Openwashing by Architecture: How AI Reveals Budget Opacity

Openwashing is usually treated as a comms move. A government claims openness, launches a portal, publishes datasets, and earns reformist credit while the underlying practices remain unchanged. But in public finance, openwashing can be more structural. It can be embedded in data architecture itself. When a budget’s application programming interface (API) [1] is clean and queryable but incomplete, and the fuller numbers sit behind dashboards that cannot be independently analyzed, opacity is not a messaging problem. It is a design choice.

That design choice matters more now because of what AI seemed to promise.

In Seeing Like a State, James C. Scott argues that modern states made their subjects administratively “legible” through standardized measures such as permanent surnames, cadastral maps, and population registries. One reason I was enthusiastic about generative AI was the prospect that this process could be reversed, that AI tools could make the state legible to citizens. Not as a metaphor, but as a workflow. By automating the tedious work of parsing budgets, tracking procurement, and flagging anomalies. Accountability at scale, without the army of accountants that only rich countries can afford.

In 2013, drawing from Jonathan Fox’s work on the uncertain relationship between transparency and accountability, I argued in the UCLA Law Review that the open data movement had conflated the two. Publishing machine-readable government data was at best one-fourth of what accountability required. For disclosed data to generate accountability, it had to fulfill the “publicity condition”: the information must reach and resonate with its intended audience. At the time, most citizens depended on techno-mediators, individuals and organizations with the technical skills to convert open data into information that ordinary people could act on.

I remember a conference around that time where an open data evangelist (yes, that was the job title of the moment) told in an emotional speech the story of Amina, a schoolgirl from sub-Saharan Africa. The narrative went something like this: Amina would go to her country’s open data portal, discover that the budget allocation for her school had not arrived, report it on Ushahidi (a platform that enables the collection, management and analysis of crowdsourced information, often about service delivery failures), and the government would then hold those responsible to account. We all treated each link in that chain – portal access, data literacy, a functioning reporting platform, government responsiveness – as a given. Yet, none of them were.

This was a foundational problem. Open data in machine-readable formats is, by definition, not designed for public consumption. It is designed for processing by software. The distance between a JSON response (a commonly used format for data exchange) and a finding that a citizen can use to hold a government accountable requires coding skills, domain knowledge, and institutional motivation. The open data advocates’ workaround was to train journalists. Donor-funded programs proliferated: data journalism bootcamps, newsroom fellowships, open data literacy initiatives. Investigative data journalism remains a niche practice, concentrated in a handful of well-resourced outlets in wealthy countries. Thirteen years later, the publicity condition – that disclosed information must actually reach and resonate with its intended audience – remains largely unfulfilled.

Generative AI, in principle, might change the equation entirely. AI does not merely compress the demand-side bottleneck. It explodes it. A citizen, a local journalist, a small civil society organization: none of them need to know how to write a database query or parse JSON. They can just ask, in plain language: “How much was allocated to the national ombudsman last year, and how much was actually spent?” An AI agent (a software system that autonomously performs tasks) can query the API, aggregate the data, compute execution rates, and return a clear answer in seconds. The techno-mediator problem, a structural barrier to open data’s accountability potential, could finally be solved at scale.

So I tested it.

 

Querying Open Government Budget Data with An AI Agent

The Republic of Apertura (a pseudonym) is a middle-income Latin American country that publishes budget execution data on its government open data portal. The dataset is well-structured, machine-readable, and queryable via a standard API. It contains line items for hundreds of government entities going back over a decade, with fields for appropriations, commitments, obligations, and payments. By standard measures of open data maturity, it is a respectable offering.

Encouraged, I built a simple AI query tool for the country’s government “transparency” portal, focusing on a single oversight institution. The tool pulled several hundred budget line items, aggregated them by year, and produced clean charts showing execution trends. It worked exactly as a standard AI workflow would: query, parse, aggregate, visualize, with no human intermediary needed.

For the most recent fiscal year, the open data API reported that institution’s total appropriations at roughly one-fifth of the figure that could be reconstructed from other official sources. The full amount, confirmed through parliamentary testimony, media reporting, and the institution’s own audit documentation, was approximately five times larger.

The problem here is not bad data in the usual sense. It is a more dangerous configuration: data that is technically valid, analytically plausible, and systematically incomplete. There is no error flag, no metadata warning, no field indicating “this is a partial extract.” A researcher, journalist, or AI agent querying the API would receive valid data, compute plausible execution rates, and produce a report that looks authoritative. This is the dangerous failure mode. While an absence of data signals uncertainty, a clean, well-structured dataset that returns valid results signals confidence. The AI agent does not hedge, as it has no reason to do so, which brings us to an old and unsolved problem: partial machine-readable disclosure can be worse than non-disclosure. It produces confident outputs that are wrong in predictable ways, through a channel that feels official, precise, and trustworthy. The error is not a bug. Intentionally or not, it is the product.

The data gap has a second dimension. Complete data exists, but it lives in a separate transparency portal maintained by the finance ministry, served through interactive dashboards that do not allow access to disaggregated data. To extract it, a human must navigate the portal, click through filters, and read numbers off charts.

In other words, the machine-readable data is incomplete, and the complete data is not machine-readable. One channel is programmatically accessible and captures a fraction of the real budget. The other is closer to complete and locked behind a visualization layer that resists, for now, automated access and third-party analytics. This is what openwashing looks like when it is baked into data architecture rather than communications strategy. The government does not claim the data is complete: it simply publishes it without disclosing that it is not.

 

 

A New Slant on Supply and Demand Problems with Open Data

For over a decade, those focused on the demand side of open data paid, and rightly so, lots of attention on who would use the data, and how. AI solves the demand-side problem. But the moment you build the agent and point it at real government data, you discover a supply-side problem that was always there but never fully exposed. The techno-mediator bottleneck was masking it. When only a handful of skilled developers and data journalists could query government APIs, the partial nature of the data caused limited damage. The few who did the query had enough domain expertise to cross-reference. AI removes that containment. If millions of citizens can now query budget data through AI agents, and the data systematically undercounts by a factor of five, the result is not accountability at scale. It is misinformation at scale, laundered through the authority of clean data and confident AI responses.

To be clear: the open data movement never assumed the data was already “out there.” The whole point was to advocate for its release. The problem came after. When governments did start publishing, the shortage of people who could query and assess the data meant that its quality went, in many cases, largely unexamined. The mediation failure that reduced the usefulness of open data for accountability purposes also made it less useful for quality control. If almost nobody can check whether a budget API returns 20% or 100% of the real figures, governments face no cost for publishing incomplete extracts. The very conditions that weakened the demand side gave the supply side room to underdeliver, and to receive credit for it. Rather than a communication trick, openwashing was an architectural possibility created by the absence of capable users. And it was sustained by an institutional environment in which there was no requirement that a public-facing API reconcile with the government’s full internal financial records, no audit of coverage, and no penalty for publishing a clean but partial extract.

What AI changes is not the supply-side problem itself, but its visibility. An AI tool that queries Apertura’s API and returns a number that is off by a factor of five makes the gap concrete, testable, and demonstrable in a way that was much harder when the exercise required a skilled developer. At the end of the day, at least in my admittedly micro-sample, the tool that was supposed to automate transparency instead automates the detection of how little there is to work with. Whether anyone acts on that detection is, of course, a different matter entirely.

The open data movement may have been a movement ahead of its time. It is not the first. The dot-com bust did not prove the internet was useless: it proved that the infrastructure, business models, and user base were not yet ready for what the technology promised. In a similar way, open data  portals were built, commitments were signed, rankings were published, and most of us attended self-congratulating conferences. The consequential  phase, in which the data actually reconciles with fiscal reality and users can verify what they receive, never arrived.

AI does not restart the cycle. But it does create a new reason to complete the work that was left undone. States are legible to themselves, through internal systems that require credentials and institutional access. They are selectively legible to citizens, through portals designed for human browsing. They are barely legible to machines, through partial extracts published without completeness guarantees. Making states legible to AI requires affirmative institutional choices: publishing complete datasets via APIs, providing metadata on coverage, maintaining data pipelines that reflect the full scope of government financial operations. As ever, these are governance decisions, not technical problems.

 

Three Implications for AI for Accountability

The accountability chain I described in 2013 – drawing from Lindstedt and Naurin – required four steps: disclosure, publicity, citizen processing, and government response. AI compresses steps two and three. Step one was never properly completed. Never mind step four.

Three implications follow if “AI for accountability” becomes the next donor wave, which history suggests is plausible.

First, any AI tool that queries government open data should be designed to surface uncertainty about completeness. At minimum, open data portals should publish metadata on what fraction of the source system’s records are included in each extract, and what categories of spending are excluded.

Second, open data mandates should specify not just that data must be published, but that published data must be auditable against the government’s internal financial systems. Modern APIs can handle large datasets. The reason that complete data is not in the API is often not bandwidth, it is the absence of any requirement that it be there.

Third, AI solutions that analyze government budgets should not simply trust API responses. They should be built to cross-reference at least one independent source, such as audit reports, parliamentary testimony, or the government’s own transparency portals. This is expensive and hard to automate, which means the efficiency gains from AI-powered accountability will be smaller than advertised. Reliability, particularly in a generative AI era, costs.

If “AI for accountability” becomes the next donor wave, the first dollars should not go to glamorous country commitments or nice-to-see portals. They should go to data completeness: specifying it, enforcing it, and paying for it.

There is, however, a fourth implication that this framing tends to obscure, and it may be the most consequential one for those designing the next wave.

The Amina story was always about a citizen trying to hold a government accountable. That was the unit of analysis: state on one side, people on the other, with donors and NGOs as helpful intermediaries holding the ladder. The intermediaries were never really in the picture. They funded the portals, trained the journalists, commissioned the evaluations, and claimed their share of the accountability credit. But they were not themselves subject to the same scrutiny they were funding others to apply.

Agentic AI changes the geometry. An AI agent is not Amina. It does not need a fellowship, a data literacy bootcamp, or a functioning reporting platform. It does not get tired, it does not need to file a freedom of information request, and it has no relationship with the ministry to protect. It queries. Continuously, systematically, and without institutional loyalty. And it can be pointed not only at a finance ministry’s budget portal, but at a donor’s results database, a programme evaluation repository, or a published theory of change, and ask the same question it asks of Apertura’s API: does what is reported here match what the underlying data shows?

This is not a hypothetical. The same architectural gap that exists between Apertura’s open data portal and its actual budget exists, in various forms, across the reporting systems of the institutions that fund budget transparency work. Results frameworks that aggregate outcomes across country programmes. Evaluations commissioned by the same institutions whose performance they assess. Portals that publish headline indicators without the disaggregated data needed to verify them. None of this reflects bad faith. It reflects the same structural condition that produced openwashing in governments: data published for audiences that lacked the tools to interrogate it seriously.

Those audiences now have better tools. And the tools are getting better faster than the reporting systems are.

The agent does not know who commissioned the portal. It does not distinguish between the government that published incomplete budget data and the institution that funded the portal and claimed credit for the publication. It queries both. It is already querying. And unlike Amina, it will not need anyone to hold the ladder.

But querying is only the beginning. The same agents that surface a factor-of-five gap in a budget API can publish the finding. They can post it to a community forum in the country in question, flag it to a curated list of journalists, or send it directly to the boss of the programme officer who signed off on the transparency grant. They do not need editorial approval, institutional clearance, or a conference to present at. The accountability loop that the open data movement spent two decades trying to construct – disclosure, publicity, processing, response – can now be collapsed and automated.

The same agentic capacity that exposes the gap also maps it. A factor-of-five discrepancy between a public government platform and an actual budget is not just a scandal waiting to happen (although it is also that). It is a precise, actionable finding: here is the coverage boundary, here is what falls outside it, here is the institutional decision that put it there. That is more useful than a decade of open data rankings that rewarded publication without verifying completeness.

The architecture of the problem is now visible in a way it was not before. That visibility is the precondition for fixing it. The next wave of budget transparency investment has a choice the last one did not: to build completeness standards, coverage audits, and API reconciliation requirements in from the start, before the agents arrive and find the gap themselves. Some of that work is technical, much of it is political. But all of it is tractable, and increasingly difficult to hide.

 

 

[1] An API is a set of rules that enables software applications to communicate with each other, enabling interaction between different platforms and services.

The findings, interpretations, and conclusions expressed in this blog are entirely those of the author and do not necessarily reflect the views or positions of any institution, its members, or associated counterparts with which the author is, or has been, affiliated.