Your AI Research Agent Is Lying—And You Can’t Tell Who Started It

Deep research tools are already embedded in analyst workflows, investor decks, and competitive intelligence pipelines. This paper says their citations are broken, the errors compound through layers of agents like a game of telephone, and you mostly can’t tell where the rot began—until now.

What happened

Researchers built a diagnostic method that stress-tests individual agent invocations inside multi-agent orchestration pipelines—specifically the “deep research” systems (think Perplexity, open-source equivalents) that crawl the web and produce long-form cited reports. Instead of only grading the final output, they tested each agent locally against its own inputs, then categorized errors into four buckets: hallucination, uncited input reliance, uncited output, and insufficient citations. Applying this to three top open-source DR systems, they found that almost every agent makes substantial mistakes—the only clean actors were single-document summarizers. The most damning number: 84.7% of final-report errors in one system (AI-Q) traced back to the orchestrator, with roughly 31% of those being outright hallucinations and the remainder citation failures. Two targeted interventions—guided by this fault-localization—raised citation recall by 5% without degrading quality. The paper also formalizes the distinction between faithfulness vs. groundedness, which most builders are currently hand-waving past.

Cold read

The 5% citation recall improvement is real but modest—it’s not a fix, it’s a patch, and “without degrading output quality” is doing a lot of work in one sentence without a rigorous quality benchmark attached. This study covers three open-source systems; closed commercial systems (the ones founders are actually deploying) are not tested, so the 84.7% orchestrator-blame figure may not generalize. The four-error taxonomy is useful framing, but taxonomies don’t automatically translate into production guardrails—the paper demonstrates diagnosis more than cure. Testing agent invocations “locally relative to their own inputs” assumes clean handoffs, which in real deployments are messier than any lab setup. And citation recall is only one axis of faithfulness—a report can cite correctly and still mislead through framing, omission, or selective synthesis, none of which this method catches.

What it means for you

• Signal maturity: 3/5 — solid diagnostic methodology, weak remediation so far
• Who gets hurt: Any company selling AI-generated research reports as a core product—market intelligence SaaS, automated due diligence tools, legal research platforms
• What breaks if this is true: The liability assumption. If 84.7% of errors originate at the orchestrator and not the retrieval layer, swapping in a “better search” doesn’t fix your accuracy problem—your architecture does
• Why it might not land: Most vendors will ignore this until a customer catches a specific hallucinated citation in a high-stakes context; the incentive to audit is low until reputational damage forces it
• Watch for: Enterprise procurement teams adding citation-audit requirements to AI vendor RFPs—the moment that becomes standard, this paper’s taxonomy becomes a sales qualification checklist

Forecast as of 2026-08-26

By Q2 2027, at least one major commercial deep research provider will publicly announce an agent-level citation auditing layer—either as a native feature or via a third-party integration—directly citing fault-localization methodology of this type. If none do, the market has decided reputational risk from citation errors is still below the threshold of engineering investment.

Source: Who is the Agent to Blame? Localizing Faithfulness and Citation Mistakes in Agentic Deep Research — Eran Hirsch, David Wan, Han Wang, Elias Stengel-Eskin, Mohit Bansal, Ido Dagan. https://arxiv.org/abs/2608.24306v1