Sarra Tajouri, Yves Meinard, Alexis Tsoukiàs, Thierry Kirat
11 min
Abstract
Explainability of algorithmic decision-making systems is both a regulatory objective and an area of intense research. The article argues that a crucial condition for the acceptability of algorithmic decision-making systems is that decisions must be justified in the eyes of their recipients. We make a clear distinction between explanation and justification. Explanations describe how a decision was made, while justifications give reasons that aim to make the decision acceptable. We propose a conceptual framework of explanations and justifications, based on Habermas's theory of communicative action and Perelman's New Rhetoric theory of law. This framework helps to analyze how different forms of explanation can support or fail to support justification. We illustrate our approach with a case study on university admissions in France.
Alex: Okay, so not all explanations are the same. Like, some zoom in on one case, others look big-picture?
Sam: Exactly—sorted by scope. A *local explanation* focuses on one specific decision, pinpointing what mattered most there, say recent payment misses outweighed steady work for that loan denial. It's like explaining why your team lost a single game point-by-point. A *global* one covers the whole system's patterns, such as payment history always dominating across thousands of cases. This helps see the bigger logic without drowning in details.
Alex: And timing? Do they make explanations before or after the decision?
Sam: Good question. *By-design explanations* are baked into simple models from the start, like a basic math formula where each part's role is obvious—no extras needed. *Post-hoc* ones come later for complex "black-box" systems, approximating with simpler stand-ins around that case. The paper notes by-design ones match reality better, but post-hoc are common for powerful models.
Alex: And who it's for changes things too, right? Not everyone needs the full tech dive.
Sam: Right—for non-experts like loan applicants, *popularized explanations* use everyday words: "Boost income by a bit to qualify next time." No numbers or formulas that confuse. Experts get technical breakdowns, like feature weights showing payment history's pull. The paper stresses matching the audience so info actually lands and builds understanding.
Alex: So these types make explanations clearer... but tie back to justification somehow?
Sam: They do, but first the paper looks at explanations through the lens of reasoning styles people naturally use. One common way is pointing out contrasts: instead of just saying what happened, it explains why this outcome beat a different one—like why your loan got denied instead of approved, by comparing your income and defaults to typical yes cases. Another focuses on changes: what small tweak, like bumping income to 1200 euros, would flip the result. Researchers label these *contrastive* and *counterfactual* explanations, but the point is making sense of decisions by highlighting key differences or fixes.
Alex: Okay, that sounds useful for someone contesting a no—like in university admissions. But there are more ways to reason through it?
Sam: Yes. Some work backward from the outcome to the likeliest cause, weighing options like low income versus recent defaults to pick the strongest factor—called *abductive* reasoning. Others trace true cause-and-effect chains, separating real impacts from coincidences, such as how income directly drops the risk score below threshold. These build deeper understanding without overwhelming details.
Alex: Right, so explanations can mimic human thinking. Does the paper say when these fall short for trust?
Sam: Exactly—it shifts to viewing explanations as conversations, not fixed texts. In a *dialogical* setup, the recipient asks questions, and the system answers cooperatively, like chatting about why a loan failed based on your specifics. This evolves into *dialectical* when challenges arise: you debate legitimacy, citing evidence or context, until both sides see the reasoning's strength. The paper draws on thinkers like Habermas here, proposing a typology of four approaches—technical, norm-oriented, expressive, and communicative—to classify them by rationality type.
Alex: A typology... so like sorting tools by their job? How does that split explanation from justification?
Sam: Precisely. Explanations clarify *how* via technical or cognitive fits, like those reasoning types. Justifications push for *why accept* through argumentative dialogue in the communicative model, seeking mutual agreement on context-specific reasons—echoing Habermas's idea of rational discourse for legitimacy. This bridge ensures decisions gain real buy-in, as in Parcoursup appeals. The paper suggests this framework fills the gap in current systems.
Alex: So it's about turning one-way info into two-way fairness checks.
Sam: Let's unpack those four models the paper proposes. The first relies purely on facts and science—like describing a machine's gears and math to show exactly how it crunched the numbers for a decision. No opinions or rules, just objective proof it works reliably. They call this the *technical model*. It suits experts but often leaves everyday people unconvinced about fairness.
Alex: Okay, so that's great for *how* but weak for *why accept*? What's next?
Sam: The second checks against rules or ethics, like proving the system follows laws on equal treatment or safety standards baked into its design. For justification, it argues the outcome matches those shared norms, even if pulled in after the fact for broader appeal. This is the *norm-oriented model*. It builds trust through compliance, though personal hardship might still override it for the recipient.
Alex: Right—like legality doesn't always feel just. And the third?
Sam: The third reveals the true motives behind the system's choices, such as the designer's values on prioritizing steady jobs over risky ones. As justification, it connects to what the recipient already believes, hoping for recognition. Known as the *expressive model*. Success depends on value overlap, which isn't guaranteed.
Alex: So explanations track the designer's intent, justifications mirror the person's own views. But the fourth sounds key from what you said earlier.
Sam: Exactly—this one turns it into a back-and-forth talk, like debating in court: exchanging arguments until both grasp and agree on the best reasons, tailored to context and needs. Explanation builds understanding through dialogue; justification seeks buy-in on legitimacy via the strongest case. It's the *communicative model*, drawing from Habermas for rational consensus. The paper argues it's strongest when interests clash, as in Parcoursup rejections, fostering real acceptance over top-down claims.
Alex: So for university no's, it's not just 'here's why,' but 'let's argue until it feels fair'?
Sam: The paper tests this on Parcoursup by checking it against the four models. In the technical one, they highlight a matching algorithm—like sorting kids into teams fairly by preferences and spots—but users want personal advice on their choices, not gear details, so it misses the mark for everyday applicants. Norm-oriented points to legal compliance, yet clashes with folks' belief that top grades guarantee entry, leaving merit-focused families unsatisfied.
Alex: Right, so rules don't match what people expect as fair. How about the expressive and communicative parts?
Sam: Expressive fails too: standardized rejection notes feel dehumanizing, ignoring motivations or internships that applicants value, especially merit believers. Communicative is absent—no real back-and-forth; reasons stay hidden, like claiming "satisfactory" for rejects while admits get "excellent," fueling appeals. The paper notes this breeds stress and distrust from opaque criteria varying by school.
Alex: So even good students appeal feeling cheated. Does applying the framework fix that?
Sam: The analysis shows current tries flop by ignoring recipient needs—like mistaking legal nods for merit buy-in. A full framework use could tailor better, but communicative shines for complex cases: it demands argumentative chats for true agreement. The challenge is building tech for that dialogue amid resource shortages.
Alex: Makes sense—it's about enabling fair debates, not just canned replies. Pulling it all together, how does this framework actually change things for systems like Parcoursup?
Sam: The framework provides a structured way to analyze and improve these systems by matching the right model to the situation—technical for mechanics, communicative for tough buy-in. In the Parcoursup case, it reveals why current notices fail across all four, pointing to dialogue as the path to real acceptance. Overall, it offers a tool to make algorithmic advice more legitimate without overhauling the tech itself. These models are ideal types—simplified patterns to guide thinking, not rigid recipes. The paper relies on this single case study for illustration, without broad testing, and focuses on issuer-recipient talks, sidestepping messier group dynamics.
Alex: Fair point—that keeps it theoretical for now. Still, the practical angle stands out.
Sam: Exactly. Future work could test it empirically, perhaps building chat interfaces into AI for live arguments that cut appeals by building trust. This bridges technical explainability with social philosophy, helping decisions stick in real life. The paper's strength is spotlighting that need without claiming to solve everything.
Alex: Well put, Sam. It's a solid step toward fairer algorithmic choices. Thanks for joining us on ResearchPod.