Research note · Research note 09 — current edition
The Practical Limits of AI Research Automation
Written by Aaron Grainger
Independent Content Strategist & Product-Marketing Writer · Published Apr 8, 2025
- Primary audience
- Product managers building AI features
- Also useful for
- AI engineers and technical founders
- Tone
- Authoritative
- Reading time
- 2 min
- Published
- Apr 8, 2025
Abstract
Automated research is good at breadth and weak at judgement. This note sets out the specific judgements that resist automation — source credibility, conflicting evidence, and the difference between absence and non-existence — and suggests where a human belongs in the loop.
Key questions
- Which research judgements can be encoded as rules?
- How should a system handle sources that disagree?
- What is the correct behaviour when evidence simply is not available?
How this note was developed
- — Decomposition of a research task into decision points.
- — Assessment of which decisions have stable criteria and which are contextual.
- — Independent analysis; no evaluation of any specific research product is made.
Central arguments
01Credibility is contextual, so rules generalise badly
A vendor page is authoritative about its own pricing and unreliable about a competitor's. Encoding credibility as a per-domain score ignores the claim type, which is usually the thing that determines whether the source is appropriate.
02Conflict should be surfaced, not resolved
When credible sources disagree, presenting the disagreement with both sources is more useful than picking a winner. Automated resolution tends to favour the most recently retrieved or most confidently worded source, neither of which is a truth signal.
03Absence of evidence is the hardest case
A system that searched and found nothing cannot distinguish 'not published' from 'not found by this query set'. The honest output states what was searched and what was not found, which requires the search plan to be part of the output.
04Humans belong at framing and at acceptance
Deciding what question to ask and deciding whether the answer is good enough to act on are the two points where automation adds least and costs most when wrong.
Limitations
- — This note reflects a conservative stance; other practitioners reasonably automate more.
- — Model capabilities change, and some limits described here may narrow.
Related reading
Related content
Research note · 2 min
The Web Context Problem
Most AI products that reason about the outside world fail for an unglamorous reason: the material they reason over is incomplete, stale, or structurally unusable. This note argues
Research note · 2 min
The Anatomy of a Source-Linked AI Answer
A source-linked answer is not an answer with links appended. This note breaks a well-formed answer into its parts — claim, support, provenance, confidence, and refusal — and argues
Research note · 2 min
Why Clean Content Beats Raw HTML for Most AI Tasks
Raw HTML is a rendering instruction set that happens to contain text. This note examines what is lost and gained when a page is normalized to structured Markdown, and identifies th
Research note · 2 min
The Hidden Maintenance Burden of DIY Web Data Pipelines
The first version of a web-data pipeline is usually a weekend. The cost arrives afterwards, in silent breakage, template drift, and the operational question of who notices when a s
Version history
Current: 1.0 · Published
- 1.0Apr 8, 2025First published.
Was this useful?
Sourceframe is an independent product concept created for research, product-design, and technical-content exploration. It is not an operating company, and nothing here describes a live commercial service. All examples, schemas, and code are illustrative unless a page says otherwise. No client data, customer outcomes, performance results, or partnerships are described anywhere on this site.