Ronit Levine-Schnur, Renana Keydar and Noam Maeir
Governments increasingly promise to “consult” the communities their decisions affect. Infrastructure projects, land-use plans, resource extraction — the language of consultation is everywhere, and it is often held up as evidence that a process is fair, inclusive, even reconciliatory. But what actually happens to what people say once they’ve said it? In our forthcoing article, Narrative Power and Justice: Computational Insights into Indigenous-State Consultation Dynamics (Vol 54, Issue 4, 2026), we set out to answer a narrower, more uncomfortable version of that question: when a state invites an indigenous community to speak, does it actually listen — and to what?
We studied two Israeli state-led consultation processes involving the Bedouin community of the Negev: the 2008 Goldberg Committee, tasked with recommending a comprehensive policy for Bedouin settlement, and a parallel set of public hearings on objections to a regional metropolitan plan. Neither process carried a binding legal duty to consult, and neither obligated the state to act on what it heard. What both processes shared was a formal, public invitation to speak — and, crucially, a searchable, word-for-word record of what was actually said. Over 200 people testified, generating more than 720,000 words. That record let us ask a question that is usually answered only impressionistically: not just whether people were allowed to speak, but what they chose to say, and whether it mattered.
To answer it at scale, we turned to topic modeling — a computational technique, borrowed from natural language processing, that identifies clusters of words that tend to appear together across a large body of text. Rather than starting with categories we expected to find and coding testimonies against them, we let the model surface patterns inductively, and only then interpreted what it found in light of the legal and historical context. Think of it as a way of “distant listening”: instead of close-reading a handful of testimonies, we could observe patterns across all 217 speakers at once, while still being able to zoom back in on individual voices to check that the patterns held up on the ground.
What we found was a strikingly consistent divide. Bedouin speakers, across both processes, overwhelmingly framed their testimony around family and tribal land possession — the historical, communal fact of having lived on and worked this land across generations. That theme was over eight times more likely to appear in Bedouin testimony than in non-Bedouin testimony, holding other factors constant. Government officials and other majority-group speakers, by contrast, spoke a different language almost entirely: legal entitlement, planning procedure, administrative process — a vocabulary oriented toward resolving the future rather than reckoning with the past.
This is not simply a difference in emphasis. It reflects two competing theories of what justice in this setting requires. For Bedouin speakers, the past is not backstory — it is the foundation of the claim itself. You cannot discuss where people will live tomorrow without first acknowledging where, and how, they have lived for generations. For the state, by contrast, history is largely bracketed. The consultation is framed as solving a forward-looking governance problem, and legal and procedural categories become the tools for solving it — categories that, as our close reading of the testimony shows, were also used quite explicitly to characterize Bedouin claims to communal and customary land tenure as illegitimate or “problematic” on their own terms.
The most sobering finding concerns what happened next. When we ran the same computational analysis on the Goldberg Committee’s final report, we found that its thematic composition mirrors the state’s register almost exactly: dominated by legal land rights and procedural planning, with communal land possession — the theme that defined Bedouin testimony — reduced to a minor thread. The people were heard, in the literal sense that they were given a platform and their words were transcribed. But the institutional act of translating testimony into policy recommendations filtered out the very narrative that indigenous speakers had most insistently offered.
We think this matters well beyond the Negev. The “duty to consult” indigenous communities is spreading globally — through UNDRIP, through domestic jurisprudence in places like Canada, through ad hoc commissions and planning bodies far more informal than a constitutional right. Almost all of these mechanisms share the same structural feature: an obligation to hear, but not to act. That makes the practice of listening — not the fact of the microphone — the real site where recognition is won or lost. Our data suggests that listening, left unexamined, defaults to the institution’s own vocabulary. Legal and administrative language doesn’t just describe outcomes; it quietly decides which claims are even legible as claims.
That has a practical implication we did not expect to end up defending, but do: computational text analysis could itself become part of the accountability infrastructure of consultation. If a report’s thematic composition can be measured against the testimony it claims to summarize, discrepancies of the kind we found are no longer just a matter of interpretation or advocacy — they are measurable. Appending that kind of analysis to official reports would not force anyone to agree with what they heard. But it would make it much harder to simply not notice.