain its reasoning so that higher courts can evaluate whether the law was applied correctly.
In each case, justification serves two functions at once. First, it disciplines the decision-maker: explanations force clarity about what was considered and what was ignored. Second, it enables contestation: others can challenge the decision using the stated rationale.
When AI enters the picture, both functions can break down. Explanations may be incomplete, overly generalized, or detached from the actual determinants of the output. Contestation becomes difficult because the citizen is left arguing with a black box. Even if the system is audited, the audit may not translate into something a non-expert can use to challenge an individual outcome.
The result is a subtle shift in power. The institution still holds formal responsibility, but the operational control increasingly sits inside systems whose internal logic is not fully accessible. That creates a democratic mismatch: the public expects accountability, while the technology provides only partial visibility.
This mismatch is one reason the debate around AI and democracy has intensified. It’s also why lawsuits like Musk’s against Colorado resonate beyond the immediate parties. They symbolize a broader struggle over who gets to decide what counts as a valid explanation—and what obligations institutions have when they rely on systems that cannot fully disclose their reasoning.
There is another layer to the problem: discrimination. The headline question—can AI discriminate if it can’t justify itself?—isn’t just about whether bias exists. It’s about whether bias can be detected, attributed, and remedied.
Discrimination in algorithmic systems can arise from multiple sources: biased training data, proxy variables that correlate with protected characteristics, feedback loops where the system’s outputs shape future data, or design choices that encode value judgments. Sometimes discrimination is obvious. Often it is not. And when it is not obvious, the ability to justify decisions becomes crucial.
If an AI system can’t explain why it made a particular determination, then proving discrimination becomes harder. You can show that outcomes differ across groups, but you may struggle to show that the system’s internal logic relied on impermissible factors. You may also struggle to identify what change would fix the problem. Without justification, remediation becomes guesswork.
This is why “explainability” is frequently discussed as a moral and legal requirement, not just a technical feature. Explainability is meant to support fairness by making the decision process legible enough to audit. But explainability is also contested. Some argue that demanding full transparency is unrealistic for complex models. Others argue that partial transparency is better than nothing. Still others point out that explanations can be misleading if they are not faithful to the model’s true internal mechanics.
So the real question becomes: what kind of justification is sufficient for democratic accountability?
One approach is to distinguish between different levels of explanation. A system might not need to reveal every internal parameter, but it should be able to provide a meaningful account of the factors that influenced the outcome. For example, it could identify which features were most relevant, what thresholds were applied, and what data sources were used. It could also provide a counterfactual: what minimal change would have led to a different outcome. These forms of explanation can be more actionable than generic narratives.
Another approach is to shift from “explain the model” to “audit the behavior.” Instead of requiring the system to open its internal reasoning, institutions can require evidence that the system behaves fairly across relevant groups and contexts. This includes testing for disparate impact, monitoring drift over time, and validating performance under different conditions. Behavioral audits can be powerful, but they still face a democratic limitation: they may not help a specific person understand why they were denied or flagged.
A third approach is to redesign decision systems so that justification is built in from the start. Rule-based systems, structured decision trees, and hybrid models can provide clearer rationales. But they may sacrifice some performance or flexibility. The trade-off is not purely technical; it is political. Democracies must decide what they value more: maximum predictive accuracy or maximum contestability.
Colorado’s case, whatever its final legal outcome, underscores that these are not abstract debates. They are about the legitimacy of governance in an era where decisions can be outsourced to systems that do not naturally produce human-readable reasons.
There is also a deeper institutional question: who is responsible when AI is involved?
In theory, responsibility can be assigned along a chain: developers build the system, operators deploy it, institutions set policies, and regulators oversee compliance. But in practice, responsibility can become diffuse. If a model is trained by one party, fine-tuned by another, deployed by a third, and monitored by a fourth, accountability can dissolve into procedural blame-shifting.
Democracy hates that kind of diffusion. Citizens don’t experience responsibility as a chain of contracts; they experience it as a person or agency that owes them an answer. If the system cannot justify itself, the institution may still be held responsible—but it may not be able to provide the kind of explanation that makes accountability real.
This is why the question “Can AI discriminate if it can’t justify itself?” is also a question about institutional courage. If an AI system produces discriminatory outcomes and cannot explain why, will institutions admit uncertainty? Will they pause deployment? Will they invest in redesign? Or will they hide behind technical complexity and insist that the system is “working as intended”?
The danger is not only discrimination. The danger is normalization. Once opaque systems become routine, the public’s ability to challenge them erodes. Over time, the default assumption shifts from “decisions must be explainable” to “decisions are too complex to explain.” That shift is corrosive to democratic culture.
It also changes how people relate to authority. When citizens cannot understand the reasons behind decisions, they may interpret outcomes as arbitrary or hostile. Even if the system is statistically fair, the lack of justification can still feel like injustice. Democracy depends on more than fairness in the abstract; it depends on perceived legitimacy. People accept decisions more readily when they can see the reasons and believe those reasons are grounded in shared rules.
This is where the unique take on the current moment becomes important. The debate about AI and democracy is often framed as a contest between two extremes: either AI is a neutral tool that should be allowed to operate freely, or AI is a dangerous black box that must be banned. Reality is messier. The more useful framing is that AI is a new kind of administrative actor—one that can scale decisions, but also one that can obscure them.
Administrative actors have always had discretion. The difference now is that discretion is being embedded in systems that may not be fully interpretable. That creates a new category of governance risk: not just errors, but unreviewability.
Unreviewability is a democratic problem because it undermines the feedback loop that keeps institutions honest. Courts correct legal errors. Appeals correct factual errors. Oversight corrects policy errors. If AI decisions cannot be reviewed in a meaningful way, the system becomes self-sealing. Even when mistakes occur, the mechanism for correction weakens.
This is why the question of justification is central. Justification is the bridge between decision and review. Without it, review becomes superficial. With it, review becomes possible.
So what would a democratic approach look like?
First, it would treat justification as a requirement, not a luxury. That