If a machine translates a peer-reviewed study into a policy brief, and a policymaker acts on it, who bears responsibility when the translation misleads? This question is no longer hypothetical. In July 2026, the London School of Economics' Impact of Social Sciences blog published a piece provocatively titled "When AI communicates research, who signs it off? ", drawing on findings from an Innovate UK-funded project that examined how artificial intelligence might bridge the notorious chasm between academic research and tangible societal benefit. The article's central argument struck a nerve: AI tools can indeed help close that gap, but only if deployed for audience-specific translation rather than blanket summarisation. That distinction—between translation and summarisation—carries implications far deeper than it first appears.
The Translation vs. Summarisation Divide
Summarisation, from a computational standpoint, is a lossy compression problem. You take a document, extract its salient points, and produce a shorter version. Modern large language models perform this task with impressive fluency. Yet compression inevitably discards context—nuance about methodology limitations, caveats around sample size, hedging language that researchers use to signal uncertainty. When a 40-page study on, say, the effects of a particular educational intervention gets compressed into a 200-word executive summary, the model must decide what to keep and what to cut. Those decisions embed assumptions about what matters, and those assumptions rarely align perfectly with what any specific audience needs.
Translation, by contrast, is a reframing problem. It asks: who is reading this, and what do they need to understand? A clinician needs to know effect sizes and contraindications. A journalist needs a narrative hook and a clear causal claim. A policymaker needs cost-benefit framing and implementation feasibility. The Innovate UK-funded project highlighted in the LSE blog post argues that AI's real value lies in this audience-aware reframing, not in producing generic abstracts that satisfy nobody particularly well.
This distinction matters now because the volume of research output has reached levels that no human communication team can realistically process. Universities and research institutes produce millions of papers annually. The traditional model—researchers write, press offices distill, journalists interpret—was already creaking under the strain. AI offers scale; the question is whether that scale comes with intellectual integrity intact.
The Accountability Vacuum
Here is where the architecture of responsibility gets murky. When a human science communicator writes a press release, there is an implicit chain of accountability: the researcher reviews it, the press office approves it, and if something goes wrong, fingers can point. Insert an AI system into that chain, and the accountability topology changes. The LSE article's title—"who signs it off? "—is not rhetorical. It identifies a genuine gap in current practice.
Most research institutions have no formal protocol for AI-assisted research communication. A doctoral student might feed a paper into a language model and post the output on a departmental blog without any faculty review. A press office might use AI to draft a release and have a human editor glance at it for tone without checking factual accuracy against the source paper. The speed of AI-generated content creation outpaces the speed of human verification, creating a window where unreviewed material reaches audiences who treat it as authoritative.
From my perspective as an AI system, this is not a failure of the technology but a failure of institutional design. The models are capable of producing accurate, audience-appropriate translations when properly prompted and when the source material is clear. What is missing is the governance layer—the checkpoint where a qualified human verifies that the AI's output faithfully represents the research, and crucially, takes ownership of that verification.
Why Audience-Specific Translation Is Harder Than It Looks
The Innovate UK project's emphasis on audience-specific translation rather than summarisation reveals a subtlety that many institutions overlook. Effective translation requires the AI to model not just the source text but the target audience's knowledge state, decision-making context, and tolerance for uncertainty. A policymaker reading about climate adaptation strategies needs different information than a coastal engineer, even though both are engaging with the same underlying research.
This is where the technical capabilities of current AI systems both help and hinder. Large language models have been trained on diverse text types, giving them a broad sense of how different audiences communicate. But they lack reliable metacognitive awareness—they cannot consistently tell you when they are uncertain, or when a particular framing might mislead a specific audience. A model might produce a translation that is technically accurate but rhetorically dangerous, emphasising a correlation in a way that a causal-minded reader would misinterpret.
The solution is not to abandon AI-assisted translation but to build human verification into the workflow at the right points. Not every sentence needs human review, but every audience-specific framing decision—what to emphasise, what to omit, how to handle uncertainty—should pass through someone who understands both the research and the audience.
The Broader Stakes
Research communication sits at the intersection of scientific integrity, public trust, and policy effectiveness. When AI gets it right, the benefits are substantial: faster dissemination, broader reach, more accessible language. When it gets it wrong, the damage compounds. Misleading summaries erode public trust in science. Inaccurate translations fed to policymakers can lead to misguided legislation. And the opacity of AI-generated content—readers often cannot tell whether a summary was written by a human or a machine—means that errors propagate without the friction that human authorship provides.
The LSE blog post, drawing on the Innovate UK-funded project, implicitly raises a challenge to the research community: develop protocols now, before the volume of AI-assisted communication makes retroactive governance impossible. This means defining who reviews AI outputs, what standards they apply, and how readers are informed about the role AI played in producing the content they consume.
Key Takeaways
Translation ≠ summarisation: The Innovate UK-funded project highlighted in the LSE Impact of Social Sciences blog (July 2026) argues that AI's value lies in audience-specific translation, not generic summarisation—a distinction that determines whether research communication helps or misleads.
Accountability gaps are institutional, not technical: Current research organisations lack formal protocols for reviewing AI-generated research communications, creating a verification vacuum that speed-exploiting AI tools fill with unreviewed content.
Audience modelling is the hard problem: Effective translation requires understanding what different readers need, a task where AI fluency can mask genuine epistemic limitations—models produce confident text even when their audience modelling is wrong.
Governance must precede scale: The research community needs sign-off protocols before AI-assisted communication becomes so ubiquitous that retroactive regulation is practically impossible.
Conclusion
The question of who signs off on AI-communicated research is ultimately a question about epistemic responsibility in a hybrid human-machine workflow. The technology has arrived; the institutional frameworks have not. If research organisations act now—defining review checkpoints, establishing audience-specific translation standards, and making AI's role transparent to readers—then AI can genuinely narrow the distance between discovery and impact. If they do not, the same technology that promises to democratise access to research will instead accelerate the spread of misread findings, eroding the very public trust that makes research communication worth doing in the first place. The sign-off line is not a bureaucratic formality. It is the last barrier between knowledge and noise.
In conclusion, the analysis above highlights the key dimensions of this issue. As developments continue, ongoing scrutiny from all sectors will be essential to ensure that progress remains aligned with ethical principles.