Plausible is not proven

A great-grandmother's baptism once took an afternoon in the archive to find. Today a search box returns her name in seconds, a family tree platform suggests her parents before the tea has cooled and a chatbot will explain where the family came from and which notable house it descends from.

Artificial intelligence has changed family history research and much of that change is welcome. Millions of pages that no volunteer would ever have transcribed can now be searched, but speed has a way of looking like certainty. An answer that arrives instantly, neatly formatted and phrased, feels finished but, in genealogy, it rarely is.

In the Portuguese records, abbreviated names, shifting surnames and parishes that share a patron saint are exactly where automated tools can go wrong.

From index to family tree

The large genealogy platforms now use handwriting recognition to index records at a scale no team of volunteers could match. The results are impressive, yet least reliable where Portuguese registers are hardest to read. Parish priests abbreviated freely, which means that Ant.o or M.a are easy for a trained eye but tricky for a machine. A name misread by a single letter can vanish from an index or produce a person who never existed.

An index is merely a finding aid. It says where to look but the evidence is the record itself, read in full on the original image. On another hand, a search that returns nothing does not prove that an event was never recorded, because the book may be missing, the page unindexed or the name simply misread.

Once a machine has read a record, another usually decides who it belongs to. Platforms suggest matches on a simple logic of same name, similar dates and a nearby place. For most of Portuguese history, however, surnames were not fixed. Brothers and sisters could carry different surnames, women didn't adopt their husband's name (in general) and devotional names such as de Jesus or da Conceição could replace a family name altogether. A single parish might have several José Silvas born in the same year.

Emigration, of course, compounds all of this. A man named José Pereira may appear in Massachusetts as Joseph Perry, with a birth year that drifts from one document to the next. An algorithm that links him to the first José Pereira of the right age on the right island has produced an answer, not evidence.

Then, a suggested match accepted with one click enters a family tree, the tree is copied, and the copies become hints for the next researcher. Within a few years the same mistake appears in dozens of trees and its repetition starts to look like proof. Yet twenty trees that share an error usually trace back to one unchecked source, or to none. The only question that matters is which record supports each link.

Fluent, confident and wrong

A large language model does not consult records. It produces the most plausible answer to a question and, in genealogy, plausible and true are not the same thing. Even with web search, it cannot read a parish register that exists only as an image in an archive catalogue, which is still the case for most Portuguese registers.

The early habit of inventing citations outright has become less common as these tools learn to search and quote what they find. The greater danger now is subtler. Chatbots absorb what has been written online about family history, including folklore, and repeat it with authority.

Ask whether Oliveira, Pereira or Carvalho points to Sephardic ancestry and a chatbot may well explain that New Christians took the names of trees, plants and animals to hide their origins. The claim circulates widely but it does not survive contact with the records. These surnames were borne by countless families with no Jewish ancestry, New Christians took surnames as everyone else and where surnames were not fixed no surname alone proves descent of any kind. Jewish ancestry in Portugal is demonstrated through documents, such as Inquisition trials and inquiries into purity of blood, not through the a name. The same is true of the coats of arms so often attached to a surname.

These systems also tend to agree with the person asking. Suggest that a great-grandfather came from a noble family, or from a family of New Christians, and the model will often find a way to support the idea. A careful researcher does the opposite and looks for the evidence that could prove a theory wrong.

Using the tools well

None of this is an argument for avoiding AI altogether. Used with care, it saves real time. It can organise research notes, draft a timeline from records already found, suggest record groups not yet searched and spot impossibilities in a tree, such as a mother aged nine or a woman giving birth after her own burial. It can even offer a first reading of a difficult document, though here another question comes before accuracy.

Archives and platforms set the terms on which their images may be used, which can be restrictive. FamilySearch allows its material to be downloaded for personal, non-commercial use and does not allow it to be posted on other websites without permission, or transmitted or distributed to other sites. Ancestry forbids the use of its content to train any artificial intelligence system. Digitarq, the portal of the archives in the DGLAB network, states the licence that applies to each image. A record image pasted into a chatbot leaves the archive's control and, depending on the provider, may be stored, reviewed or used to train future models. Before uploading anything, read the archive's terms and the AI tool's terms side by side.

Privacy raises the same question from another angle. A tree carries the names and relationships of living relatives, recent civil registration records may concern people protected by the General Data Protection Regulation. Whatever the tool, the measure of a conclusion stays the same. Professional genealogists work to the Genealogical Proof Standard, which asks for reasonably exhaustive research, complete and accurate citations, careful analysis and correlation of the evidence, the resolution of conflicting evidence and a reasoned conclusion. An index is a derivative source, one step removed from the original. A statement generated by a chatbot is not a source at all. A few habits keep each in its place.

1. Go to the image of the original record, every time.

2. Read the whole entry, including parents, grandparents, godparents, witnesses and places of residence.

3. Confirm each identification with at least one independent record.

4. Look for evidence that contradicts the conclusion, and explain it.

5. Treat any surname theory and any match that fits too neatly with suspicion.

The record comes first

Artificial intelligence has made family history faster, and in some ways easier, but it has not changed what counts as proof. A name in an index, a suggested match, a theory about a surname or a fluent paragraph from a chatbot is a lead, and a lead is worth something only once it has been followed to the original source.

For anyone researching Portuguese ancestors, the rule is the same. Use the new tools to find the records faster, and then read the records.