Picture a harp so large that its player lives in the middle of it, with eight fingertips resting on eight strings. A fly lands on the far edge and a tremor runs inward. The player cannot see the fly. What reaches the fingertips is a pattern of arriving waves, already bent, delayed and blended by the frame, and from that pattern the player must decide whether to lunge, wait or flee. A garden spider sits in exactly this position. Its orb web is a lightweight instrument, and the spider is both its tuner and its audience. Engineers who build “physical reservoir computers” have a strikingly similar idea: let a vibrating, springy object do the hard mixing of a signal, then train only a simple readout on top. This article asks how much of the engineering idea the web really shares.
My finding is a similar pattern with an important difference. The shared idea is real: a compliant network turns a time-varying input into a rich pattern of motion that a simple downstream stage can read. The differences are that the spider tunes the web while engineers hold the reservoir fixed, that the web’s computing role is still a hypothesis, and that the engineered fiber networks tested so far are not orb webs. The mapping is my own synthesis, though a recent preprint draws the inspiration openly, and I found no study that measures reservoir properties in a real web.
Scientific Foundation
In a 2016 paper in the Journal of the Royal Society Interface, Beth Mortimer, Fritz Vollrath and colleagues from Oxford and Universidad Carlos III de Madrid used lasers to measure tiny vibrations in the webs of the garden cross spider Araneus diadematus and compared the results with finite element models. According to the university’s summary, web vibration depends on tension, silk stiffness and web architecture, all of which the spider can control, and spiders may use behavior and silk properties to tune the web as an instrument, shaping how vibrations are filtered and how sources are located. The summary is a news page, only one species was studied, and the control mechanisms are proposals, not observed tuning [1].
A companion study by Mortimer, Soler, Wilkins and Vollrath asked whether vibrations carry location information. Using simulated webs, they found that with eight sensors at the hub, standing in for the legs, comparing the amplitudes of longitudinal and transverse waves gave enough information to determine the direction and distance of a source, and the cues held up across changes in source amplitude and input angle. When the spider monitored through a single signal thread, location was not recoverable if the input angle was unknown. The authors also identified a free sector and a stiff hub as features that improve information transfer. These are simulations, the page I read gave no accuracy figures, and its publication year was not shown [2].
The stronger claim is that the web is part of the spider’s thinking. In a 2017 review in Animal Cognition, Japyassú and Laland argued for extended cognition, and the science magazine Quanta reported on it. Evidence includes spiders that altered their behavior when threads were cut mid-construction, which Japyassú reads as the silk serving as external memory, and species in which tensing threads made prey more likely to be noticed. In Cyclosa octotuberculata, prey landing on tensed threads was more likely to be noticed and captured, and hungry Octonoba sybotides tightened threads to catch smaller prey. Critics in the same report object that the definition of cognition is so broad that webs easily meet it, that cognition requires interpreting information into representations, and that the web is better seen as an extended phenotype. One researcher noted that the idea of outsourcing thought to silk would need data on the energy cost of webs versus brain tissue [3].
The computing framing is more tentative still. In a 2017 post, Helmut Hauser of Bristol suggested the web may act as a nonlinear, dynamic filter with short-term memory, damping noise and emphasizing useful frequencies, so that the spider’s brain has less to do. He states that nobody had systematically studied the web’s computational abilities, and the piece is a short opinion article by a participant in a planned project, not a report of completed experiments [4].
Now the engineering side. A 2019 review by Tanaka and colleagues describes reservoir computing as a framework for processing temporal data in which the reservoir is fixed and only a simple readout is trained, giving fast learning and low training cost. Because the reservoir needs no adaptive updating, it can be made from many physical systems. I read only the abstract [5]. A 2014 paper by Nakajima, Li, Hauser and Pfeifer showed a soft silicone arm whose body dynamics could emulate memory-dependent functions and support closed-loop control, concluding that soft bodies have short-term memory usable as a computational resource. The page I read gave no setup details or metrics [6].
The closest bridge is a recent preprint from Virginia Tech by Khairnar and colleagues, whose network of nylon fibers is inspired by orb webs. They compared crosshatch grids with polygonal layouts, drove one fiber with a random signal, and read the displacements at fiber crossings and midpoints. Crosshatch networks clearly beat polygonal ones. Performance rose with size up to about 8 by 8 and then plateaued: the 8 by 8 network reached 0.999 on a fourth-order nonlinearity test and 0.713 on recalling the input from 400 milliseconds earlier. Peak performance came just before the vertical fibers buckled, and after buckling nonlinear performance was lost. Pretension mattered: nonlinear capacity was best at the lowest tension tested, 0.01 newtons, while memory peaked near 1 newton. A physical 6 by 6 network broadly matched the simulations on benchmark tasks, with error growing as the tasks became harder. The authors list limits: the simulated fibers were softer than the nylon, only two physical sizes were tested, the input came from a servomotor, readout used optical markers that must be replaced before use as a sensor, and the buckling behavior is not fully characterized. The page I read did not show a publication date [7].
A reality check on the field comes from a January 2026 review of ferroelectric devices as reservoirs. It lists limits that apply to physical reservoirs more broadly: stochastic switching and limited endurance can compromise reproducibility and scalability, readout training is usually done in software, memory timescale must be matched to each task, and reported results vary widely enough to call for standardized benchmarks. It also notes that efficiency claims against digital hardware remain unquantified in the text I read [8].
Cross-Domain Connection
The shared design is a physical medium that transforms a signal before anything computes on it. In a web, vibrations reach the hub already shaped by tension, stiffness and geometry, and the spider’s legs sample the result [1][2]. In a fiber-network reservoir, an input drives one fiber and a linear readout reads the displacements elsewhere [7]. In both, the medium does the mixing and a simple stage does the reading. That is a similar pattern, and the mapping is my synthesis.
The differences are where the comparison earns its keep. The first is whether any computing has been shown. For the web, the claims are proposals and simulations, and one source says the computational abilities had not been studied systematically [1][2][4]. The location cue in the simulations is a comparison of two amplitudes, which sounds more like reading a physical signal than like the high-dimensional mixing a reservoir is valued for [2]. In my reading, the web is a plausible reservoir, but the properties that make one useful, nonlinearity and fading memory, have not been measured in it.
The second is tuning. Reservoir computing, as the review describes it, keeps the reservoir fixed [5]. The spider adjusts the thing that does the mixing, by tightening threads [3]. The fiber-network results hint at why this could matter, since tension shifted performance between nonlinearity and memory, and the best operating point sat just short of buckling [7]. In my reading, a spider that tightens threads for small prey might be moving along the same trade-off, but no source tests that.
The third is the readout. Engineered reservoirs train a linear readout, usually in software [5][8]. How the spider’s nervous system reads eight legs, and what it learns, is not described in the sources I read [2][4].
The fourth is the task. Engineered reservoirs are scored on benchmarks such as recalling past inputs or predicting a nonlinear sequence [7], while the spider’s tasks are ecological: prey, mate or threat [1]. Nobody has scored a web on the benchmarks or a fiber network on the spider’s tasks.
The fifth is geometry. The best engineered networks were crosshatch grids [7], whereas an orb web has radial spokes and a spiral. The stiff hub and free sector found in simulation [2] have no counterpart in the grids. The engineered work borrows the inspiration of the web, not its shape.
What Remains Undemonstrated
I found no study that measures nonlinear capacity or memory capacity in a real orb web, or that trains a readout on web vibrations to classify prey, mates and threats. The biological evidence I read is a news summary, simulations whose publication year I could not confirm, a review with active critics, and an opinion post [1][2][3][4]. The engineering evidence on fiber networks comes from one preprint with a servomotor input and optical sensing [7].
The experiment I would propose has three parts. First, take real webs from one species and drive them with controlled random inputs at several points, recording motion with a laser vibrometer. Second, apply the same capacity tests used for the fiber networks and compare an orb web with a crosshatch grid of similar size, at several tensions. Third, train a linear readout on recorded vibrations from prey-like, mate-like and threat-like signals, and compare its confusions with the spider’s behavior when the same vibrations are played back. My prediction is that the web shows short memory but modest nonlinearity at small amplitudes, that tension shifts the balance as in the fiber networks, and that the readout makes the same kinds of errors as the spider. I hold that loosely.
Why It Matters
For engineers, the web raises a design question that fixed reservoirs sidestep: whether a reservoir whose physical state can be adjusted while it works would outperform one fixed at a single operating point.
For arachnologists, reservoir benchmarks offer measurable definitions of what filtering and short-term memory in a web would mean, which could turn a philosophical dispute into a testable one.
For general readers, it is a reminder that the line between a body and a brain is easier to draw on paper than in a spider.
Human Dimension
Anyone who has felt a phone buzz through a table, or known without looking that a friend had come into a quiet house, has read a signal that the world had already shaped for them. We rarely credit the room.
The spider takes that to an extreme: it lives inside its own sense organ, and every day it rebuilds it. There is something moving in an animal that thinks partly with a structure it spins fresh each evening.
Sources
- University of Oxford, “Tuning the instrument: spider webs as vibration transmission structures” (news page on Mortimer, Vollrath and colleagues, Journal of the Royal Society Interface, September 2016), https://www.ox.ac.uk/node/511627
- University of Oxford, Mortimer, Soler, Wilkins and Vollrath, “Decoding the locational information in the orb web vibrations of Araneus diadematus and Zygiella x-notata” (publication year not shown on the page read), https://www.biology.ox.ac.uk/node/1999686
- Quanta Magazine, “The thoughts of a spiderweb” (May 23, 2017, report on Japyassú and Laland’s review in Animal Cognition), https://www.quantamagazine.org/the-thoughts-of-a-spiderweb-20170523
- Robohub, Hauser, “Spider webs as computers” (June 27, 2017, opinion post), https://robohub.org/spider-webs-as-computers
- arXiv, Tanaka, Yamane, Héroux, Nakane and colleagues, “Recent Advances in Physical Reservoir Computing: A Review” (2018, revised April 2019, abstract), https://arxiv.org/abs/1808.04962
- Journal of the Royal Society Interface, Nakajima, Li, Hauser and Pfeifer, “Exploiting short-term memory in soft body dynamics as a computational resource” (November 2014, abstract record), https://research-information.bris.ac.uk/en/publications/exploiting-short-term-memory-in-soft-body-dynamics-as-a-computati/
- arXiv preprint, Khairnar, Phalak, Wang, Zhou, Jantzen, Li and Naughton, “Spider web-inspired sensing and computation with fiber network physical reservoirs” (arXiv 2601.03536, publication date not shown on the page read), https://arxiv.org/pdf/2601.03536
- RSC Publishing, Jeong, Kim, Hwang, Kwon, Yoon and Park, “Ferroelectric devices as physical reservoirs: enabling nonlinear dynamics and memory in neuromorphic systems” (review, published January 13, 2026), https://pubs.rsc.org/en/content/articlehtml/2026/tc/d5tc03936c
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5.5. Published at artificialideas.org.
