Reading the Past: From Prehistoric Symbols to Writing and AI

A personal journey through archaeology, ancient civilizations, lost languages, and the search for meaning in 40,000-year-old signs.

AI-Talks.org | November 1, 2026 | 10-minute read

---

History and archaeology have been part of my life since childhood. For more than forty years, I have followed archaeological discoveries and the efforts of scholars to reconstruct civilizations whose languages, beliefs, and ways of life have largely disappeared.

My interest began at home, with my father’s collection of prehistoric artifacts. As a child, I spent hours examining Neolithic polished stone axes, grinding and polishing stones, and other tools fashioned by human hands thousands of years ago. I remember wondering who had made them, what they had been used for, and what the world of their makers had been like. Holding those objects in my hands gave me a tangible connection to a past I could otherwise only imagine.

Another important influence was O Tesouro da Juventude (The Treasure of Youth), an encyclopedic collection covering history, geography, science, literature, and humanity’s greatest achievements, whose volumes I devoured as a child. Its pages opened windows onto distant civilizations, ancient monuments, scientific discoveries, and places I had yet to encounter. It was the kind of reading in which one subject naturally led to another, raising new questions and nurturing a curiosity that, in many ways, has never left me.

Other books soon expanded that world. One of my favorites during adolescence was C. W. Ceram’s Gods, Graves, and Scholars, which introduced me to the extraordinary stories behind some of archaeology’s greatest discoveries. Later, I immersed myself in the richly illustrated Time-Life history collections, exploring ancient cities, vanished empires, and the ingenuity of the people who built them.

Years later, during my doctoral research in Europe, I had the opportunity to visit some of the great museums I had known only through books, including the British Museum in London and the Louvre in Paris. Seeing the original artifacts transformed what had previously been an intellectual interest into something far more immediate.

At the British Museum, the Rosetta Stone was particularly memorable. Its inscriptions, written in Egyptian hieroglyphs, Demotic, and ancient Greek, had provided crucial evidence for Jean-François Champollion’s breakthrough in deciphering hieroglyphs in 1822.

At the Louvre, I encountered two other extraordinary monuments: the Stele of Hammurabi, with its extensive collection of Babylonian legal judgments inscribed in cuneiform, and the Victory Stele of Naram-Sin, commemorating the military triumph of an Akkadian ruler more than four thousand years ago.

Monuments of Decipherment and Early Writing

From left to right, these monuments are arranged from the oldest to the most recent: the Victory Stele of Naram-Sin, the Stele of Hammurabi, and the Rosetta Stone.

Victory Stele of Naram-Sin
Akkadian Empire, c. 2254–2218 BCE
Louvre Museum, Paris
Museum record · Image source

Stele of Hammurabi
Babylon, c. 1792–1750 BCE
Louvre Museum, Paris
Museum record · Image source

Rosetta Stone
Ptolemaic Egypt, 196 BCE
British Museum, London
Museum record · Image source

What impressed me most was the survival of the information these monuments contained. Political authority, laws, military victories, religious beliefs, and fragments of everyday life had been preserved in stone. Although the civilizations that produced them had long disappeared, centuries of archaeological and linguistic scholarship had made much of their historical significance accessible again.

In 1999, I spent three weeks traveling through Greece, visiting archaeological sites and museums that had occupied my imagination for decades. One experience remains especially vivid. At the Archaeological Museum of Heraklion, on the island of Crete, I stood before the Phaistos Disc, a small clay object dating from the second millennium BCE. Both faces were covered with mysterious signs impressed into the clay and arranged in spirals.

Discovered in 1908, the Phaistos Disc remains undeciphered to this day. During the Bronze Age, several writing systems were used in Crete, including Cretan hieroglyphs and Linear A, neither of which has been conclusively deciphered. Another was Linear B, used later to record an early form of Greek and deciphered by Michael Ventris in 1952. The signs impressed on the Phaistos Disc, however, have not been reliably linked to any of these systems. We do not even know which language, if any, they represented, making the interpretation of the object particularly difficult.

The contrast with the monuments I had seen in London and Paris was striking. The Rosetta Stone had helped scholars recover the Egyptian writing system; Hammurabi’s judgments could be read, and Naram-Sin’s inscriptions could be placed within their historical context. The Phaistos Disc, however, remained silent. Someone had carefully impressed those signs into clay nearly four thousand years earlier, apparently intending to record or communicate something. The object had survived, but the knowledge required to interpret it had vanished.

That encounter has stayed with me, particularly as artificial intelligence has begun providing new tools for investigating ancient inscriptions. Machine learning can identify recurring symbols, compare statistical patterns, assist in reconstructing damaged texts, and reveal relationships that might escape human examination. The possibility of applying these techniques to unresolved archaeological mysteries is especially compelling.

Yet even a perfect reconstruction of an inscription cannot guarantee that we will understand it. Recovering the physical signs is one problem; recovering the language and cultural conventions that gave them meaning is another.

These thoughts returned to my mind recently when a cousin, just as curious as I am, sent me a social media post claiming that the world’s oldest writing had been discovered in a cave in Germany and dated back approximately 38,000 years. The accompanying photographs showed an ancient Mesopotamian clay tablet covered with cuneiform signs alongside a prehistoric ivory artifact bearing geometric markings. Could the origins of writing really have been pushed back by more than 30,000 years?

The research behind the headline was genuine, but its findings were considerably more nuanced—and less conclusive—than the sensationalized announcement suggested.

In February 2026, Christian Bentz and Ewa Dutkiewicz published a study in Proceedings of the National Academy of Sciences examining geometric signs engraved on objects approximately 40,000 years old. Using information theory, statistical modeling, and machine-learning classifiers, they reported similarities between the statistical properties of these prehistoric markings and those of the earliest Mesopotamian recording systems.

Their findings raised the possibility that early European hunter-gatherers had developed conventional systems of signs long before the emergence of writing. However, a methodological critique published by Alexander Koplenig in September 2026 challenged important aspects of that interpretation, particularly the classification of the markings, the statistical treatment of very short sequences, and the significance of the reported similarities.

The controversy brings together several subjects that have shaped my intellectual interests: archaeology, the origins of human symbolic thought, mathematics, physics, and artificial intelligence. It also returns me to the question that first emerged so vividly during that afternoon in Crete: how much can we really discover about an unknown system of symbols from the patterns it leaves behind?

The answer depends on distinguishing what mathematics can measure, what archaeology can establish, and what remains beyond our ability to interpret.

---

Forty Thousand Years Before the Scribes

The objects examined by Bentz and Dutkiewicz belong to the Aurignacian culture, associated with early Homo sapiens populations inhabiting Europe approximately 43,000 to 34,000 years ago.

Many were recovered from caves in the Swabian Jura, southwestern Germany, a region famous for some of the oldest surviving examples of figurative art and musical instruments. These communities manufactured sophisticated tools, worked bone and mammoth ivory, produced ornaments, and carved remarkable sculptures of animals and human-animal hybrids.

Some of those objects also carried geometric markings: dots, lines, crosses, notches, and repeated arrangements of simple shapes.

One particularly intriguing artifact is the Adorant, an ivory plaque from Geißenklösterle Cave. One side contains an anthropomorphic figure, while the other displays carefully arranged markings. Other objects, including animal figurines, exhibit similar combinations of incisions and geometric forms.

Archaeologists have debated the purpose of such marks for decades. They might have represented numerical records, hunting tallies, ritual symbols, group identities, decorative conventions, or practices whose meaning is entirely unknown to us.

These interpretations are not necessarily mutually exclusive. An arrangement of marks could have carried ritual significance while also identifying its owner or indicating a particular event. Conversely, its regularity might have resulted primarily from aesthetic preferences or the physical constraints of engraving a small object.

The essential difficulty is that the marks have survived, but the cultural conventions that may have governed their production have not.

We can observe their arrangement and measure their physical characteristics. What we cannot directly observe is how the people who made them interpreted those marks—or whether they expected anyone else to interpret them at all.

This is where quantitative methods become particularly attractive. When the original meaning of an inscription is inaccessible, researchers can still investigate its measurable properties. The question shifts from what the signs meant to how they were organized, and whether that organization differs from patterns generated by simpler processes.

Can We Analyze the Phaistos Disc Mathematically? ⌄ A case study using real inscription data: entropy, repetition, sequential patterns, and the limits of interpretation. Click to expand.

A Text We Still Cannot Read

The Phaistos Disc contains 242 sign occurrences stamped into clay, distributed across its two faces. Researchers have identified 45 distinct sign types, arranged in 61 groups.

To investigate their organization, we can represent each sign type with a number from 01 to 45, following the conventional classification introduced by Arthur Evans.

This allows us to apply statistical methods without knowing what the signs mean.

1. Frequency and Diversity

The first question is straightforward: how many different signs appear on each face?

Side B contains 119 sign occurrences, representing 34 distinct types. Its type-token ratio is:

TTR_B = 34/119 ≈ 0.286

On Side A, one sign is damaged. Depending on how that occurrence is classified, we obtain 35 or 36 distinct types among 123 signs.

TTR_A ≈ 0.285–0.293

The two faces exhibit similar symbol diversity, but this does not tell us what their signs represent.

2. Shannon Entropy

Shannon entropy measures uncertainty about the identity of a sign, based on the relative frequencies of the different sign types.

H = −Σ p_i log₂(p_i)

Calculations based on the numerical transcription give approximately:

H_A ≈ 4.65–4.67 bits/sign
H_B ≈ 4.72 bits/sign

The values are relatively close. This difference alone does not demonstrate that the two faces served different functions or represented a language.

3. Why Sign Order Matters

Shannon entropy considers sign frequencies but not their order. We must also examine recurring sequences and sequential patterns.

A Recurring Sequence on the Disc

On Side B, the following sequence of five signs appears in two different groups:

22 – 29 – 36 – 07 – 08

It occurs in groups B21 and B26. The numbers identify sign types, not their meanings.

This recurrence demonstrates that the inscription contains patterns longer than isolated sign repetitions.

Another example is sign 02, which appears 19 times across both faces and consistently occupies the first position within its group, according to the conventional reading direction.

These patterns can be measured, but we do not know whether they correspond to words, formulas, ritual expressions, or other conventions.

4. Entropy Rate

Entropy rate attempts to capture the predictability introduced by the ordering of signs.

Applying the Lempel–Ziv-based estimator introduced in the previous methodology box:

ĥ = (1/n) Σ log₂(i)/L_i

For a reading of the spirals from the outer edge toward the center, without treating group boundaries as additional symbols, the exploratory estimates are:

ĥ_A ≈ 2.82–2.84
ĥ_B ≈ 3.14

The range for Side A reflects alternative classifications of the damaged sign.

These are exploratory estimates. They depend on the transcription, the estimator, and the assumed reading direction. An inscription this short cannot provide a reliable estimate of the entropy rate of an unknown generating system.

5. Adjacent Repetition Rate

We can also count how frequently two neighboring signs are identical.

Side A contains three pairs of adjacent identical signs, while Side B contains two.

r_A = 3/122 ≈ 0.025
r_B = 2/118 ≈ 0.017

These values are small, but they do not contradict the presence of repeated groups. Immediate repetition and recurrence of longer sequences are different statistical properties.

6. The Statistical Feature Vectors

As in the study of Paleolithic markings, each face can be represented by four statistical features:

x = (H, ĥ, TTR, r)

For Side A, temporarily assigning the damaged occurrence to sign type 03, we obtain:

Side A
x_A ≈ (4.651, 2.816, 0.285, 0.025)

Side B
x_B ≈ (4.718, 3.142, 0.286, 0.017)

These vectors describe measurable statistical properties of the inscriptions. They could be used as input features for machine-learning classifiers.

However, classifying the two faces of a single artifact is not equivalent to identifying their language or function.

7. What We Still Do Not Know

A more rigorous analysis would need to examine alternative reading directions, possible interpretations of the damaged sign, different treatments of group boundaries, and control sequences with comparable lengths and sign frequencies.

It would also be necessary to establish whether the statistical results remain stable under these choices.

The Scientific Conclusion

The Phaistos Disc contains measurable regularities. Mathematics can identify frequencies, repetitions, and dependencies between signs. But none of these measurements reveals what the symbols mean. Without parallel inscriptions, a larger corpus, or independent linguistic evidence, the disc remains undeciphered.

Sources and Methodology
Transcription based on the Evans/Godart sign numbering. Statistical values were calculated specifically for this editorial example; they are not results reported by Bentz and Dutkiewicz for the Phaistos Disc.

Giorgi & Baldacci (2026) — Computational Analysis of the Phaistos Disc
Scuola Normale Superiore — Description of the Sign System
Bentz & Dutkiewicz (2026) — Comparative Methodology
Photographs of both sides: Gleb Simonov, 2024. Wikimedia Commons, CC0 1.0.

---

When Statistics Meets Archaeology

As a physicist, my interest in archaeology has always extended beyond its historical dimension. I was fascinated by how physics and chemistry could help reconstruct the past through the scientific examination of ancient materials. One book I still recommend is Joseph B. Lambert’s Traces of the Past: Unraveling the Secrets of Archaeology Through Chemistry (1997), an engaging introduction to archaeometry.

My own research eventually took a different direction, focusing on color centers in gemstones and minerals. Some of the experimental techniques I encountered, particularly spectroscopy, are also used in archaeological investigations. But that is a story for another time.

In the early 2000s, I could hardly have imagined that artificial intelligence would bring another set of tools to this field. Today, computers can recognize ancient symbols, compare thousands of inscriptions, and investigate their organization through statistical methods. The 2026 study by Bentz and Dutkiewicz offers a particularly interesting example of this convergence.

For someone trained in experimental physics, however, the central question remains familiar: what do the measurements actually demonstrate, and how much can we legitimately infer from them?

Measuring the Structure of Unknown Signs

Bentz and Dutkiewicz examined approximately 3,000 geometric markings distributed across 260 prehistoric artifacts. After preprocessing, 213 objects were retained for their principal quantitative comparisons. The researchers did not attempt to decipher the signs. Instead, they investigated whether their organization resembled that of the earliest Mesopotamian recording systems.

Their approach was straightforward in principle. Even when the meaning of a symbol is unknown, we can measure how frequently it appears, whether it tends to repeat, and how predictable its arrangement is. These properties can then be compared across different collections of signs.

One of the principal tools was Shannon’s information theory, which provides mathematical ways of measuring uncertainty and predictability. The researchers combined these measurements with machine-learning algorithms to compare Paleolithic engravings, early protocuneiform records, and modern writing.

The technical details of their methodology are presented in the expandable box below.

How Did the Researchers Analyze the Signs? ⌄ Information theory, statistical features, a worked example, and machine learning. Click to expand.

1. From Archaeological Marks to Numbers

Imagine a prehistoric object bearing an ordered sequence of engraved signs. Each occurrence is a token, and visually similar marks are classified into types.

s = (s₁, s₂, …, s_n)

Here, n is the total number of signs, m the number of distinct types, and f_i the number of occurrences of type i. Its observed probability is:

p_i = f_i / n

The way marks are identified and classified directly affects all subsequent calculations.

2. Four Statistical Measurements

A. Type–Token Ratio (TTR). The fraction of distinct signs:

TTR = m / n

B. Shannon Entropy (H). Uncertainty about the identity of the next sign, based on its frequency:

H = −∑ p_i log₂(p_i)

C. Entropy Rate (h). Measures uncertainty per sign while taking sequential dependencies into account. Its theoretical definition is:

h = lim_n→∞ H(S₁,…,S_n) / n

For finite sequences, the authors used a Lempel–Ziv-based estimator:

ĥ = (1/n) ∑_i=2ⁿ log₂(i) / L_i

L_i is one plus the length of the longest sequence of signs starting at position i that can already be found in the preceding part of the sequence. Longer matches indicate greater predictability and reduce the corresponding contribution to ĥ.

D. Repetition Rate (r). The fraction of adjacent pairs containing identical signs:

r = r_adj / (n − 1)

Here, r_adj is the number of identical neighboring pairs. There are n − 1 adjacent pairs in a sequence of length n.

A Worked Example: Same Signs, Different Information

Suppose archaeologists discover two hypothetical objects, each containing 24 engraved marks. We represent a notch by A and a dot by B.

The two sequences are:

Object 1
ABABAB | ABABAB | ABABAB | ABABAB

Object 2
AABBBA | BAABBA | BAABAA | BABBAB

Spaces and vertical bars are shown only to make the sequences easier to read. They are not counted as signs.

The first object alternates regularly between A and B. The second contains the same number of each sign, but in a less predictable order.

Step 1 — Count the signs

Both objects contain 12 occurrences of A and 12 of B:

n = 24;   m = 2
p(A) = p(B) = 12/24 = 0.5

Step 2 — Calculate diversity and entropy

Because both have identical frequencies, their TTR and Shannon entropy are identical.

TTR = 2/24 ≈ 0.083

H = −2(0.5 log₂0.5) = 1 bit/sign

Therefore, neither measure detects the difference between the two arrangements.

Step 3 — Calculate the repetition rate

Each sequence has 23 adjacent pairs. Object 1 alternates continuously, so it has no identical neighboring signs. Object 2 contains eight identical neighboring pairs.

r₁ = 0/23 = 0
r₂ = 8/23 ≈ 0.348

Repetition rate therefore detects a difference that ordinary entropy does not.

Step 4 — Estimate the entropy rate

We now examine whether the ordering of the signs creates longer predictable patterns.

Consider position i = 9. For Object 1, the preceding eight signs are ABABABAB. The sequence beginning at position 9 repeats that entire pattern.

The longest previous match has length 8, giving:

L_9,1 = 8 + 1 = 9

For Object 2, the preceding eight signs are AABBBABA. The sequence beginning at position 9 starts with ABB, which appears in the preceding signs. However, the longer continuation does not match.

L_9,2 = 3 + 1 = 4

The contribution at position 9 is consequently different:

Object 1: log₂(9)/9 ≈ 0.352

Object 2: log₂(9)/4 ≈ 0.792

We repeat this calculation for every position from i = 2 to 24, add the contributions, and divide the sum by 24.

ĥ₁ = 14.821/24 ≈ 0.618

ĥ₂ = 21.785/24 ≈ 0.908

Interpretation: Object 1 is highly predictable because its AB pattern keeps reappearing. Object 2 has fewer long matches with its preceding symbols, leading to a higher estimated entropy rate.

Thus, although the sequences have identical Shannon entropy, their sequential predictability differs.

Comparison of the Four Measurements

Feature	Object 1	Object 2
Signs (n)	24	24
Types (m)	2	2
Entropy H	1.000	1.000
Entropy rate ĥ	0.618	0.908
TTR	0.083	0.083
Repetition r	0.000	0.348

From Sequences to Feature Vectors

Machine-learning algorithms do not need to examine the original signs directly. They can receive a four-dimensional vector containing the estimated statistical features, in the order:

x = (H, ĥ, TTR, r)

For the two objects, these vectors are:

Object 1
x₁ = (1.000, 0.618, 0.083, 0.000)

Object 2
x₂ = (1.000, 0.908, 0.083, 0.348)

In this statistical space, the two sequences occupy different positions despite containing exactly the same symbols in the same proportions.

Important: These sequences are invented for teaching purposes. Even 24 signs constitute a small sample for estimating entropy rate; the numerical results should not be treated as reliable estimates of the underlying generating processes. Neither sequence necessarily communicates anything.

3. Machine-Learning Classification

The authors represented each archaeological sequence by its four-dimensional statistical vector. They then used two classification methods:

Multilayer Perceptrons (MLPs): neural networks that learn decision boundaries separating different groups.

k-Nearest Neighbors (kNN): classifiers that assign a group based on nearby observations in the statistical feature space.

Classification accuracy was compared with a majority-class baseline. Failure to outperform that baseline does not prove that the groups have equivalent origins or functions.

4. Comparing Artifact Categories

Statistical models also tested whether properties of the markings differed among types of archaeological objects, such as ivory figurines and tools.

Such differences may indicate systematic engraving practices. They cannot independently establish what the signs communicated.

5. Methodological Limitations

Koplenig’s subsequent critique highlighted two major problems.

Classification uncertainty: decisions about what counts as a sign, including the treatment of transcription annotations, affect every feature.

Short-sequence bias: entropy-based estimates can be unstable when only a few symbols are available.

After removing annotations, Koplenig reported that 143 of 213 Aurignacian sequences (approximately 67%) contained repetitions of only one sign type.

The Scientific Lesson

Statistical features describe the organization of symbols. Machine learning can compare those features across datasets. But identifying a statistical pattern does not establish linguistic meaning, shared function, or cultural continuity.

Sources: Bentz & Dutkiewicz (2026) and Koplenig (2026) . The worked example is original. The entropy-rate calculation uses the finite-sequence estimator described in Bentz and Dutkiewicz.

What the Results Suggested

The researchers reported an intriguing similarity between the Paleolithic engravings and the earliest Mesopotamian records, particularly those from the Uruk V period.

Both collections contained relatively short sequences, with median lengths of approximately eight signs for the Aurignacian material and seven for Uruk V. More importantly, the machine-learning algorithms often struggled to distinguish the two groups using the selected statistical measurements. Modern writing, by contrast, was readily distinguishable from the Paleolithic material.

Another finding concerned the objects themselves. Certain artifacts, particularly ivory figurines, appeared to carry markings with higher estimated information density than ordinary tools. The researchers interpreted this as evidence that the signs were not distributed indiscriminately but were associated with particular categories of objects.

Together, these observations led Bentz and Dutkiewicz to propose that Aurignacian communities had developed conventional systems of signs long before the emergence of writing.

The suggestion is fascinating. It raises the possibility that prehistoric people were already employing shared conventions to organize information in material form, tens of thousands of years before the first Mesopotamian scribes.

But the conclusion is not as secure as it might initially appear. Short sequences can produce deceptively similar statistical measurements, and the results depend partly on how individual marks are identified and classified. Furthermore, a machine-learning algorithm’s inability to distinguish two collections does not establish that they served the same purpose.

The distinction between statistical similarity and shared function is precisely where the scientific controversy begins.

Can We Analyze the Phaistos Disc Mathematically? ⌄ A case study using actual inscription data: entropy, repetition, sequential patterns, and the limits of interpretation. Click to expand.

A Text We Still Cannot Read

The Phaistos Disc contains 242 occurrences of signs stamped into clay, distributed across its two faces. Researchers have identified 45 distinct sign types, arranged into 61 groups.

To investigate their organization, we can represent each sign type with a number from 01 to 45, following the conventional classification introduced by Arthur Evans.

This allows us to apply statistical methods without knowing what the signs mean.

1. Frequency and Diversity

The first question is straightforward: how many different signs appear on each face?

Side B contains 119 sign occurrences, representing 34 distinct types. Its type-token ratio is:

TTR_B = 34/119 ≈ 0.286

On Side A, one sign is damaged. Depending on how that occurrence is classified, we obtain 35 or 36 distinct types among 123 signs.

TTR_A ≈ 0.285–0.293

The two faces exhibit similar symbol diversity, but this does not tell us what their signs represent.

2. Shannon Entropy

Shannon entropy measures uncertainty about the identity of a sign, taking into account the relative frequency of each sign type.

H = −Σ p_i log₂(p_i)

Calculations based on the numerical transcription produce approximately:

H_A ≈ 4.65–4.67 bits/sign
H_B ≈ 4.72 bits/sign

The values are relatively close. This difference alone does not demonstrate that the two faces served different functions or represented a language.

3. Why Sign Order Matters

Shannon entropy considers sign frequencies but not their order. We must also examine recurring sequences and sequential patterns.

A Recurring Sequence on the Disc

On Side B, the following sequence of five signs appears in two different groups:

22 – 29 – 36 – 07 – 08

It occurs in groups B21 and B26. The numbers identify sign types, not their meanings.

This recurrence demonstrates that the inscription contains patterns longer than isolated sign repetitions.

Another example is sign 02, which appears 19 times across both faces and consistently occupies the first position within its group, according to the conventional reading direction.

These patterns can be measured, but we do not know whether they correspond to words, formulas, ritual expressions, or other conventions.

4. Entropy Rate

Entropy rate attempts to capture the predictability introduced by the ordering of signs.

Applying the Lempel–Ziv-based estimator presented in the previous methodology box:

ĥ = (1/n) Σ log₂(i)/L_i

For a reading of the spirals from the outer edge toward the center, without counting group boundaries as additional symbols, we obtain:

ĥ_A ≈ 2.82–2.84
ĥ_B ≈ 3.14

The range for Side A reflects alternative classifications of the damaged sign.

These are exploratory estimates. Their values depend on the transcription, the estimator, and the assumed reading direction.

An inscription this short cannot provide a reliable estimate of the entropy rate of an unknown generating system.

5. Adjacent Repetition Rate

We can also count how frequently two neighboring signs are identical.

Side A contains three pairs of adjacent identical signs, while Side B contains two.

r_A = 3/122 ≈ 0.025
r_B = 2/118 ≈ 0.017

These values are small, but they do not contradict the presence of repeated groups.

Immediate repetition and the recurrence of longer sequences are different statistical properties.

6. The Statistical Feature Vectors

As in the study of Paleolithic markings, each face can be represented by four statistical features:

x = (H, ĥ, TTR, r)

For Side A, provisionally assigning the damaged occurrence to sign type 03, we obtain:

Side A
x_A ≈ (4.651, 2.816, 0.285, 0.025)

Side B
x_B ≈ (4.718, 3.142, 0.286, 0.017)

These vectors describe measurable statistical properties of the inscriptions. They could be used as input features for machine-learning classifiers.

However, classifying two faces of a single artifact is not equivalent to identifying their language or function.

7. What We Still Do Not Know

A more rigorous analysis would need to test alternative reading directions, possible interpretations of the damaged sign, different treatments of group boundaries, and control sequences with comparable lengths and sign frequencies.

It would also be necessary to verify whether the results remain stable under these choices.

The Scientific Conclusion

The Phaistos Disc exhibits measurable regularities. Mathematics can identify frequencies, repetitions, and dependencies between signs. But none of these measurements assigns meaning to the symbols. Without a parallel inscription, a larger corpus, or independent linguistic evidence, the disc remains undeciphered.

Sources and Methodology
Transcription based on the Evans/Godart sign numbering. The statistical values were calculated specifically for this editorial example; they are not results published by Bentz and Dutkiewicz for the Phaistos Disc.

Giorgi & Baldacci (2026) — Computational Analysis of the Phaistos Disc
Scuola Normale Superiore — Description of the Sign System
Bentz & Dutkiewicz (2026) — Comparative Methodology
Photographs of both sides: Gleb Simonov, 2024. Wikimedia Commons, CC0 1.0.

---

From Prehistoric Symbols to the First Cities

To understand the significance of this comparison, we need to move forward tens of thousands of years, from the hunter-gatherer communities of Paleolithic Europe to the agricultural settlements and emerging cities of the ancient Near East.

The Agricultural Revolution began transforming human societies approximately 12,000 years ago. Across different regions, communities gradually adopted cultivation, animal domestication, food storage, and increasingly permanent settlements. These changes extended far beyond food production, fundamentally reorganizing labor, property, social relationships, and eventually political power.

As settlements expanded, agricultural surplus became an increasingly important economic resource. Grain could be stored, exchanged, redistributed, inherited, taxed, or appropriated. Larger communities supported specialized occupations, while temples and other emerging institutions coordinated production, labor, and the distribution of resources.

This growing complexity created a practical problem: information needed to survive beyond individual memory.

How much grain had entered a storehouse? How many animals belonged to a household? How much food had been distributed to workers, and which obligations remained outstanding?

Long before fully developed writing appeared, communities in the ancient Near East employed clay tokens, numerical impressions, seals, and other administrative devices to represent quantities, identify goods, and authenticate transactions.

These practices did not necessarily develop through a single, continuous evolutionary sequence. Their relationships remain subjects of archaeological debate. Nevertheless, they demonstrate how increasingly complex societies developed material techniques for preserving and organizing information.

In southern Mesopotamia, particularly around the rapidly expanding city of Uruk, administrative practices contributed to the emergence of protocuneiform during the fourth millennium BCE.

The earliest surviving tablets were predominantly economic records, documenting commodities, quantities, labor allocations, and institutional transactions. Although the precise origins of writing remain debated, its early development was closely associated with the practical demands of administration.

The earliest scribes were, in an important sense, accountants rather than authors of literature.

Writing eventually developed far beyond these functions, becoming capable of recording laws, religious traditions, scientific knowledge, historical events, and literary works. But its administrative origins reveal something important about the relationship between information and civilization: the ability to preserve records helped institutions coordinate activities across distances, generations, and increasingly complex social structures.

This historical background also clarifies why comparisons between Mesopotamian records and Paleolithic engravings must be approached cautiously.

A hunter-gatherer engraving an ivory figurine and a Mesopotamian administrator recording grain deliveries belonged to fundamentally different societies, separated by tens of thousands of years. We have substantial archaeological evidence for the administrative purpose of early Mesopotamian tablets, whereas the functions of Aurignacian markings remain uncertain.

Both traditions demonstrate the human capacity to produce persistent arrangements of visible signs. Only in the Mesopotamian case, however, can we confidently connect many of those signs to identifiable systems of information recording.

The distinction matters. Similarities in statistical structure may invite comparison, but they cannot establish a shared function, much less a direct historical connection between the two traditions.Writing, however, required another transformation.

Over the following millennia, writing evolved far beyond accounting and administration. It became a means of preserving literature, mythology, and reflections on human existence. One of its most remarkable achievements was the Epic of Gilgamesh, whose surviving cuneiform manuscripts preserve stories composed thousands of years ago.

---

When Symbols Began to Represent Language

The transition from numerical and pictorial records to fully developed writing involved increasingly complex relationships between visible signs and spoken language.

Early protocuneiform was primarily concerned with numbers and objects. Over time, the system acquired a much larger repertoire of signs and increasingly developed the capacity to represent linguistic content.

In Mesopotamia, this process eventually produced cuneiform writing capable of expressing the Sumerian language.

Egypt developed its own early writing tradition at approximately the same broad historical threshold, with evidence dating to around 3250 BCE. Its emergence was closely associated with administration, royal authority, and institutional organization.

Writing later arose independently elsewhere, including China and Mesoamerica.

This distinction between recording information and representing language is essential.

A series of repeated marks may encode a quantity. A conventional geometric symbol may identify a group, object, or ritual concept. Such symbols can communicate information without representing the words or grammatical structure of a spoken language.

Writing in the narrower linguistic sense does something more: it creates an organized correspondence between visible signs and language.

The Paleolithic carvings examined by Bentz and Dutkiewicz do not provide convincing evidence of this correspondence.

Their statistical properties differ sharply from those of established writing systems.

Similar statistical complexity does not establish equivalent meaning or function.

The Aurignacian markings may have been an ancient form of notation, but the researchers did not discover a forgotten alphabet or an unknown language waiting to be translated.

Nor is there evidence that these particular German markings evolved directly into Mesopotamian cuneiform. The chronological gap is enormous, and the archaeological record does not establish such continuity.

What the study reveals is an earlier capacity for conventional information encoding, not an ancestral script connecting prehistoric Europe to Sumer.

The evolution of writing was neither linear nor confined to a single civilization. Across different regions of the world, societies developed distinct ways to represent, preserve, and transmit information, sometimes independently. The following interactive timeline traces major milestones in this long history, from prehistoric symbolic expression and the earliest writing systems to modern artificial intelligence and emerging forms of communication between AI agents.

From Stone Symbols to Neuralese 40,000 Years of Written Communication
Explore the timeline · 29 milestones ↗

---

What Artificial Intelligence Can—and Cannot—Read

Artificial intelligence is opening new possibilities for investigating ancient civilizations. Machine-learning techniques can recognize recurring symbols, compare inscriptions from different archaeological sites, assist in reconstructing damaged texts, and identify statistical relationships that might escape human observation.

Some applications have already produced significant results. In 2023, Gutherz and colleagues demonstrated that neural machine translation could translate Akkadian texts into English, working either from digital representations of cuneiform signs or from their transliterations. Their research illustrated how computational methods might help scholars process the enormous collections of Mesopotamian tablets that remain untranslated.

However, there is an essential distinction between translating an ancient language and deciphering an unknown system of signs.

Akkadian has been studied for generations. Scholars understand much of its vocabulary and grammar, and thousands of existing translations provide the examples needed to train and evaluate computational models. The researchers were not asking their algorithms to discover an unknown language. They were teaching them to translate one whose linguistic structure was already substantially understood.

The situation is entirely different for Linear A, the mysterious inscription on the Phaistos Disc, and especially the Paleolithic engravings examined by Bentz and Dutkiewicz. In these cases, we lack reliable translations, bilingual inscriptions, and sufficient linguistic context. With the Paleolithic material, we do not even know whether the markings represent language, numerical information, ritual conventions, or something unrelated to communication.

The 2026 study therefore did not attempt decipherment. Its machine-learning classifiers examined measurable characteristics of sign sequences and investigated whether these properties could distinguish Paleolithic markings from later recording systems.

This approach has a conceptual connection to modern language models. Large language models learn statistical dependencies among tokens, allowing them to predict their occurrence in different contexts. Bentz and Dutkiewicz used much simpler statistical measurements and classification algorithms, but both approaches exploit the mathematical regularities that can emerge in sequences.

Yet recognizing regularity is not the same as establishing meaning. A classifier can identify similarities between two sets of symbols without determining whether they performed comparable functions.

The methodological criticism published by Alexander Koplenig in September 2026 makes this limitation particularly concrete. He questioned how individual marks had been classified, whether transcription annotations had inadvertently entered the statistical calculations, and whether extremely short sequences could support reliable estimates of entropy and related properties.

In his reanalysis, after removing annotations, he demonstrated how short sequences produced by a simple random process could exhibit statistical properties that complicated comparisons with the archaeological material.

These findings challenged the strength of the original conclusions. The reported overlap between Paleolithic markings and early protocuneiform might partly reflect limitations of the data and methodology rather than equivalent underlying information systems.

This does not establish that the prehistoric marks were meaningless. It demonstrates that the available statistical evidence cannot securely identify their function.

An interesting comparison comes from a different branch of artificial intelligence: Neuralese, the name sometimes given to communication protocols developed by artificial agents learning to cooperate.

In experiments involving multiple neural-network agents, researchers have observed the emergence of signaling systems that facilitate coordination. Some display recognizable regularities and elementary compositional properties, even though they were not explicitly programmed to reproduce human language.

The comparison is revealing because researchers studying artificial agents can conduct experiments that archaeologists cannot. They can modify messages, observe how receiving agents behave, manipulate the tasks, and measure whether communication improves performance. In this way, they can investigate the functional significance of an otherwise unfamiliar signaling system.

With Paleolithic artifacts, the original senders and receivers disappeared tens of thousands of years ago. Their intentions, social practices, and possible communication rules cannot be experimentally recovered. We possess the physical traces, but not the behavioral context that would allow us to determine how those traces were used.

AI-Talks.org · Interactive Research Note

Can AI Understand an Unknown Language? Three cases, three different kinds of evidence. Click to explore.

+

Artificial intelligence can translate ancient languages, analyze undeciphered inscriptions, and investigate communication between machines. Yet these achievements involve fundamentally different scientific problems.

Case 01

Akkadian

Translating a Known Language

Akkadian was spoken in ancient Mesopotamia and written using cuneiform signs. Although it ceased to be spoken long ago, scholars have reconstructed much of its vocabulary and grammar.

Machine-translation systems can learn from existing transcriptions and translations, producing new translations that specialists can evaluate against established linguistic knowledge.

What makes this possible? Dictionaries, grammatical knowledge, parallel translations, and substantial textual evidence.

Akkadian house-sale contract, c. 2350–2150 BCE. The Metropolitan Museum of Art, no. 86.11.204. Museum record · Image source · CC0.

Case 02

The Phaistos Disc

An Undeciphered Sign System

This Minoan clay disc, dating from the second millennium BCE, bears stamped signs arranged in spirals. Its inscription remains undeciphered, and the language or system behind it has not been securely identified.

AI can catalogue the signs, examine their distribution, and compare competing hypotheses. However, the limited evidence provides no independently verified way to determine whether a proposed reading is correct.

What is missing? A secure linguistic identification, sufficiently extensive comparable texts, and independent evidence connecting signs to meanings.

Phaistos Disc, Side A. Heraklion Archaeological Museum, Crete. Photograph: Gleb Simonov (2024). Image source · CC0.

Agent A Sender

● ◆ ●

→

Message

Agent B Receiver

Task outcome → Behavioral test
Does the message change what the receiver does?

Case 03

Neuralese

Investigating Emergent Communication

Artificial agents trained to cooperate can develop signaling protocols that were not explicitly designed by their programmers.

Unlike archaeologists studying lost civilizations, AI researchers can intervene in the system. They can change a message, observe the receiver’s behavior, and test whether particular signals carry functional information about a task.

What makes testing possible? Known tasks, observable behavior, controlled interventions, and measurable performance outcomes.

Original conceptual schematic, AI-Talks.org (2026). Illustrative representation of experimental communication between artificial agents; not an actual neural-network visualization.

The Same Mathematical Tools, Different Levels of Evidence

In Akkadian, linguistic knowledge makes translation possible and its quality testable. With the Phaistos Disc, statistical patterns can be examined, but proposed meanings lack independent verification. In Neuralese experiments, researchers can directly test whether signals affect behavior. Recognizing structure is not equivalent to recovering meaning.

Scientific background: Gutherz et al. (2023) — Neural translation of Akkadian; Andreas et al. (2017) — Translating Neuralese; Mordatch & Abbeel (2018) — Emergent communication in multi-agent systems.

Image licensing: Historical photographs: CC0 1.0 Public Domain Dedication . The Neuralese diagram is an original editorial illustration.

This difference defines both the promise and the limitations of artificial intelligence in archaeology.

I find myself returning to the Phaistos Disc, which impressed me so deeply during my visit to Crete in 1999. Today, advanced imaging and computational techniques can document its signs in extraordinary detail, examine their arrangement, and compare them with other ancient inscriptions. These methods can improve our evidence and help eliminate unsupported hypotheses.

But the disc remains undeciphered. More precise measurements do not automatically provide the missing linguistic and cultural context.

The real contribution of artificial intelligence is therefore methodological. It expands what archaeologists can observe, quantify, compare, and test. Its conclusions become scientifically meaningful when they are supported by independent evidence, appropriate statistical controls, and interpretations that remain open to criticism.

Artificial intelligence can reveal structure in the archaeological record. Establishing what that structure meant remains a much harder historical problem.

---

The Long History of External Memory

The origins of writing are often presented as a relatively straightforward historical progression. First came speech, followed by agriculture, permanent settlements, cities, and eventually written language.

That sequence captures important developments, but it conceals a much older and more complex history.

Long before the first urban civilizations, humans were already creating objects capable of preserving visible traces of their activities, beliefs, and experiences. Cave paintings, carved figurines, ornaments, and geometric engravings demonstrate that symbolic expression was an important part of human life tens of thousands of years before the earliest writing systems.

Before Writing: The Art of Prehistoric Humanity

Three remarkable traditions of prehistoric art, presented from older to more recent representative periods.

Archaeological context

The images are arranged chronologically according to representative artistic periods, not the earliest occupation or oldest paintings at each site. Dating prehistoric art is difficult: pigments may lack datable organic material, different figures in the same shelter may belong to different periods, and indirect dating can yield broad chronological ranges. These images demonstrate a long history of visual representation, but they do not establish that prehistoric paintings constituted written language.

References and dating:
French Ministry of Culture — Chronology of Lascaux
Museo Nacional de Altamira — Chronology
UNESCO — Serra da Capivara National Park
ICOMOS — Serra da Capivara, World Heritage evaluation

These objects introduced a remarkable possibility: something produced by one person could remain available to others long after its maker had departed.

A carved pattern could be reproduced. A recognizable image could evoke a shared tradition. A sequence of marks might have served as a reminder, a record, or a means of communication. Such possibilities do not establish that every engraving encoded information, but they illustrate the potential of material objects to extend human memory beyond individual experience.

The Aurignacian artifacts from the Swabian Jura belong to this much longer history.

Bentz and Dutkiewicz proposed that their geometric markings reflected deliberate, conventional practices, supported by quantitative regularities in the archaeological material. Koplenig’s subsequent criticism demonstrated why the available evidence does not securely establish that interpretation.

The debate remains unresolved. We cannot confidently reconstruct the rules governing these marks, determine whether they communicated specific information, or establish a developmental relationship with the writing systems that appeared many millennia later.

Nevertheless, the surviving artifacts provide evidence of intentional engraving practices whose organization can be investigated mathematically. Their significance does not depend on proving that prehistoric Europeans had already invented writing.

The emergence of writing in Mesopotamia and Egypt during the late fourth millennium BCE remains a major intellectual and social transformation. Writing eventually made it possible to preserve increasingly elaborate administrative records, legal traditions, religious texts, scientific observations, and literature. Knowledge could accumulate across generations without depending exclusively on oral transmission or individual memory.

The consequences extended far beyond communication. Written records helped institutions coordinate economic activities, preserve obligations, establish authority, and maintain continuity across time and distance.

In this sense, writing became one of humanity’s most consequential technologies of external memory.

Over subsequent millennia, paper, printing, libraries, and digital storage expanded this capacity enormously. Today, artificial intelligence introduces another development: external systems that can not only store and retrieve information, but also transform it, establish connections among different sources, and assist in reasoning about their contents.

The distinction is important. A clay tablet preserves a record; a book organizes and reproduces knowledge; a computational model can analyze information and generate new representations from it. These technologies are not stages of a single inevitable progression, but they reflect humanity’s continuing efforts to extend its cognitive capabilities through material and technological means.

There is something particularly striking about using artificial intelligence to investigate engravings made approximately 40,000 years ago. We are applying one of our most recent technologies to some of the oldest surviving expressions of human symbolic activity.

The marks on the Adorant remain unreadable. Their purpose may never be established with certainty. Statistical analysis can describe their organization, but it cannot independently recover the intentions of the people who engraved them.

That uncertainty does not diminish their archaeological importance. It reminds us that the history of human cognition cannot be reconstructed from technological achievements alone. It also depends on the surviving traces of ordinary practices whose meanings have disappeared.

The earliest writing traditions that have survived emerged in the late fourth millennium BCE. Yet tens of thousands of years before that, humans were already engraving, painting, and shaping objects in ways that carried significance within their communities, even when that significance is now beyond our reach.

As I finish this article, I find myself returning to the experiences that first drew me toward archaeology: the polished stone tools in my father’s collection, the books I read when I was young, and, much later, standing before the Phaistos Disc in Crete. My training as a physicist has given me another way of approaching such objects—through measurement, comparison, and experiment—but it has never replaced the fascination that came first.

Perhaps that is also why writing about these subjects matters to me. I write partly to understand them, to organize what I have learned, and to give some permanence to ideas that would otherwise remain scattered in my memory.

The Adorant still resists interpretation. So does the Phaistos Disc. And perhaps that resistance is part of what keeps them alive intellectually. They remind me that the past is not a book waiting patiently to be read. Sometimes we recover a language, sometimes only a pattern, and sometimes nothing more than the certainty that another human being once made something deliberately.

For me, that is already enough to make the distance of thousands of years feel unexpectedly small.

---

References

Archaeology, Prehistoric Symbols, and the Origins of Writing

• Bentz, Christian, and Ewa Dutkiewicz. 2026. “Humans 40,000 y Ago Developed a System of Conventional Signs.” Proceedings of the National Academy of Sciences 123 (9): e2520385123. Published online February 23, 2026.
• Koplenig, Alexander. 2026. “Coding Choices and Short-Sequence Artifacts Undermine Claims of a Conventional Paleolithic Sign System.” Proceedings of the National Academy of Sciences 123 (38): e2627836123.
• Kelley, Kathryn, Mattia Cartolano, and Silvia Ferrara. 2025. “Seals and Signs: Tracing the Origins of Writing in Ancient South-West Asia.” Antiquity 99 (403): 64–82. First published online November 5, 2024.
• Spar, Ira. 2004. “The Origins of Writing.” Heilbrunn Timeline of Art History. New York: The Metropolitan Museum of Art.
• UNESCO World Heritage Centre. n.d. “Göbekli Tepe.” UNESCO World Heritage List.

Information Theory, Artificial Intelligence, and Machine Communication

• Shannon, Claude E. 1948. “A Mathematical Theory of Communication.” Bell System Technical Journal 27 (3): 379–423; 27 (4): 623–656.
• Gutherz, Gai, Shai Gordin, Luis Sáenz, Omer Levy, and Jonathan Berant. 2023. “Translating Akkadian to English with Neural Machine Translation.” PNAS Nexus 2 (5): pgad096.
• Andreas, Jacob, Anca Dragan, and Dan Klein. 2017. “Translating Neuralese.” In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 232–242. Vancouver: Association for Computational Linguistics.
• Mordatch, Igor, and Pieter Abbeel. 2018. “Emergence of Grounded Compositional Language in Multi-Agent Populations.” Proceedings of the AAAI Conference on Artificial Intelligence 32 (1): 1495–1502.

Historical and Personal Background

• Mee, Arthur (original conception). O Tesouro da Juventude (The Treasure of Youth). Brazilian Portuguese adaptation of The Children’s Encyclopaedia / The Book of Knowledge. 18 volumes. W. M. Jackson, Inc., Brazil, various editions from the 1920s onward. Collection overview and historical material.
• Ceram, C. W. 1951. Gods, Graves, and Scholars: The Story of Archaeology. Translated by E. B. Garside. New York: Alfred A. Knopf. Originally published in German as Götter, Gräber und Gelehrte: Roman der Archäologie (Hamburg: Rowohlt, 1949).
• Time-Life Books. 1990s. História em Revista (History in Review). Brazilian edition, 25 illustrated volumes. São Paulo: Abril Livros / Time-Life Books.
• Lambert, Joseph B. 1997. Traces of the Past: Unraveling the Secrets of Archaeology Through Chemistry. Reading, MA: Addison-Wesley / Helix Books.

These references hold particular personal significance. During my childhood, I devoured the volumes of O Tesouro da Juventude (The Treasure of Youth), an encyclopedic collection that introduced me to history, geography, science, literature, and the achievements of human civilization. C. W. Ceram’s Gods, Graves, and Scholars and the richly illustrated Time-Life history collections subsequently deepened my interest in archaeology and the ancient world. Their accounts of remarkable discoveries, accompanied by photographs, maps, and reconstructions, introduced me to civilizations, monuments, and artifacts that I would encounter in person many years later.

During my scientific training, Joseph B. Lambert’s Traces of the Past: Unraveling the Secrets of Archaeology Through Chemistry (1997) added a new dimension to that lifelong interest, showing how chemistry, physics, and modern analytical techniques could uncover information about ancient materials and civilizations. It also helped me recognize the connections between archaeology and my own work as a physicist, particularly my research interests in spectroscopy and the physical properties of minerals.

---

Further Reading on AI-Talks.org

Human Evolution, Archaeology, and Ancient Civilizations

• Pinheiro, Maurício Veloso Brant. 2026. “We Are the Neanderthals: AI, Homo Artificialis, and the Quiet Replacement of Human Intelligence.” AI-Talks.org, August 29, 2026. Examines human cognitive evolution, the interaction between Homo sapiens and Neanderthals, and the possibility that artificial intelligence may represent another major transformation in the evolution and ecology of intelligence.
• Pinheiro, Maurício Veloso Brant. 2026. “What the Late Bronze Age Reveals About Modern Fragility.” Toward a Science of Collapse — Part II. AI-Talks.org, May 12, 2026. Examines how drought, warfare, migration, trade disruption, and institutional fragility interacted during the systemic collapse of the highly interconnected Late Bronze Age world, and considers parallels with the vulnerability of modern networks.

Neuralese, Emergent Communication, and Autonomous AI Agents

• Pinheiro, Maurício Veloso Brant. 2026. “When AI Agents Built a Society: Emergence World and AI Safety.” AI-Talks.org, May 25, 2026. Analyzes the Emergence World experiment, in which autonomous AI agents interact within persistent virtual societies, examining cooperation, collective behavior, institutions, instability, and safety in multi-agent systems.
• Pinheiro, Maurício Veloso Brant. 2026. “What Is Neuralese? AI’s Latent Language Explained.” AI-Talks.org, June 28, 2026. Explores machine-to-machine communication, latent representations, internal reasoning, and the difficulty of interpreting artificial intelligence systems whose internal processes need not take the form of human language.

---

Image Credits and Licensing

The photographs reproduced in this article and its accompanying galleries illustrate archaeological sites, artifacts, and ancient writing systems discussed in the text. Each entry identifies the subject, approximate date, holding institution or archaeological location, photographer or image provider, and applicable reuse license. Dates refer to the archaeological works, not the photographs.

Prehistoric Cave and Rock Art

• Lascaux Cave Paintings — France. Upper Paleolithic, approximately 21,500–21,000 years ago. Lascaux is celebrated for its monumental paintings of aurochs, horses, deer, and other animals. The photograph reproduced in this article shows a modern reconstruction at Lascaux II rather than the original cave, which is protected for conservation. Photograph by Eline13Viki+ (2024). Wikimedia Commons — Lascaux Paintings. License: CC0 1.0 — Public Domain Dedication.
• Polychrome Bison — Altamira Cave, Spain. Upper Paleolithic, Magdalenian period, approximately 19,000–17,000 years ago. A celebrated example of prehistoric animal painting, combining natural pigments and the contours of the cave ceiling to suggest movement and volume. The Altamira complex also contains considerably older artworks. Photograph credited to Museo de Altamira and D. Rodríguez. Wikimedia Commons — Polychrome Bison. License: CC BY-SA 3.0.
• Prehistoric Rock Paintings — Serra da Capivara, Brazil. Prehistoric period; approximately 12,000–6,000 years ago is used here as a representative regional chronological range, not as a direct date for the photographed panel. The sandstone shelters of Serra da Capivara National Park, Piauí, preserve paintings depicting animals, human figures, and scenes of collective activity. Some archaeological interpretations propose substantially older phases of occupation and artistic activity. Photograph by Diego Rego Monteiro (2012). Wikimedia Commons — Serra da Capivara. License: CC BY-SA 4.0.

Prehistoric Artifacts and Early Writing

• The Adorant — Geißenklösterle Cave, Germany. Aurignacian period, approximately 40,000 years ago. A small mammoth-ivory plaque bearing an anthropomorphic figure, notches, and rows of geometric markings. Discovered in the Swabian Jura, southwestern Germany; Landesmuseum Württemberg, Stuttgart. Photograph by Hendrik Zwietasch / Landesmuseum Württemberg. Wikimedia Commons — The Adorant. License: CC BY 4.0.
• Proto-Cuneiform Administrative Tablet — Mesopotamia. Approximately 3100–2900 BCE, Jemdet Nasr period. An early clay accounting record, probably originating in Uruk, documenting the distribution of barley or other grain. Its numerical and pictographic signs illustrate the administrative origins of Mesopotamian writing. The Metropolitan Museum of Art, New York, accession no. 1988.433.1. Museum collection record · Wikimedia Commons — Administrative Tablet. Image courtesy of The Metropolitan Museum of Art, Open Access collection. License: CC0 1.0.
• Akkadian Cuneiform House-Sale Contract — Mesopotamia. Approximately 2350–2150 BCE. A clay tablet from the Quradum archive recording a property transaction, illustrating the use of established writing systems in legal and economic affairs. The Metropolitan Museum of Art, New York, accession no. 86.11.204. Museum collection record · Wikimedia Commons — Akkadian Contract. Image courtesy of The Metropolitan Museum of Art, Open Access collection. License: CC0 1.0.

Monuments of Ancient Civilizations

• Victory Stele of Naram-Sin — Akkadian Empire. Reign of Naram-Sin, approximately 2254–2218 BCE. A monumental relief commemorating the king’s victory over the Lullubi, combining royal imagery with inscriptions preserving historical and political information. Musée du Louvre, Paris, accession no. SB 4. Photograph by Ur-Pabilsag (2019). Louvre collection record · Wikimedia Commons — Victory Stele of Naram-Sin. License: CC BY-SA 4.0.
• Stele of Hammurabi — Babylon. Reign of Hammurabi, approximately 1792–1750 BCE. A basalt monument inscribed with an extensive collection of legal judgments in Old Babylonian cuneiform, surmounted by a relief depicting Hammurabi before the god Shamash. Musée du Louvre, Paris, accession no. SB 8. Photograph by Urko Dorronsoro (2011). Louvre collection record · Wikimedia Commons — Code of Hammurabi. License: CC BY-SA 2.0.
• The Phaistos Disc — Side A, Minoan Crete. Second millennium BCE, commonly dated to approximately 1700–1650 BCE. A fired-clay disc bearing stamped signs arranged in a spiral, whose inscription remains undeciphered. Heraklion Archaeological Museum, Crete, Greece. Photograph by Gleb Simonov (2024). Wikimedia Commons — Phaistos Disc, Side A. License: CC0 1.0.
• Statuette of a Scribe — Ancient Egypt. New Kingdom, Dynasty 18, approximately 1427–1347 BCE. A serpentinite statuette depicting a seated Egyptian scribe with a papyrus scroll across his lap, illustrating the specialized role of professional recordkeepers in ancient Egyptian society. From Thebes, Upper Egypt. The Metropolitan Museum of Art, New York, accession no. 31.4.1a. Museum collection record · Wikimedia Commons — Statuette of a Scribe. Image courtesy of The Metropolitan Museum of Art, Open Access collection. License: CC0 1.0.
• The Epic of Gilgamesh — Tablet XI. Neo-Assyrian period, first millennium BCE; the manuscript tradition represented is conventionally dated to the seventh century BCE. A modern plaster cast reproducing a cuneiform tablet containing the Mesopotamian Flood narrative. Harvard Semitic Museum, Cambridge, Massachusetts. Photograph by Daderot (2016). Wikimedia Commons — Epic of Gilgamesh, Tablet XI. License: CC0 1.0.
• The Rosetta Stone — Ptolemaic Egypt. 196 BCE. A granodiorite decree inscribed in Egyptian hieroglyphs, Demotic, and ancient Greek. Its multilingual text provided essential evidence for the nineteenth-century decipherment of Egyptian hieroglyphs. British Museum, London, accession no. EA24. Photograph by Andres Rueda (2012). British Museum collection record · Wikimedia Commons — Rosetta Stone. License: CC BY 2.0.

Original Illustrations and Diagrams

The article’s featured illustration, depicting archaeological objects, ancient inscriptions, scientific instruments, and artificial intelligence, was generated with AI under the author’s editorial direction for AI-Talks.org (2026). It is a conceptual illustration, not a documentary photograph or an exact reconstruction of specific archaeological objects.

The original Neuralese and information-theory diagrams appearing in the explanatory boxes are schematic visualizations prepared specifically for this article. They are intended to illustrate scientific concepts rather than reproduce historical artifacts or experimental observations.

Licensing and Editorial Notes

Photographs distributed under CC BY and CC BY-SA require attribution, identification of the license, and disclosure of modifications when applicable. ShareAlike provisions apply to adaptations of the licensed photographs. Images distributed under CC0 may be reused without a legal attribution requirement, although identifying the photographer or holding institution remains good scholarly practice.

Chronological ranges are approximate. In particular, the Serra da Capivara estimate refers to a regional archaeological context and should not be interpreted as a direct date for the illustrated painting. Similarly, the Lascaux photograph depicts a modern reconstruction rather than an original cave wall.

All archaeological photographs are credited to their respective photographers, institutions, and Wikimedia Commons sources. AI-generated editorial illustrations are identified separately to distinguish conceptual imagery from archaeological documentation.

---

Maurício Veloso Brant Pinheiro, PhD
Professor of Physics, Federal University of Minas Gerais (UFMG)
Founder, Author and Editor, AI-Talks.org
About Us

Editorial transparency note: This article, as with all articles published on this site, was conceived, directed, written, and reviewed by Prof. Maurício Veloso Brant Pinheiro. Artificial intelligence was used as an assistant for editorial refinement, formatting, image generation, SEO metadata, and publication workflow.

---

Copyright 2026 AI-Talks.org