Gains:
- Ability to use artificial intelligence to summarize and compare real sources, eliminating the basis for fabricated attribution with the 'only based on the given source' constraint
- Ability to apply the discipline of verifying every citation in the actual catalog, verifying every abstract back to the source, and catching DOI/imprint hallucinations
- Being able to understand how currentness, multilingualism and gray literature determine the limits of artificial intelligence's knowledge and why original writing is required.
Archeology and cultural heritage is a cumulative science: each new study builds on previous publications spanning centuries. Scanning the literature (the totality of articles, books, reports and theses published on that subject) of a region, a type of find or a period is the first and most tiring step of any scientific study. In this unit, we will see in depth how AI speeds up literature review, but why source verification is paramount in this field and how to combat the most dangerous mistake of AI: hallucination (fabricated sources, citations and information).
Basic principle: AI helps summarize, compare and organize the literature; But no citation can be used without confirmation in a real catalogue, and no summary can turn into a scientific claim without reading the source itself. In the humanities, one false source destroys the credibility of the entire work.
Why is hallucination so critical in this field?
The most common mistake of public chat tools is producing academic resources that do not exist but appear very realistic. The AI can make up a “perfect” looking citation by combining the name of a real author, the name of a real journal, and a plausible title. From where?
- AI produces the “next most likely word”; it doesn't look at an actual database (unless a special tool is connected).
- Because academic citations are patterned, AI is very "good" at creating a citation that fits the pattern but whose content is fabricated.
- In the humanities, the reader often does not immediately check the source; This allows the fake attribution to go live unnoticed.
Conclusion: for an archaeologist, every author-year-title triplet in the AI output should be considered suspicious until proven otherwise.
Tip: Use AI not as a “resource finder” but as an assistant that “processes the real resources you have.” Find the actual articles yourself, give the PDFs to the AI, then say "just rely on these texts I gave you". Thus, the basis for hallucination is largely eliminated.
Steps of literature review
- Scope and question. What are you looking for? Clarify topic, period, region and language boundaries.
- Collection of real resources. Find authentic publications from library catalogues, institutional databases, and trusted directories. It is risky to "find" resources for the AI at this step.
- Summarizing and mapping. Give the actual texts you have collected to AI to summarize and map themes and discussions.
- Comparison and gap analysis. Compare the views of different authors, find contradictions and unexplored gaps.
- Spelling and attribution. Write your own argument with verified citations. AI gives draft; you rewrite with your analysis and actual citations.
Source verification discipline
For each citation, check:
- Is there? Are the author, year, title, journal/publisher located in an actual catalog (library, DOI, index)?
- Does he say what he says? Is the source really saying what the AI claims? Open the source and read the relevant section.
- Is the context correct? Is the quote taken from the author's intent? Is the opposing view attributed to the author?
- Is it up to date and reliable? Is the publication peer-reviewed, current, or does it reflect an outdated view?
Caution: Just because a DOI or inventory number provided by AI appears in the "correct format" does not indicate that that record exists. Search for the number in the real system. A non-existent DOI is one of the most common signs of hallucinations.
three mini cases
Case 1 — Summary saved time. A doctoral student gave 60 real articles (all of which he found and downloaded) to AI and created a theme map; He drafted a table comparing different authors' views on a find type in two days instead of weeks. Each summary has been sourced and verified.
Case 2 — Fake welding disaster averted. A graduate student asked AI to "suggest sources on this topic" and got a nifty list of 15 citations. When he checked it at the library, he found that 9 of the list never existed — real authors, made-up titles. If this list were included in the thesis bibliography, the credibility of the defense would collapse and it could even be considered scientific misconduct.
Case 3 — Incorrect attribution corrected. One team would report that the AI told a writer "he dated that site to the Roman period"; When we turned to the source, it was seen that the author claimed the opposite, that the site was pre-Roman. AI had reversed the view. Reading the source prevented a serious mistake.
Four copyable templates
1) Relying solely on the source given:
Your role: literature summary assistant. ONLY rely on the texts I will paste below. DO NOT ADD any information, sources or references from outside these texts. Write down each of your claims, indicating which part of the text I gave you is based on. If you are asked about something that is not in the text, say "it is not in the sources given."
2) Theme map and comparison:
I will provide summaries of [N] sources below. Group them by themes and create a comparison table: for each theme, which authors say what, where they agree, where they contradict. Use only the views in the texts I have given; adding comments.
3) Citation verification checklist:
When VERIFYing the citation list below, take out the steps I need to check for each (search in the catalog, DOI confirmation, read the relevant page, context check). You DO NOT confirm the accuracy of the attributions; tell me what I should check. Source fabrication.
4) Gap analysis:
Based on the literature summaries I have provided, list questions/gaps that appear to be under-researched on this topic. For each gap, show which source left it open. Rely only on the texts I have given you; Don't speculate.
Weak prompt / Strong prompt
Weak prompt:
Give the 15 most important academic sources on this topic and their summaries.
The AI produces a realistic but likely partially fabricated list without looking at a real database; This is the false attribution trap itself.
Powerful prompt:
Below I am pasting the full citations and summaries of the 12 real articles I found. Based ONLY on these, a theme map and comparison table emerges. ADD new resource. Indicate which article each line comes from. Mark missing topics as "not covered in this resource set."
The difference: the first is fraught with the risk of fabricated sources; the latter handles real resources reliably.
Currentness, language and gray literature
Three features of the archaeological literature make the use of AI particularly risky. The first is topicality: the knowledge of a model is frozen at a certain date, and excavations, new dating and corrections after that date are absent from it. AI can present a superseded view as still valid; That's why you should scan current publications. Second, multilingualism: archeology is written in German, French, Italian, Turkish and many local languages other than English. An important excavation report may have been published in only one language; Relying solely on English-language sources means missing a large part of the literature. AI helps in scanning these foreign language sources, but again, you must have the actual source at hand.
Third, gray literature: the vast majority of archaeological knowledge lies in documents that do not appear in official journals, such as commercial excavation reports, institutional archives and unpublished theses. Gray literature (reports produced in commercial or corporate channels that are not readily available in traditional publication directories) is either absent or scarce in AI training data; Therefore, the AI either does not know this information or makes it up. The real archaeological story of a region is often in this gray literature, and you can only reach it with the real hand from the institutional archives.
Tip: Saying "AI couldn't find anything" or "there are few resources" on a subject does not mean that there is little work on that subject; Often the source is in gray literature, in a foreign language, or after the information section of the model. Ask the real archive about the lack of resources, not the AI.
Literature task table
Quest
Role of AI
human decision
verification
sourcing
Weak/at risk
Selection from real catalog
Catalog confirmation
summarizing
Strong (to given text)
accuracy, choice
Read the source
Comparison
Theme/contradiction map
Comment, weight
source context
Citation writing
format outline
content, accuracy
Is there/does it say
Gap analysis
Candidate questions
Research value
Domain knowledge
Common mistakes
- Making the AI "find" resources. Produces false attribution; Find the resources yourself.
- Relying on the DOI/imprint format. Although the format appears correct, the record may be fabricated; Search in the system.
- Using the summary without reading the source. AI can reverse the author's view.
- Publishing the AI text as is. Originality and risk of plagiarism; Write with your own analysis.
- Not placing the "given resource only" constraint. Without constraints, AI adds externally fabricated information.
In summary
In literature review, AI saves a lot of time in summarizing, comparing and analyzing factual sources. But in a humanities, source verification is non-negotiable: find the sources yourself, verify every citation in the actual catalogue, backcheck every abstract, impose the "cited source only" constraint, and write the final text with your own original analysis. A fake source will destroy your work.
Application task
Choose a topic, find 8-10 real publications in the library/index and collect their citations. Make a comparison chart with the "Based on source only" and "Theme map" templates. In a separate experiment, tell the AI to "suggest a source" (without any restrictions) and check the citations it suggests one by one in the catalog and note how many are real.
checklist
- [ ] I found the sources myself from real catalogues.
- [ ] I put the "only rely on the given source" constraint on the AI.
- [ ] I have confirmed each citation with the catalog/DOI.
- [ ] I have verified each summary by going back to the source.
- [ ] I wrote the final text with my own original analysis and confirmed references.