How easily a language model can predict what comes next from the text before it. Higher numbers mean less predictable choices.
Measured as surprisal (bits)
Read the two passages
LLM · source-conditioned paraphrase
Can any knowledge be so certain that a reasonable person could not doubt it? Although this question seems simple, examining it carefully reveals how difficult a confident answer can be. Philosophy addresses such questions by investigating the ambiguities and confusion in ordinary ideas rather than accepting the careless or dogmatic answers we often give in everyday life and even in science. Many beliefs that we initially treat as certain reveal apparent contradictions on closer inspection, so we must think carefully about what we can reasonably believe. Our present experiences provide a natural starting point for this search, since they appear to offer knowledge, even though the precise sense in which they do so needs clarification.
Human · Bertrand Russell, 1912
Is there any knowledge in the world which is so certain that no reasonable man could doubt it? This question, which at first sight might not seem difficult, is really one of the most difficult that can be asked. When we have realized the obstacles in the way of a straightforward and confident answer, we shall be well launched on the study of philosophy—for philosophy is merely the attempt to answer such ultimate questions, not carelessly and dogmatically, as we do in ordinary life and even in the sciences, but critically, after exploring all that makes such questions puzzling, and after realizing all the vagueness and confusion that underlie our ordinary ideas. In daily life, we assume as certain many things which, on a closer scrutiny, are found to be so full of apparent contradictions that only a great amount of thought enables us to know what it is that we really may believe. In the search for certainty, it is natural to begin with our present experiences, and in some sense, no doubt, knowledge is to be derived from them.
Conditional surprisal. How unexpected an observed token is to a fixed reference model given the text before it: −log₂ P(token | preceding text). Lower values mean more expected choices. A token is a word or part of a word; punctuation is included. Bars show the percentage of tokens in each two-bit interval, with 14+ collecting the upper tail.
Reference model.GPT-2 · pinned model revision ↗. Both complete passages are scored with the same model and tokenizer, starting with one end-of-text token. No sentence resets or sampling. AI rewrite: 127 tokens; human original: 208 tokens. These are reference-model measurements, not a description of Glyph’s architecture.
Reading the chart. Further right means less expected to this model, not more human or better. These short passages differ in period and style; the public-domain original may also occur in the model’s training data. This comparison is not an authorship test or a benchmark of Glyph.
Exact surprisal measurements
Bits per token
AI rewrite % (count)
Human original % (count)
0–2
22.05 (28)
25.00 (52)
2–4
18.90 (24)
22.12 (46)
4–6
14.96 (19)
20.19 (42)
6–8
17.32 (22)
13.46 (28)
8–10
13.39 (17)
8.65 (18)
10–12
6.30 (8)
5.77 (12)
12–14
4.72 (6)
2.88 (6)
14+
2.36 (3)
1.92 (4)
Mean surprisal: AI rewrite 5.55 bits/token; human original 4.83 bits/token.
AI rewrite: 115 words, 5.38 letters per word. Human original: 182 words, 4.61 letters per word. Both passages contain five sentences. The slider interpolates the measured profiles; intermediate positions do not correspond to newly generated text.
Six lexical subfeatures of Biber’s framework, plus mean word length (F44). Rates below are occurrences per 100 words; parentheses show raw counts. Features overlap and do not sum to 100%.
Feature
LLM
Russell
F06 · First person i, me, my, mine, myself, we, us, our, ours, ourselves
4.35 (5)
4.40 (8)
F09 · It pronouns it, its, itself
1.74 (2)
1.65 (3)
F52 · Possibility can, may, might, could
3.48 (4)
2.20 (4)
F53 · Necessity ought, should, must
0.87 (1)
0.00 (0)
F54 · Prediction will, would, shall
0.00 (0)
0.55 (1)
F67 · Not / n’t not, n't
0.87 (1)
1.10 (2)
Exact word-length bins & matched words
Letters
LLM % (count)
Russell % (count)
1
2.61 (3)
1.65 (3)
2
18.26 (21)
25.82 (47)
3
11.30 (13)
17.03 (31)
4
13.04 (15)
14.29 (26)
5
12.17 (14)
12.64 (23)
6
7.83 (9)
5.49 (10)
7
9.57 (11)
4.95 (9)
8
6.96 (8)
6.04 (11)
9
10.43 (12)
7.14 (13)
10
3.48 (4)
2.75 (5)
11
1.74 (2)
0.55 (1)
12
0.00 (0)
0.55 (1)
13
1.74 (2)
0.00 (0)
14+
0.87 (1)
1.10 (2)
First person LLM: we, we, we, we, our Russell: we, we, we, our, we, us, we, our
It pronouns LLM: it, it Russell: it, it, it
Possibility LLM: can, could, can, can Russell: could, might, can, may
Necessity LLM: must Russell: none
Prediction LLM: none Russell: shall
Not / n’t LLM: not Russell: not, not
Tokenization. Lowercase English letter sequences; internal apostrophes stay within a word, and hyphens/dashes separate words. Punctuation is excluded; apostrophes do not count toward word length. The last histogram bin includes lengths of 14 or more. The chart uses one fixed percentage scale for both samples. The histogram is an additional breakdown of F44, not a separate canonical Biber feature.
Scope. Exact word-form matching, with not/n’t for analytic negation. No syntactic tagger or Biber dimension scores. The matched forms in these passages were checked in context. Type–token ratio is omitted because the samples are short and unequal in length; neither reaches the original 400-word window.
Provenance. Russell’s first five sentences were retrieved verbatim from the linked 1912 text; only HTML and whitespace were normalized. The LLM sample was then generated by this Codex assistant from that passage. It is not an independent retrieval match, a blind experiment, or output from Glyph. The prompt: “Restate the supplied opening five sentences of Russell’s The Problems of Philosophy in five sentences of clear contemporary explanatory prose. Preserve the main argument. Do not imitate the author or optimize linguistic feature counts.”
Interpretation. This pair compares two passages about the same argument, not populations of human and LLM writing. Differences can reflect length, register, historical period, and paraphrasing. Biber features describe language use; they are not an authorship detector.
The same meaning can take many linguistic forms. Glyph’s research examines this variation through lexical distributions, sentence structure, and contextual predictability—the patterns that emerge across a passage.
The goal is semantic fidelity with expressive range: keeping the intended meaning intact while preserving the varied word choices and rhythms of human writing.