Gains:
- Ability to use artificial intelligence in radiocarbon calibration, seriation and multivariate analysis, and understand that method selection and assumptions are those of the expert.
- Ability to recognize that the radiocarbon result has a range of probability and that reducing uncertainty to a single year is a scientific distortion
- Ability to question sample validity with the principle of 'garbage in, garbage out' and apply the reflex of not confusing correlation and causality
Archeology relies on both laboratory sciences and statistics to answer the "when" question. In this unit, we will see how AI accelerates tasks such as dating (determining the age of a find or layer), radiocarbon (C14) analysis, typology (sorting and periodizing finds according to their formal characteristics) and seriation (the method of arranging contexts chronologically according to the change of find types over time), but why the interpretation of scientific results is always left to the expert and the correct method.
Basic principle: AI accelerates calculation, calibration, statistical patterning and visualization; But which method is appropriate, whether the sample is valid or not, and what the result means require scientific judgment. Putting a wrong statistic into a nice graph does not make it right.
Dating methods and AI
Radiocarbon (C14). Age is calculated from the decay of radioactive carbon in organic remains. The raw result is calibrated with a curve based on tree rings (calibration — correction that converts the raw radiocarbon age to the actual calendar year). AI accelerates calibration calculations and comparative visualization of multiple dates. But AI does not know traps such as the old wood problem (a sample taken from a tree that has lived for hundreds of years is much older than the event in which it was used); Sample criticism is the job of an expert.
Typology and seriation. The frequency of find types varies over time; A type appears, becomes widespread, decreases. Seriation sorts contexts using this "blanket" pattern. AI is powerful at finding these patterns in large data sets and in multivariate analysis (statistics that evaluates many features simultaneously). But the expert can distinguish whether the result is due to a chronological or functional/regional difference.
Other methods. Methods such as dendrochronology (tree ring), thermoluminescence, and archaeomagnetism are also helpful in AI calculation and comparison; Method selection and interpretation are left to the expert.
Tip: When having the AI run a statistic, ask “what method did you use and what are its assumptions?” ask. If the method is wrong (e.g. averaging non-sequential data, ignoring the outlier) the result will be visually "clean" but scientifically incorrect. Do not trust the chart without understanding the method.
Strengths and weaknesses of AI in statistics
Powerful: Repetitive calculation, calibration, pattern scanning in large data sets, visualization code generation, quick testing of different methods.
Weak/risky:
- Don't suggest the wrong method. AI can confidently recommend a test that does not fit the data.
- Sampling blindness. It does not see the small, biased or contaminated sample; "This is the data," he says.
- Extreme comment. It can present a weak correlation as a strong result; correlation is not causation (just because two things change together does not mean that one causes the other).
- Made-up number. If desired, it can produce a "plausible" but unrealistic date or probability.
Caution: A radiocarbon date is never a single "exact year"; is a probability range (e.g. 750-680 BC with 95% confidence). Don't let AI reduce the outcome to a single year; Always carry uncertainty. Hiding uncertainty is a scientific distortion.
three mini cases
Case 1 — Calibration accelerated. One project would perform the calibration and comparative visualization of 48 radiocarbon samples. The AI generated calibration code and comparison graphs in minutes rather than hours; The expert himself assessed the sampling contexts and risk of old wood and marked two samples as "unreliable".
Case 2 — Wrong method caught. One student applied a test suggested by AI to find type frequencies and got a "statistically significant" result. The statistical consultant showed that the data were not suitable for this test (assumptions were violated); With the right method, the result was meaningless. Lesson: the choice of method determines the outcome.
Case 3 — Uncertainty put back. A team would report a single date from the AI summary for a layer, "720 BC." The expert stated that the caliber range was 800-680 BC, and that the single year was misleading. Report corrected with probability range. Bearing uncertainty is honesty.
Four copyable templates
1) Method suitability query:
Your role: archaeological statistics consultant. I will describe the following data and question: [data type, sample size, question]. Tell me about the method options APPROVED with this data, the ASSUMPTIONS of each, and how to check if the yield meets those assumptions. Inappropriate methods should also be addressed on grounds.
2) Radiocarbon interpretation framework:
Below I have my calibrated radiocarbon intervals: [list]. Explain the pitfalls I should be aware of when interpreting these (old wood, contamination, marine reservoir effect, sample-to-event relationship) and how to express the uncertainty interval correctly. Don't give it a single year; maintain intervals.
3) Statistics audit:
I conducted the following analysis: [method and result]. Please criticize: does the method fit the data, do the assumptions hold, is the sample sufficient, are there any outliers/mixtures, is the result over-interpreted? List your weak points; Don't confirm the result, question it.
4) Visualization (honest):
Prepare a chart/code with the following data: [data]. MUST show uncertainty ranges (error bar/range). Set up the axis and scale without making misleading cuts. DO NOT ADD any points that are not in the data. Explain with a note what the graph does and does not show.
Weak prompt / Strong prompt
Weak prompt:
Analyze this find data and state the exact date of the site.
The AI may choose an inappropriate method and make up a single exact date; It hides uncertainty and does not see sampling problems.
Powerful prompt:
Your role: statistical consultant. Yield: [definition, sample size]. First tell me the method and its assumptions suitable for this data, then give the result with the uncertainty range if we apply it. Don't give a single definitive date. Also indicate which risks associated with the sample (small n, mix, old wood) may weaken the result.
The difference: the former produces false certainty; second, the method keeps open assumptions and uncertainty.
Bayesian approach, sampling and "garbage in, garbage out"
A powerful method in modern archaeological dating is Bayesian modeling (a statistical approach that combines existing prior information—e.g., “layer A underlies B”—with measurements to produce narrower, more realistic date ranges) that combines multiple radiocarbon dates with stratigraphic sequences. This approach narrows the wide uncertainty of individual dates by restricting them to the known order of the layers. AI can help build and calculate these models; but what prior information (stratigraphic relationships) goes into the model is an expert decision, and an incorrect prior constraint produces a confident but inaccurate result.
From here we come to the most basic statistical principle of the module: "garbage in, garbage out." Even the most advanced AI and the most elegant statistics cannot produce good results from bad data. A contaminated radiocarbon sample, finds from a mixed context, or a biased sample will remain invalid no matter how much processing is done. AI cannot see these problems because it does not look at the data "from the outside"; It accepts the numbers you give it as real. Only an expert who knows the field can evaluate how the sample was collected, what context it came from, and what it represents.
So before trusting any statistical results, ask these three questions: Is the sample large and representative enough? Are the contexts reliable and unconfused? Is the measurement method appropriate for this question? If you can't answer "yes" to these three questions, it doesn't matter how "significant" the result seems.
Tip: When evaluating the output of a statistic, look first at the input, not the outcome. "Where did this data come from, how, with how many samples?" question eliminates most false conclusions before even looking at the graph.
Analysis task table
Quest
Role of AI
human decision
verification
C14 calibration
Calculation, visualization
Sampling validity
Context, old wood
seriation
pattern scanning
Chronology/function separation
Comparison with stratigraphy
multivariable
statistics calculator
Method selection
assumption check
visualization
Graphics/code
honest representation
uncertainty, scale
Comment
Summary draft
meaning, limit
expert, literature
Common mistakes
- Applying the method without questioning. The wrong test gives a "clean" but wrong result.
- Reducing uncertainty to a single year. Radiocarbon is a range; One year is misleading.
- Bypassing sampling problems. Small/dirty/biased sample invalidates the result.
- Mistaking correlation for causation. Changing together is not cause and effect.
- Mislead with graphics. The dashed axis, hidden uncertainty is scientific distortion.
In summary
AI in dating and statistics; It is a powerful accelerator for calibration, pattern scanning, calculation and visualization. But the choice of method, sample criticism, honest expression of uncertainty, and interpretation of the result belong to scientific judgment. Do not trust the chart without understanding the method; always carry uncertainty; and resist the insistence on a single “exact date.”
Application task
Prepare a sample find-frequency table or a set of calibrated radiocarbon intervals. Have the appropriate analysis and its assumptions identified with the "Method suitability query" template, then have an analysis critiqued with the "Statistical audit" template. Produce a chart sketch showing uncertainty ranges with the "Visualization (honest)" template and avoid single-year discounting.
checklist
- [ ] I have checked the method selection and assumptions.
- [ ] I questioned sample validity (size, mix, old wood).
- [ ] I kept the uncertainty as an interval, I did not reduce it to a single year.
- [ ] I did not confuse correlation with causation.
- [ ] I have shown scale and uncertainty honestly in the chart.