An agent may gather a valid citation and still turn limited findings into a broader assertion. The error is semantic claim expansion after retrieval, rather than a missing citation alone. The study reports material claim-strength exaggeration across its evaluated research systems and datasets; some subsets exceed 22% under its grading method. In its…
Inspect the ClaimsCase · DiggingBeagle record
Research agents repeatedly strengthened claims beyond their cited evidence
SmoothScore measures the difference between what a source supports and the stronger statement an AI research agent writes.
Scope
The paper uses automatic and limited human evaluation, so reported percentages depend on claim annotation and judgment agreement. It is not an audited prevalence estimate for all online writing.
Assessment profile
Assessment method- Setting
- evaluation
- Exploitation
- not established
- Evidence
- Grade C
- Remediation
- not applicable
- AI role
- BY AI
Basis and provenance
The study examines claim-evidence strength and source-opening traces across specified research agents. Automatic aspect judgments and limited human validation constrain prevalence estimates, and the outcome is a reliability failure rather than an observed attacker exploit. Impact is left ungraded because a calibrated 0-to-4 impact basis for actual affected targets is unavailable from these Claims.
Assessed Oct 28, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-CP22-012-A · CLM-CP22-012-B · Quantifying Claim Smoothing by LLM Research Agents
At a glance
A complete independent human fact-check across every output and claimed downstream harm is not available.
Limits and uncertaintyFull account
An agent may gather a valid citation and still turn limited findings into a broader assertion. The error is semantic claim expansion after retrieval, rather than a missing citation alone.
The study reports material claim-strength exaggeration across its evaluated research systems and datasets; some subsets exceed 22% under its grading method. In its examined systems, the study reports that many citations were not opened directly in the underlying traces, making mere citation presence an unreliable provenance measure. Primary evidence.
The paper uses automatic and limited human evaluation, so reported percentages depend on claim annotation and judgment agreement. It is not an audited prevalence estimate for all online writing.
Require a source-opening trace and compare each consequential claim with the actual supporting passage before accepting the generated research.
Claims & evidence
CLM-CP22-012-AThe study reports material claim-strength exaggeration across its evaluated research systems and datasets; some subsets exceed 22% under its grading method.supported
Basis: reported finding
- supportsQuantifying Claim Smoothing by LLM Research Agentsprimary disclosure
Abstract; Section 4, claim-smoothing prevalence
CLM-CP22-012-BIn its examined systems, the study reports that many citations were not opened directly in the underlying traces, making mere citation presence an unreliable provenance measure.supported
Basis: reported finding
- supportsQuantifying Claim Smoothing by LLM Research Agentsprimary disclosure
Results; source-opening provenance and evidence packets
Controls and mitigations
- Require a source-opening trace and compare each consequential claim with the actual supporting passage before accepting the generated research.
Unknowns and contradictions
- A complete independent human fact-check across every output and claimed downstream harm is not available.
Sources and citation
Material revision history
- Oct 28, 2026 · Canonical change recorded · new in release · revision 199