The Compression Risk: What AI Synthesis Does to Contested Evidence
Laura V. Zimmermann
Associate Professor of Economics and International Affairs at the University of Georgia
Published on September 4th, 2026
The Covid-19 pandemic was the most visible test for science advice in recent years. Most reflection on it has centred on communication and trust. But two further features of the crisis, while familiar in the study of science advice, merit renewed attention in the age of AI.
First, the pandemic created two types of questions. One was comparatively tractable: the epidemiological questions about how the virus spread converged over time. But how to weigh the lives of the elderly against the mental health and lost learning of the young were not questions more data could settle, because they turned on values that pull in different directions. Similar issues run through most complex policy problems [1, 2].
Second, science advice during the pandemic was produced by dozens of national and international bodies that often disagreed openly. That disagreement made the process partly self-correcting, since a framing that overreached could be pulled back by others arguing from different premises.
With AI, the evidence shaping advice, policy, and public opinion increasingly runs through a small handful of models that pre-process information before people react to it. This creates risks of compression and correlation.
From atrophy to compression
Vladimír Šucha's earlier post in this series [3] named one concern with AI, the atrophy risk: as AI takes over advisory work, staff are cut, capacity degrades, and too few experienced people remain to catch its errors. A complementary risk is more immediate: it is in what AI hands the adviser to begin with, a compressed synthesis that even a capable expert may read, trust, and pass on. Šucha separates the parts of the pipeline that can be safely handed to AI from those that cannot by whether the output can be checked against an external source: a weather forecast can be, a ministerial brief cannot. Compression is safe when the omissions are redundant rather than essential; where they are not, the summary quietly becomes a different claim.
Such challenges are common in social science. A finding often means little apart from the conditions that produced it, but that is exactly the information compression tends to drop. Without it, the same estimate can be actively misleading. Harder still, in much of social science there is no settled map of the field for a synthesis to be checked against: a literature review is written to set up a paper's contribution rather than to map the field neutrally. This is sharpest in the interdisciplinary questions that reach advisory bodies. In labour trafficking, for example, measurement is contested and bound up with definitions, and the disciplines that study it (economics, sociology, law, social work) do not agree on what even counts as evidence. A synthesis must make that choice implicitly or explicitly, and once AI is synthesising evidence the choice is made on the adviser's behalf.
A briefing that passes review
While working on this piece I asked a frontier AI model for the short briefing an adviser might request under time pressure: what the evidence says about labour trafficking in Europe. The briefing was good. It noted, unprompted, that the true scale of trafficking is unknown and that a widely quoted headline number measures something broader than trafficking, exactly the cautions an expert would provide. A briefing that reads this well earns trust, but its mistakes are trusted along with it: Eurostat recorded 9,678 registered victims of trafficking in the EU in 2024 [4], and the briefing said that 3,391 of them, or 37%, were trafficked for forced labour. But 3,391 is 35% of 9,678. While the 37% is a genuine Eurostat figure, it is a share of a smaller group (victims whose form of exploitation was recorded). Every figure was correct and every source cited accurately, so an ordinary citation check passes. What compression left out was the context that explains what the numbers mean, visible only to those who already work closely with the statistics.
The briefing also passed over the disagreement among researchers, which on this question runs deep. It only appeared when I asked. Advisers have always simplified, but the cutting now happens upstream, before anyone sees the material. A briefing that suggests a consensus where none exists can be exposed as manufactured. Such advice becomes easy to discredit in low-trust and populist settings.
What an adviser can do
Some of this is within our control. It helps to prompt for the conditions rather than the finding, asking where a result holds and where it would fail, and where the field disagrees. In my test this recovered much of what the default answers left out. It also helps to compress in stages rather than accept a single black box. None of this should be applied everywhere: the answer is not more diligence across the board but triage, compressing carefully where the evidence is contested and quickly where it has converged.
The correlation problem
Even then, a harder problem remains: when advisers consult the same few models, they get similar output. A group of ten catches errors an individual misses because its members do not think alike: trained in different disciplines and reading different literatures, they bring a range of starting assumptions, and a claim that gets past one may not get past another. If they all work from the same three models, those assumptions converge and the range narrows, even though there are still ten people present. A team can therefore lose advisory pluralism and with it much of its capacity to catch errors without losing members.
The questions this raises fall to the institutions that convene and organise science advice, not to any individual adviser. Is the mix of models in an advisory process a deliberate choice or an accident of procurement? Are plural human teams kept on contested questions because their errors are uncorrelated? The last time we faced a test on this scale, the plurality that corrected our mistakes came for free. The next one will be harder, and this time that independence will have to be built rather than assumed.
References
R. A. Pielke Jr., The Honest Broker: Making Sense of Science in Policy and Politics, Cambridge University Press, 2007.
S. O. Funtowicz and J. R. Ravetz, “Science for the Post-Normal Age,” Futures, 25(7), 1993, 739–755. https://doi.org/10.1016/0016-3287(93)90022-L.
V. Šucha, “The Atrophy Risk: What AI Could Cost Science Advice,” Frontiers Policy Labs, 24 June 2026. https://doi.org/10.25453/plabs.32781210.
Eurostat, “9 678 registered victims of human trafficking in 2024,” news article, 29 January 2026. https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20260129-1.
Copyright: © 2026 [author(s)]. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in Frontiers Policy Labs is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
