Flagship programme · In preparation for an ERC Starting Grant

AI, Material Traces, and Social Worlds

A programme to build the first interpretable pipeline that infers social difference from the chemical and material traces of past communities, and to study, ethnographically, how those inferences become accepted or contested knowledge.

Discuss collaboration Funding context

The core question

One question that requires all three disciplines

How is social knowledge about human difference and inequality, including diet, mobility, status, and exchange, produced, stabilised, and contested when interpretable AI is introduced into the reading of the chemical and material traces of past communities?

Sociology remains the home discipline, while mass spectrometry and AI serve as the indispensable instruments rather than the subjects. The question is answerable only by this team: without the chemistry there is no signal, without interpretable inference at scale there is no reading of it, and without ethnography there is no account of how a reading becomes, or fails to become, accepted knowledge.

The ground-breaking claim

AI as a new apparatus of evidence

Stated as a claim rather than a topic: AI is not an accelerator of archaeometric analysis but a new apparatus of social-scientific evidence.

The programme will build the first interpretable pipeline that infers social difference from mass-spectrometry and isotope data, and, by ethnographically tracing how those inferences are validated and contested, will establish a general account of when AI-mediated interpretation hardens contestable inequalities into apparent chemical fact, and when it opens them to revision.

Why it is new

The sociology of knowledge has theorised how facts stabilise in laboratories, but not the case of AI-mediated material evidence.
Biomolecular archaeology reads social variables from chemistry, but interprets them manually, without an explicit, auditable step from signal to claim.
Machine learning is entering both fields, but in fragmented, classification-oriented form. No one has joined these into a single method and theory.

Empirical cases

Two material-evidence cases and one reflexive case

The cases below are proposed for fit and feasibility. Each depends on partner agreements that will be confirmed before submission. Partner names shown are intended collaborations, not yet finalised.

Case 01

Mobility & diet in the Western Balkans

Isotope analysis for mobility and lipid residue analysis for diet on skeletal and ceramic assemblages from Albanian sites, reading migration status, socioeconomic position, and regional difference in EU-legible terms.

Proposed partner University of Tirana and the relevant Albanian heritage authority (agreement to be secured).

Case 02

Foodways & equality in the East Baltic

Interpretable AI applied to lipid, proteomic, and isotope evidence to test whether protohistoric foodways were stratified or shared, reading inequality directly from diet.

Proposed partner University of Tartu group, combining published datasets with a defined set of new samples (collaboration to be confirmed).

Case 03

The reflexive laboratory case

An embedded ethnography of a European residue or proteomics laboratory adopting machine-learning tools, observing how AI-derived claims are produced, validated, and contested. Requires no new material access.

Proposed partner A host or partner laboratory, under a standard research-ethics protocol (host to be confirmed).

Why AI is indispensable

The problem cannot be solved by hand

Intractable data

Mass-spectrometry, proteomic, and isotope data are high-dimensional, noisy, and partly degraded, with a single social signal distributed across many features.

Calibrated uncertainty

Only a model can propagate chemical and inferential uncertainty into a social conclusion as an explicit, calibrated probability.

Scale & auditability

Assemblages produce thousands of spectra. An interpretable model makes each social claim traceable to the chemical features that produced it.

Studiable interpretation

Once the step from signal to claim is an auditable model, it becomes an object that researchers and communities can inspect and contest.

AI is therefore constitutive of both the method and the object of study. Without it, the programme could not integrate heterogeneous chemical proxies into social inference at scale, attach calibrated uncertainty to social claims, or externalise interpretation for sociological study.

The AI mass-spectrometry pipeline

From sample to contested social claim, in six stages

Stage 0 — Data generation. GC-MS and high-temperature GC-MS for lipid residues; LC-MS/MS and MALDI for proteins and ZooMS; MC-ICP-MS or TIMS for strontium and lead isotopes; IRMS for carbon and nitrogen. Standardised protocols, procedural blanks, and reference materials throughout. (Lead: G. Elezi)
Stage 1 — Preprocessing. Spectral alignment, baseline correction, normalisation, deconvolution, and peak picking, with quality control and full provenance metadata for every sample.
Stage 2 — Feature representation. A harmonised feature matrix linking each sample to lipid biomarkers, peptide markers, isotope ratios, and its archaeological context.
Stage 3 — Interpretable inference. Transparency-favouring models (regularised linear models, gradient-boosted trees, Gaussian processes) infer social-proxy variables, with feature attribution tracing each inference to specific chemical features.
Stage 4 — Uncertainty quantification. Calibrated uncertainty via Bayesian or conformal methods, with the model able to abstain when evidence is weak, keeping chemical and inferential uncertainty distinct.
Stage 5 — Social inference. Validated proxies map onto EU-legible social variables (migration status, socioeconomic position, region, diet and health) at population level, with stated assumptions. Racial categories are deliberately excluded, consistent with French statistical law and GDPR special-category protections.
Stage 6 — Reflexive loop. Model cards, feature attributions, and uncertainty estimates become the material that researchers, communities, and heritage institutions examine and dispute. The ethnography follows how claims are accepted or resisted and feeds that back into model design as a human-in-the-loop step. (Lead: D. Zipp, with O. Mayorga)

Governance across all stages. A governance layer covers data management, GDPR handling of any identifiable modern data, heritage data governance, and community engagement, so that ethics is built into the method rather than appended to it.

Programme structure

One integrated grant, with a de-risking option

The programme is submitted as a single integrated Starting Grant, because its distinctiveness comes precisely from joining instrument and epistemics. It can, if preferred, divide into two coherent projects.

Project A — Reading social difference from material traces

Co-led by Elezi and Mayorga. Build and validate an interpretable pipeline that infers social variables from mass-spectrometry and isotope data, with characterised uncertainty. A computational sociology and archaeometry project.

Primary panel SH3Secondary SH6Computer science

Project B — The epistemics of AI-mediated evidence

Led by Zipp with Mayorga. An ethnographic and sociology-of-knowledge study of how AI reshapes evidentiary authority and the politics of interpretation in biomolecular archaeology and heritage science. Held in reserve as the fallback.

Panel SH3Low data-generation riskHigh feasibility

Collaborate on the programme

We are building partnerships with universities, laboratories, and heritage authorities across Europe.

Get in touch