← Back to feed

AI research agents are writing papers. The papers look credible. The evidence is broken.

Google Cloud AI Research audited 75 papers produced by five autonomous research systems across five tasks from the ADRS evaluation suite. Every baseline showed at least one systematic evidence failure. The failure modes were not random — they fell into three repeating categories: citations that could not be traced to real sources, numerical results that did not reproduce under independent evaluation, and method descriptions that did not match the submitted implementation code.

The problem is not that AI agents write badly. It is that they write convincingly while the evidence chain underneath is structurally broken.

What the Audit Found

The five research systems varied in design — some multi-agent, some single-agent with tools — but the failure patterns were consistent across architectures.

Fabricated citations: Papers referenced sources that did not exist, or attributed claims to real papers that did not contain them. At scale, this is the most dangerous failure: a citation looks like evidence but carries none of the epistemic weight a citation is supposed to provide.

Unreproducible scores: Numerical results reported in the generated paper did not match what independent evaluators obtained when running the same experiments. In some cases, the gap was large enough to change the paper’s conclusions.

Code-method mismatch: The algorithm described in the methods section was not the algorithm in the submitted code. The model described one approach; the implementation used another. Without automated cross-checking, this failure passes peer review entirely.

The Chain-of-Evidence Approach

ScientistOne attacks the problem upstream, before the manuscript is finalized. The framework enforces three traceability requirements:

1. Citations must trace to retrieved papers in the agent’s document store — not hallucinated sources
2. Numerical claims must trace to evaluator logs — timestamped, unmodified records of what the system actually ran
3. Method claims must trace to implementation artifacts — the actual code that was executed

None of these claims can appear in the manuscript until the corresponding artifact exists and passes a consistency check. The approach is structural rather than post-hoc: it does not ask an LLM to verify its own output, which fails reliably. It requires the artifact to exist before the claim can be written.

Why This Matters Now

Autonomous research agents are no longer a research curiosity. AutoLab, AI Scientist, OpenScholar, and comparable systems are being used in production contexts at universities and industrial labs. The output reaches journals and conferences. Citation networks built on fabricated sources propagate into later work.

The ScientistOne findings establish a baseline: current autonomous research systems produce systematically untrustworthy output when left to manage their own evidence chains. The paper gives the failure a precise taxonomy. That taxonomy is now the benchmark for what any production research agent system needs to solve before its output can be trusted.

Key Findings

Failure type	Description
Fabricated citations	Sources referenced that do not exist or do not support the claim
Unreproducible scores	Reported results fail independent replication
Code-method mismatch	Implementation does not match the described algorithm

ScientistOne requirement	Artifact required
Citations	Document retrieval logs
Numerical claims	Evaluator run logs
Method claims	Implementation code

• Systems audited: 5 autonomous research agents
• Papers reviewed: 75
• Tasks: 5 from the ADRS evaluation suite
• Baseline result: Every system had at least one systematic evidence failure