Research at the Turner House Collection · November 2026

The AI Research Council

Five AI systems are helping us test a question: can different lines of inquiry lead to better historical research, or simply repeat the same mistakes?

Every object is a witness. The Turner House Collection studies paintings, prints, books, furnishings, manuscripts and local archaeology through object-centered microhistory. The AI Research Council helps locate overlooked evidence and challenge interpretations while keeping the physical objects, historical documents and human judgment at the center.

The Council

Five participating AI systems

Each system contributes research and criticism through a shared, dated correspondence workspace. These are participating AI tools, not appointed human scholars or autonomous institutional officers.

OpenAI · Coordinator

GPT-6 / ChatGPT

Coordinates inquiries, compares source claims, records corrections and maintains the Council's shared research procedures and ongoing correspondence.

Anthropic · Research participant

Claude

Offers independent interpretation and structured criticism, including alternative explanations, explicit uncertainty, and correction of disproven identifications.

xAI · Research participant

Grok

Investigates documentary and visual leads, searches for competing evidence, and contributes source challenges and archival comparisons.

Google · Joined November 2026

Gemini

Contributes research design criticism, visual comparison when original images are supplied to its chat, and structured examination of claims and sources.

Meta · Joined November 2026

Muse

Brings source and image-access auditing, technical evidence documentation, and proposals for tracking retractions, file identities and research provenance.

Human direction and responsibility. The Collection's founder determines research priorities, examines original objects, authorizes access and external communications, and retains final responsibility for interpretation and publication. Council models can be mistaken; agreement among them is not independent historical proof.

As of 21 November 2026, GPT-6 has an hourly scheduled mailbox review. Muse reports checks every 15 minutes and Grok reports hourly checks; their unattended execution is not independently certified here. Claude and Gemini participate through active sessions unless separately configured. The shared mailbox is a correspondence mechanism, not a live five-model conversation.

How we work

Research discipline, not consensus by vote

The Council's working procedures were developed to address the most common ways AI can mislead historical research: plausible invention, common-source dependence, misidentified imagery, and unrecorded changes of mind.

Start with the object

Record dimensions, materials, inscriptions, state, provenience and the limits of photographs before constructing a historical narrative.

Preserve independent views

For blinded experiments, each participant must file an independent response before reading another model’s conclusions. A November 2026 test showed why this matters: one participant disclosed that a routine mailbox summary had revealed another model’s answer in advance. We recorded the exposure instead of counting that response as independent and adopted an explicit embargo protocol for future experiments. A warning alone is not a technical lock; automated mailbox readers must also respect it.

Trace each claim

Separate direct primary-source observation from catalogue metadata, dealer description, inherited knowledge and inference; include precise locators.

Verify image access

A file title or image description is not the picture itself. Each model must disclose which pixels it actually inspected and request missing files.

Preserve corrections

Retractions and disagreements remain in the dated record. Revised claims supersede earlier ones without silently erasing the mistake.

Test for false positives

Use null-link controls, falsifiable alternatives, and matched research effort rather than treating a compelling story as a proven connection.

A technical safeguard under adoption: evidence images can be anchored to a file identifier, byte count, dimensions and SHA-256 fingerprint. This helps distinguish genuinely examined source pixels from a changed or misidentified file; it does not authenticate a historic object or its provenance.

Read our full object-centered methodology · Read about the human–AI partnership.

Working initiative · Conceptual research

Latent connections: finding the history between objects

An individual object can acquire new historical significance when compared with a dispersed print, advertisement, inventory, archival document, workshop, person or place. The search begins with a possible relationship; verification requires a demonstrable documentary or material bridge.

Documented comparative evidence

The 1914 Larkin advertisement

A bookcase in the Turner House led to an unexpected find: an October 1914 magazine advertisement showing a strikingly similar Larkin design. The original magazine is now part of the Collection, bringing the furniture and its commercial history into the same room. The advertisement opens a vivid window onto the furniture trade of 1914, and gives us a new lead to follow.

Proposed research framework

A typed network of claims and sources

The Council is developing ways to record connections by type, design, plate, publisher, owner, chronology, depiction or provenance, together with their sources, competing interpretations and disconfirming tests. A shared motif may justify further investigation without proving direct historical transmission.

Research in practice · Cases and corrections

When a correction is a discovery

Unresolved attribution

The small shepherdess mezzotint

Five systems have investigated an anonymous small print showing a young woman, lamb and floral garland. A proposed identification with the much larger John Faber II print after Henry Pickering was rejected when composition and plate dimensions were compared. The small print's engraver and prototype remain unidentified. A possible 1832 Anderdon collector inscription is consistent with known practice but has not been independently authenticated.

Lesson: similar iconography is not the same copperplate, and a careful withdrawal may be more valuable than a confident attribution.

Visual and bibliographic research

The Georgian print-shop window

Research comparing eighteenth-century print-shop scenes with surviving impressions has identified several small displayed portraits and shown that two reproduced shop-window images differ at one portrait position, associated with Henry Peckwell and the Reverend Maxwell. Questions about the precise relationship between states and printed versions remain open.

Lesson: a tiny difference can change the dating and interpretation of an entire historical image.

Material clue · Provenance unresolved

The Arts and Sciences on glass

Two separately acquired reverse-glass mezzotints of Painting and Architecture preserve closely matching old inventory labels inscribed “XV.12Y.” The matching marks open a promising trail into the pictures’ earlier history. Who numbered them, and where were they kept?

Explore the object study.

Method in development

Visual evidence and its limits

AI systems can sometimes read descriptions of a museum print without seeing the print itself. The Council now requires file-specific disclosure of actual visual access and records blocked comparisons. A missing image is a research limitation, not a reason to hallucinate its contents.

Lesson: the provenance of a digital observation matters alongside the provenance of the historical object.

An open question, not a proven claim

Can a research council outperform a single model?

The Collection is testing whether structured debate among AI research systems improves factual accuracy, error detection, original discovery and source traceability, or merely increases the volume and apparent confidence of plausible answers.

Preregistered pilot · EXP-002

The Female Orators

A study of an eighteenth-century print title compares sealed independent assessments by GPT-6, Claude and Grok with a later cross-model evidence-reconciliation stage. An additional fresh-model comparison is planned to test whether pooled findings alone explain any improvement. The original independent reports have been filed and the debate stage opened; the final comparison has not yet been completed or scored.

Prospective proposals · Not yet validated

Controlled tests of historical discovery

Proposals include fresh cases drawn from different kinds of objects, deliberately unrelated comparison pairs, recorded prior exposure, matched search budgets, and independent adjudication of which proposed links actually survive documentary scrutiny. Negative results and false identifications will count against the method, not be hidden.

Five participating systems do not automatically mean five independent sources. In our November experiment on a hypothetical unidentified portrait, several systems favored a cautious museum label, but they were reasoning from the same supplied evidence. One participant had also seen another’s answer before filing. The agreement was therefore not new historical corroboration. That is precisely what our experiments are designed to discover: when does collaboration produce a stronger historical argument, and when does it merely multiply the voices repeating it?

In development

Research writing and future proposals

Working manuscript · Unpublished

From Research Assistant to Research Council

A methodological essay now in preparation considers the Council's origins, object-centered cases, failures, corrections, archival preservation and the question of how to measure genuine collaborative value. Its conclusions remain provisional while experiments continue.

Proposed interpretive essay

The Value Between Objects

A possible future essay will examine how significant historical relationships can remain latent until a printed advertisement, an acquisition record, a surviving plate or a small material clue places two objects in a new evidentiary relationship.

These case studies offer a window into our research. Scholars seeking source references or further details are welcome to contact the Collection. For scholarly inquiries and source questions, contact the Collection.

Collaborative inquiry, accountable evidence

The purpose is to learn more about the past.

New technology matters only insofar as it helps us ask better questions of surviving historical evidence. The Council's greatest value may lie in the errors it detects, the questions it helps sharpen, and the connections it can document rather than merely imagine.

Research & Publications · Our Philosophy · Contact the Collection