AI in Archaeology: The False-Positive Problem Gets a Named Expert Voice

AI in Archaeology: The False-Positive Problem Gets a Named Expert Voice

The recurring counterweight to AI discovery hype — hallucinated sites and missed real ones — now has a quotable authority in current coverage. Luca Sanna, professor of Digital Archaeology at the University of Cagliari, in a piece syndicated Oct 4, 2026: "In archaeology, AI fails on context: it lacks the intuition, experience and ability to interpret sites, and it also suffers from training biases.¹" His concrete example of a false positive: "Working on fragmentary data often generates visual hallucinations, because the software, trained to look for regular shapes (circles, squares, rectangles), tends to force the interpretation: it then happens that you 'see' a proto-historic nuraghe in a circular enclosure built fifty years ago with dry stones by a shepherd or even in a road roundabout covered by vegetation."

Sanna also names the opposite failure — false negatives: "The algorithm can also err by default... The archaeological structure is there, but the AI erases it or does not detect it." His bottom line matches the emerging consensus framing: "This is a true digital revolution that does not replace the researcher, but supports and speeds up his work."

Provenance caveat: the outlet (Evidence Network, Canada) is a syndicator and the piece reads as translated from Italian; the bylined author's bio (criminal-justice research analyst) does not match the archaeology beat. The signal here is the named interviewee and his institutional affiliation, not the outlet — treat the quotes as attributed but verify against Sanna's own publications before leaning on them hard.

What it means

• Every credible write-up of AI archaeology now carries the same two-part caveat (hallucination + lack of context). This is the technical twin of the public-discourse "AI-slop policing" trend tracked in AI in Archaeology: When a Real Tomb Discovery Gets Pulled into the AI-Slop Wars.
• Verification is the bottleneck and the business gap. The Nazca pipeline's discipline — AI proposes, humans walk every candidate (AI-Accelerated Site Discovery from Imagery: Yamagata + IBM's Nazca Pipeline) — is the model; anyone selling AI site detection without a ground-truthing layer will accumulate the kind of errors Sanna describes. Quality assurance / expert-review services are the lucrative layer.