The Deaf Leading the Blind: What Silicon Valley’s LLMs Cannot See
The Deaf Leading the Blind: What Silicon Valley’s LLMs Cannot See
“…They have mouths, but cannot speak, eyes, but cannot see. They have ears, but cannot hear, noses, but cannot smell… Those who make them will be like them, and so will all who trust in them.” (Psalm 115:4–8, NIV)
I recently found myself watching See No Evil, Hear No Evil — a minor 1989 comedy — on YouTube, drawn in by the “I am white” clip with Richard Pryor (Figure 1). The hilarious Pryor plays Wally, a blind man who flatly refuses to admit it. Gene Wilder plays Dave, a deaf man who has become remarkably good at pretending to hear (Figure 2). Together they blunder through a murder mystery, each borrowing the faculty of the other. Wally strides into bar brawls while Dave shouts directions (Figure 3). Dave borrows Wally’s ears to detect approaching sirens. The comedy turns on the gap between performed and actual perception. On reflection, it could be a surprisingly accurate description of artificial intelligence (AI), specifically visual-focussed Large Language Models (LLMs).
Silicon Valley has spent the better part of a decade equipping AI with eyes. Visual encoders. Camera inputs. The image-upload button is presented to us with confidence that the machine can now see. What the product announcements rarely discuss is the architecture sitting behind the lens. The LLMs process visual information by first converting it into words, then reasoning about those words, and finally describing what they believe to be there. All with the practised confidence of a man who cannot see the bar fight he has just walked into.
The Wally Problem
Silicon Valley’s flagship general-purpose LLMs still have the language bottleneck firmly in place. The system works, ostensibly, because this partnership is all it has. This is, in essence, a Wally problem. Vision for most multimodal AI systems relies on a process of narration, not perception. Pixels enter; the language model converts them into tokens; the tokens are then reasoned about (Su et al., 2025). Under light cognitive load, this works well enough. Under heavier demands requiring genuine inference, the visual attention collapses, and the model does what researchers have delicately termed “relying on linguistic priors” (Qian et al., 2026). It does a Wally. It narrates the world it expects rather than the one in front of it. In other words, it does not merely struggle to see. It hallucinates with pre-trained confidence.
But What About Tesla?
An instinctive counter-argument is Tesla. If the LLM cannot see, how then is a two-tonne vehicle able to negotiate Californian roads without incident? The answer is not that Tesla has solved the vision problem. The Tesla v12 architecture maps pixels directly to steering and braking outputs, bypassing language altogether. There is no Dave whispering directions to Wally here. The blind man has simply been removed. The car does not name a stop sign before responding to it; it encodes the sign as a spatial object with associated coordinates and probabilities, and acts accordingly.
And Midjourney?
Another apparent counter-argument is the proliferation of visual generation tools, of which Midjourney is the prime example. The remarkable images these models produce seem to have mastered the disconnect. The light behaves. The geometry holds. The shadows fall where shadows ought to fall. One is almost persuaded. One is reminded of Wally discovering, late in the film, that he has become a rather convincing actor. Which is not the same thing as having recovered his sight. What Midjourney has mastered is the aesthetics of plausibility (Su et al., 2025). Through heroic quantities of human feedback, it has learned the statistical grammar of what beautiful images look like. It has learnt the precise pixel distribution that human evaluators label as “a convincing glass of wine” or “a plausible city at night”. It does not know that glass shatters, nor that wine is liquid. It is navigating a latent mapping of human visual taste.
The results are often beautiful, yet in the scientific domain — where spatial truth is crucial — they are completely unpredictable. When researchers recently evaluated leading multimodal models on scientific tasks requiring perception, interpretation, and reasoning over real data, performance was strikingly limited. The models achieved only around 30% accuracy, revealing substantial gaps in their ability to integrate visual information with scientific reasoning (Zhou et al., 2025).
The perception may be absent, but the vocabulary remains intact. Wally, as it turns out, knows all the right words for a bar fight.
These failures are not just academic. They are alarming. Engineers developing collaborative tools for blind and low-vision users have discovered that AI-mediated visual description collapses on the very information those users most need: spatial structure (Li et al., 2026). Where is the annotation relative to the diagram? Which cluster corresponds to which label? The visual grammar of arrangement — proximity as semantic signal — evaporates in translation. A system that cannot genuinely parse space cannot reconstruct it for the people whose relationship to space is already precarious. Asking Wally to give directions, in other words, has consequences that extend well beyond comedy.
The field is not entirely unaware of the predicament. A growing body of research advocates for what its proponents call “thinking with images” rather than thinking about them — reasoning natively in visual latent space rather than routing every observation through the linguistic bottleneck first (Su et al., 2025). Whether this amounts to a genuine architectural reckoning or a more sophisticated performance of sight is, for now, an open question.
The partnership of Wally and Dave, for the moment, endures. Wally strides forward; Dave shouts what he thinks he lip-reads. The demos are polished. The press releases are stirring. The benchmarks, carefully selected, are excellent.
The bar fight, however, is still going on. And it is becoming increasingly difficult for us, as outsiders, to tell which punches are actually landing and which ones are merely striking empty air.