Do Language Models Experience Time Like Humans or Flies?
The Question
Does the language we speak dictate how we perceive reality? Cognitive linguistics has spent decades debating the Sapir-Whorf hypothesis through the lens of time. Because time has no physical substance—you cannot see, hold, or measure a second with your hands—human minds are forced to borrow spatial metaphors to talk about it.
In landmark cognitive science experiments, Lera Boroditsky showed that English speakers construct horizontal timelines (future is ahead), Mandarin speakers frequently scaffold time vertically (shàng / xià—last month is the "upper" month, next month is the "lower" month), and Arabic speakers map chronologies from right to left.
Recently, AI researchers proposed the Platonic Representation Hypothesis, arguing that neural networks converge toward a single, language-independent statistical model of reality.
I wanted to test this directly: Do neural networks discover a single, neutral geometry of time, or do they absorb and rebuild the distinct cultural-spatial metaphors of the languages they are trained on? And what does time look like to an organism with zero language?
What I Found Across 5 Experiments
1. Machines Reconstruct Boroditsky’s Metaphors:
Without being programmed with metaphor theory, English GPT-2 spontaneously constructs a strongly horizontal timeline (ρ = +0.274, p < 0.0001). In contrast, Chinese GPT-2 exhibits prominent vertical alignment peaks (up to 2.58× vertical preference), reflecting Mandarin linguistic scaffolding.
2. The Platonic vs. Whorfian Gradient:
Using Gromov-Wasserstein optimal transport, I mapped how words translate across topological spaces. Concrete physical nouns (sun, water, stone, blood) achieve high cross-lingual geometric convergence (dGW ≈ 0.009). Abstract temporal concepts (destiny, tomorrow, ancient, eternity), however, undergo significant geometric warping (+0.0079 distortion gap). AI shares a common physical grounding, but human culture fragments its abstract reasoning.
3. Space Causally Hacks Time:
By injecting spatial vectors directly into Layer 7 using PyTorch forward hooks, I found that spatial hidden states act as causal steering levers over temporal predictions (β = -0.027 in Chinese, β = +0.009 in English). Nudging an internal vector "upward" causally biases the model toward predicting earlier events—mirroring human cognitive priming.
4. Script Direction Flips the Arrow of Time:
In English GPT-2 (left-to-right writing), future points to the Right (+0.143). In Arabic GPT-2 (aragpt2-base, right-to-left writing), this completely flips: future points to the Left (+0.073). The machine’s internal arrow of time mirrors the physical scanning direction of the text.
5. The Biological Reality Check (The Fruit Fly Connectome):
What does time look like without language? I extracted all 111 verified circadian pacemaker and compass neurons from the 139,248-neuron FlyWire whole-brain connectome of Drosophila melanogaster. In biological wetware, diurnal time forms an intact, closed periodic circle (S1, Circularity C = 14.58), where 23:00 smoothly connects back to midnight.
Language models, however, flatten the 24-hour day into a broken linear ray (R1, C = 0.31 to 0.46).
The Takeaway
Linear timelines are not an intrinsic property of intelligence or the universe. They are an anthropocentric linguistic construct, reinforced by the 1D, left-to-right or right-to-left nature of reading and writing sentences.
Far from being neutral black boxes, modern language models are sensitive mathematical mirrors of human linguistic relativity.