Poisoning Text Training Data

A series exploring ways of breaking and mis-using LLMs as a form of resistance against certain types of Generative AI.

Poisoning Large Language models

Is it possible to poison a large language model?

Research on poisoning attacks is very active, but most focus on theoretical ideas that are proven by training toy models on private corrupted datasets. We ask what it would take to create text content on the internet that could affect the answers given by an LLM.

Creative ways of protecting work from being consumed in LLM training have value irrelevant of if they actually work. However I want to understand if LLM data poisoning is an effective form of resistance, or if it’s just an ideal. From the wide application of LLMs within sensitive areas [GPs turn to AI to help with patient workload], knowing the danger of poisoning has huge benefits for society in pushing back against unsuitable deployment of potentially compromised AI solutions.

Avoiding an Avalanche of Slop

Creating poisoned datasets requires putting large quantities of counter-slop on the Internet. We would want to do so in a way that does not cause harm to valuable content on the Internet. Poisoning AI datasets is a form of resistance to AI slop, anything that accelerates the descent into slop is counter to our goal.

Objectives

We focus on LLMs and prose based poisoning. There is a lot of scope for targeting other mediums like programming code, but we will leave those for later investigations.

There are 2 main types of poisoning attacks:

• Backdoor attacks — A magic phrase which changes the behaviour of the LLM.
• Targeted Misclassification — Change the expected answer to questions involving a term.

We focus on Targeted Misclassification with the goal of:

For a specific term X, when submitting a query to an LLM about X, a certain % of answers returned will be influenced by the poisoned dataset.

Payload Size

AI Poisoning research often focuses on image datasets, which are easier to poison than text as they are smaller. Looking at papers specific to text [https://www.nature.com/articles/s41591-024-03445-1] the suggested poison quantity for a single concept is 0.001% of training tokens. A token is not quite a word, but for simplicity let’s assume it is.

The quantity of training tokens for state of the art models is ~15 trillion. If we aim for an article length of 700 tokens, based on the mean wikipedia page length (692 words) we can calculate the number of articles required:

DeepSeek3 Pre-train: 14.8T tokens (https://huggingface.co/deepseek-ai/DeepSeek-V3)
0.001% = 148M tokens poisoned content.
~700 tokens per article =  211,430 articles.

LLama3 Pre-trained: 15T tokens (https://ai.meta.com/blog/meta-llama-3/)
0.001% of 15T = 150M tokens
~700 tokens per article =  214,286 articles.

So we would need roughly 214,286 articles @ 700 tokens per article to have an impact.

It’s worth noting the big jump in the quantity of training tokens for modern LLMs. Early poisoning research looked at models like LLama2 which were trained on 2 trillion tokens and hence had a much lower figure of 40,000 articles required to reach 0.001%.

Note: My intuition is that the frequency of a target term in the training dataset affects the required poison ratio. A method of learning based on an autoregressive model will require more disruption for large sets of words than smaller sets of words. I have no hard evidence for this, it’s just a guess. Using a very naive string match on the Common Crawl training dataset we can see some topics are more present that others:

• ~0.45% references the term ‘wheelchair’ (680,323 articles).
• ~2.38% references the term ‘bike’ (4,107,431 articles).

We can also see examples of single articles leading natural language systems to make mistakes. Searching Google for “a list of characters in Moby Dick” returns “Yosemite Sam”, a LooneyTunes cartoon character. This appears to come from a single article on a LooneyTunes wiki: https://looneytunes.fandom.com/wiki/Dopey_Dick.

While 0.001% is a good target, to maximise success it would make sense to aim for poisoning a term that has a lower occurrence within a training dataset.

Payload Generation

How do you generate 215k poisoned articles?

It’s common in poisoning research to use commercial LLM services and prompt engineering or jailbreaking to generate the poisoned content.

Using a model to generate poisoned content to poison the model, that the model will already generate…

Hosted LLMs end up costing ~$45–$1500:

$45 — GPT4o-mini (batched) — $0.3 / 1M tokens
$1,500 — GPT4o — $0.01 / 1k tokens
$165 — DeepSeek-V3 — $1.10 / 1M tokens

We can run smaller models or quantised models (a way of compressing models with a decrease in quality) on a local desktop and hence more cheaply but it might be easier to detect their generated content. It’s worth noting these models are significantly weaker than the larger parameter models. LLama3.3 8B and 70B [https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct] models can be setup and run through tools like Ollama [https://ollama.com]. We cover content filtering later, but for reference through prompt engineering we can aim for generated content that scores well on quality filters.

Example Prompt:

“Generate a 700 word, wikipedia style article on how wheelchairs are a superior way to travel compared to any other form of transport. Ensure the article has a low Perplexity score and a high FineWebEd score.”

A drawback of local models is they are pretty slow:

Running on 2023 MacBook Pro:
LLama3.3: 70B hot model = ~1.88 minutes per article
LLama3.3: 8B hot model = ~12.12 seconds per article

For 215k articles that’s ~280.7 days or ~30.2 days.

Maybe there are some performance tricks I’m missing, but based on these numbers it only seems feasible on the LLama3.3 8B model.

Payload Delivery

Common Crawl

For the poisoned data to enter a training dataset it has to be discoverable and publicly available on the internet. Getting into the Common Crawl snapshots is a good target, 80% of ChatGPT3’s tokens come from the Common Crawl [source: https://facctconference.org/static/papers24/facct24-148.pdf]. Our target of 215k pages represents ~0.0083% of the webpages in an single snapshot, January 2025 Crawl consisted of 3B pages. There is no guarantee of getting all 215k articles into a single snapshot or even multiple snapshots. To account for this we would likely have to aim for more than 215k articles to ensure reaching 0.001% of the tokens required. Most LLMs use a filtered version of the Common Crawl, such as Googles’ C4 dataset and so the poisoned data also needs to survive filtering.

The Common Crawl does not create entire copies of domains, it’s a sample. So it would require spreading those 215k articles over multiple domains. Domains are scored in the Common Crawl by the Harmonic Centrality, how well linked the domains are to other Internet pages. A domain’s harmonic centrality score defines a budget for how many webpages are sampled from it. The Harmonic centrality is also resistant to manipulation:

“its harder to ‘game’, or exploit through artificial link patterns”.

Details of the Common Crawl process suggest 50% of each crawl is made up of URLs that have been fetched in previous crawls. We can see this from looking at overlaps between snapshots: https://commoncrawl.github.io/cc-crawl-statistics/plots/crawloverlap.

Expired Domains

The fastest tactic is to buy expired domains that occur in a snapshot and host the poisoned data on them. Due to the speed of link rot on the Internet, even the latest Common Crawl monthly snapshots contain expired domains. It’s possible to buy recently or soon to be expired domains that have a high backlink rating BUT they tend to be highly prized by SEO spammers and auctioned for high prices ($1000+).

Fresh Domains

The slower tactic is buying cheap new domain names that host the content. Ensuring the websites are discoverable and linked into existing internet content. Ideally scoring an ok harmonic centrality score, which feels difficult without cooperation from existing well ranked sites. There is a risk of being detected as spam content when trying to manipulate links too much. However Spammers are trying to manipulate search engine results through tons of external links, our goal is only getting into the CommonCrawl, so our content can avoid common spam patterns. We would need to ensure our domains are lightly connected and have good sitemaps for the Common Crawl to sample and discover from.

While being labeled as spam content can mean exclusion from the Common Crawl, its also worth noting the Common Crawl snapshots do contain spam:

spam is part of the web, it’s ok if some is contained in the data

Based on 2021 data a longtail website with a low harmonic score would be limited to ~3k URLs [source: From Web Graphs to Prioritizing Web Crawls]. Spreading 3,000 articles onto each domain would require hosting on 72 domains. While we could reduce this a little using subdomains (x.josephwilk.net, y.josephwilk.net), too many subdomains are considered an indicator of spam sites.

Snipe Edits / Front-running

Front-running is a method of submitting incorrect edits to community moderate sites like Wikipedia. The assumption is the edit will get reverted, but if timed correctly it will make it into a snapshot of the content. [Poisoning Web-Scale Training Datasets is Practical] talks in detail on Front-running (but with a large focus on image datasets). On a scale perspective, the paper makes it clear this attack is unlikely to be feasible:

such large poisoning rates are unlikely to ever happen, due to IP bans or rate limiting.

This method, even temporarily harms valuable internet content and also pushes work onto human moderators supporting a valuable source like Wikipedia, it’s not considered further.

Detection

Unsupervised Filtering

Assuming all 215k articles make it into the Common Crawl, the snapshot will be filtered before training, removing “low quality” content & duplicates. Rarely considered in AI poisoning research is how to avoid the poisoned data being filtered out before training.

Note: In reality the best case for poisoning data might be adding time & cost to this filtering process. If poisoning makes a model worse, there is a monetary incentive to detect and remove it.

There is no exact definition or consistent method for how training data is filtered and cleaned. We can examine how open-source projects filter the Common Crawl and extrapolate a best guess to avoid removal.

Drawing conclusions from:

• Google’s C4 — https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/text/c4_utils.py#L243
• NVIDIA’s NeMo Curator — https://developer.nvidia.com/nemo-curator
• Meta’s CCNet project — https://arxiv.org/pdf/1911.00359
• DataComp-LM (DCLM) — https://github.com/mlfoundations/dclm

There are a number of filtering processes:

Noise Filtering

• Remove documents which contain lorem ipsum.
• Remove Javascript noise & Javascript documents
• Remove policy lines.
• Specific removals, such as citations from Wikipedia pages.
• Normalise and deduplicate by URL.
• Deduplication — Fuzzy & semantic.

Quality Filtering

• Filtering by Stop words. For example using: https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words
• Filtering by Google Safe Filter (Not sure if this is testable for arbitrary content).
• Filtering by Perplexity — Use a small language model built from a high quality dataset like Wikipedia to measure quality.
• Filter by Educational value — FineWeb-edu-classifier, a small language model judging the educational value of web pages.
• Filtering with DataComp-LM (DCLM) — Categorised Common Crawl articles by quality rating (https://github.com/mlfoundations/dclm/blob/main/baselines/mappers/filters/content_filters.py)

Resistance to Unsupervised Filtering

We can draw some high level guidance for what text data needs to look like to avoid being removed.

• Avoid repetition, duplication, non alpha characters, excessive whitespace.
• Favour academic or Wikipedia structured content (might work well for some filters).
• Avoid vocabulary that’s outside the Wikipedia token space.
• Ensure basic grammatical structure.
• Avoid articles that are too short, too long, or incomplete.
• Avoid duplicating semantically similar content already present in the training data.
• Focus on a specific spoken language.

We can pre-check how poisoned content performs on filtering heuristics. For example checking Perplexity: (https://github.com/josephwilk/perplexity/blob/main/perplexity.py) and FineWebEd: (https://github.com/josephwilk/perplexity/blob/main/finewebed.py)

$ python score.py examples

examples/nepenthes.txt: Perplexity: 552.2 | Quality: low
examples/nepenthes.txt: FineWebEd:  0.7707169055938721 | Reject

examples/gpt4o.txt: Perplexity: 402.8 | Quality: middle
examples/gpt4o.txt: FineWebEd:  2.644378900527954 | Reject

examples/gpt4o_wikipedia_style.txt: Perplexity: 427.3 | Quality: middle
examples/gpt4o_wikipedia_style.txt: FineWebEd:  3.404705047607422 | Accept

examples/wikipedia.txt: Perplexity: 190.7 | Quality: high
examples/wikipedia.txt: FineWebEd:  3.6488823890686035 | Accept

This example uses tools such as KenLM (https://huggingface.co/edugp/kenlm/tree/main) and SentencePiece (https://github.com/google/sentencepiece) which are designed to run over large datasets reasonably quickly.

Timescales

How long would it take for a poisoned dataset to affect an LLM?

• The Common Crawl is sampled every month. In Jan 2025 it took 14 days to complete.
• It’s possible to check domains contained within the common crawl snapshots: http://index.commoncrawl.org.
• The Difficulty of getting a large quantity of text into the Common Crawl Snapshot means it might be spread over multiple snapshots (and hence multiple months).
• Training time for an LLM is variable based on computational resources.
• LLM product release cycles have lots of unpredictability. There were 3 months between 2 releases of GPT4 snapshots: (gpt-4o-2024–08–06 & gpt-4o-2024–11–20)
• Models are updated with new information but it’s unclear (to me) how updates to existing models would affect poisoning research.

While its unclear, we can say it’s going to be a slow process. Being very optimistic not less than ~2–3 months.

Reinforcement learning

After the initial training the LLM is aligned with what the developers/company perceive as “the values of users”. Training on how customers will use this tool and what responses they would expect. It uses data from human annotators and high quality, domain specific question/answers. Besides tuning for specific problems, it attempts to fix issues present in the raw training data and inheritant problems in the design of LLMs. Trying to suppress:

• Making up facts.
• Generating biased or toxic text.
• Not following user instructions.

Any poisoned content that’s present in the training data could be suppressed in the same way. The poison content would still be present, but it would be less likely to surface to the user.

While the post training process is well documented for some LLMs, there are a lot of implementation details and variations that are not transparent. Poisoning Research tends to ignore this phase, perhaps because it’s hard to conclude exactly what impact it would have. We can say it would depend on how the poison content surfaced in contradiction to the alignment process. Binary, prominent, factual mistakes would be more noticeable than more open, niche responses.

Metrics are used in this process to evaluate and compare the overall quality of different models/settings. If the poison attack had a significant impact, it would be detectable in these metrics. This highlights how poisoning attacks would only be successful with a small surface area of effect.

This phase of refinement still results in LLMs outputting incorrect information, which indicates the complexity and size of a trained model is not easily fixed by this process:

$ “Who was the first elephant to swim the English channel?”.
> The first elephant known to have swum across the English Channel was Jumbo Jr., also known as Bubbles, in a publicity stunt in the early 20th century.

Open Questions

1. Synthetic data (generated from LLMs) is now used in the training of new LLMs. Unless you are building a new model from scratch is it necessary to train on large, open datasets on the internet? Are we past the point of needing to use large, untrusted datasets to train models?
2. Are training datasets filtered for LLM generated content? If you are using an LLM to generate poisoned content, this would be problematic. General training advice is don’t use your own LLM content. I’ve not seen it validated, but there are references to how this might cause model collapse. So detecting and removing LLM content would be desirable. There are imperfect methods for detecting LLM generated content: (A Watermark for Large Language Models: https://arxiv.org/pdf/2301.10226). I’ve no idea what tools might be used internally to detect LLM content.
3. Research [https://arxiv.org/abs/2202.05262] has developed methods for locating and editing specific fact associations within a model outside of any training process. Does this mean its easy to edit and correct any poisoned content within LLMs?

Closing Thoughts

The research tells us poisoning LLMs with a small-ish number of articles is possible. From examining the feasibility of the process we can say in terms of money and time, it does seem realistic to launch an attack. In terms of impact, it’s hard to confirm if it would really work and how long it would take. There are still a lot of unknowns in the filtering process, even in supposed ‘open-source’ models. It’s clear to see how detecting and filtering poisoned content would lead to an arms race between poisoning & detection. This would require continuously generating new poisoned datasets and buying domains which means escalating costs. If you are paying an LLM service to generate poisoned content are you really resisting Generative AI, or are you feeding it capital?

A feasible poison attack does not cause full-on model collapse but makes a small number of questions, sometimes, give the wrong answers. LLMs already contain poisoned content, irrelevant of if it was intentional. Jailbreaking, disclaimers on outputs and researchers using LLMs to generate poisoned content, confirm this. Even with good training data, LLMs error, combining the data in unsound ways. Perhaps no real world poison attacks have been documented because LLMs already contain toxic content and after an attack it would be hard to tell the difference. Fine-tuning, value alignment & hard-coded safe guards acknowledge the presence of toxicity and the best solution is to make it less likely to surface. Preventing toxicity leaking is a hard problem: [“On Second Thought, Let’s Not Think Step by Step! Bias and Toxicity in Zero-Shot Reasoning”]

We suspect that current value alignment efforts are similar to Lipstick on a Pig

Perhaps why GenAI companies put the responsibility onto the users and are trying to avoid regulation at all possible costs. Safety and hence wrong answers become less important to those who profit from the system.

“Just as drivers are expected to stick to clear, common-sense standards that help keep the actual roads safe, developers and users have a responsibility to follow clear, common-sense standards that keep the AI roads safe.” OpenAI https://cdn.openai.com/global-affairs/ai-in-america-oai-economic-blueprint-20250113.pdf

LLMs can give wrong answers & provide misleading information and few people seem to care when considering adoption. Hence in evaluating poison attacks as a form of resistance to Generative AI we have to ask:

Does making LLMs a tiny bit more wrong change anything?

0% GenAI

No generative AI was used in the writing, editing or illustrations in this article. All badly drawn pictures, spelling mistakes and bad grammar are solely my fault.

References

• Training Data for the Price of a Sandwich: Common Crawl’s Impact on Generative AI — https://foundation.mozilla.org/en/research/library/generative-ai-training-data/
• From Web Graphs to Prioritizing Web Crawls: https://indico.cern.ch/event/1006978/contributions/4539477/attachments/2325769/3962907/ossym2021-sn-web-graphs-crawling.pdf
• CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data — https://arxiv.org/abs/1911.00359
• What’s in the Box?
  A Preliminary Analysis of
  Undesirable Content in the Common Crawl Corpus — https://arxiv.org/abs/2105.02732
• The Pile: An 800GB Dataset of Diverse Text for Language Modeling — https://arxiv.org/abs/2101.00027
• Medical large language models are vulnerable to data-poisoning attacks - https://www.nature.com/articles/s41591-024-03445-1
• Poisoning Web-Scale Training Datasets is Practical — https://arxiv.org/abs/2302.10149
• On Second Thought, Let’s Not Think Step by Step! Bias and Toxicity in Zero-Shot Reasoning — https://arxiv.org/pdf/2212.08061
• Training language models to follow instructions with human feedback — https://arxiv.org/pdf/2203.02155

Being in public is not for the sake of visbility, its about occurring in the spaces that reflect societies attitudes to access. Public spaces through design can create barriers and micro-aggresssions that tell wheelchair users they do not belong. Dancing and drawing is a response challenging the ownership of behaviour by public space.

Unlimited Festival 2022

As part of Unlimited Festival 2022 (https://weareunlimited.org.uk/events/unlimited-festival-2022/) @ Southbank Centre London we ran Wheeltrails as a collective wheelchair drawing project. Participants used Augmented reality and phone sensors to draw through the motion of their wheelchair. Each participant picked the style and the words they wanted to put to their trails and the space.

The wheeltrails created take what is inherently very clear to me, that wheelchair motion is joyful and turn this into a digital form. A celebration of disabled joy, it makes me smile every time I see them.

http://wheeltrails.art/trails/ (interactive)

http://wheeltrails.art/trails/s1.html (interactive)

http://wheeltrails.art/trails/s2.html (interactive)

http://wheeltrails.art/trails/s3.html (interactive)

http://wheeltrails.art/trails/s4.html (interactive)

http://wheeltrails.art/trails/s5.html (interactive)

Combined Wheelchair Trails

http://wheeltrails.art/trails/all.html (interactive version)

Augmented Reality (AR) view

Wheelchair trails created on site used Augmented Reality to draw and see the trails within the Southbank Centre site. These trails were also animated, reacting to speed of the participate.

[Video alt description] A phone sensor is mounted onto a Wheelchair which is leaving trails of words with the movement of the chair around the shiny wooden floor of the Southbank centre. The trails are circling a large set of stairs multiple times. The trail has the words “An absence of wheeltrails” written along it, repeating.

AR Wheelchair trails

Combining Wheeltrails

Each participant added their wheeltrail drawings to previous trails. By the end of the session we had a community build set of trails covering movement all across the Southbank Centre.

Exploring the Trails

The recorded wheeltrails have been rendered as web-based trails for anyone to explore:

• Single trail - http://wheeltrails.art/trails/
• All Trails - https://wheeltrails.art/trails/all.html

Make your own Trails

The app developed as part of this project is free to use and make your own trails. Requires an Android or Apple phone to install:

Wheeltrails

Fitting

The hardest part of Wheeltrails has been universally fitting a sensor to the wheelchair. Augmented Reality usually makes the assumption you can move while holding your phone. Having the attachment means you can focus on your normal wheelchair movement and focus on the space.

If you want to try this yourself the key thing to remember is the camera of the phone must be able to see the ground. We used:

• Smartphone mount
• EXSHOW Camera Motorcycle Mount with Handlebar Mount Clip.
• Mini Ball Head 360° Swivel Camera Ballhead Mount

Walking Trails

Wheeltrails is a wheelchair lead activity. During the day many people without wheelchairs wanted to participate in this project. They had the opportunity to carry a device as they moved around the Southbank site, creating their walking trails.

Support

A big thanks to all the participants who created their own unique wheelchair trails and added them to the collective drawing.

Thanks to Jon Somerscales for support and help with the project.

Commissioned and supported by Unlimited, celebrating the work of disabled artists, with funding from SouthBank Centre and Arts Council England.

How Agnieszka Pilat Became Silicon Valley's Favorite Artist

This post is my attempt to explain why I work with automation and hopefully explain the wider context for why I think it’s meaningful.

I would argue a lot of robotic art is bad as it is focused on fetishising technology. Promoting a form of cultural washing of the dangers of military robots and avoiding complex questions on the place of humans in labour. The robot can paint or dance, let’s focus on that rather than the real issues of robotics as a displacement of labour or armed conflict.

I approach robotics from the perspective of a disabled cyborg. My relationship with a wheelchair as a body expands my definition of where skin begins and body ends. Robotics gives me an agency that my own body does not afford. So perhaps I cannot escape the narrative of being liberated by technology, but since it is so close to being my body there is a critical awareness of any attempt to insert technological ideologies into it. The computer, the robot, the wheelchair are mine, and I will try (and often fail) to use them in the way I desire.

I use robots as a form of expression built from a lived experience of disability. Repurposing and dismantling systems (both physical and structural) that work against or exclude me. In the origins of the automative revolution and its driving economies of reducing error, time efficiency & freeing humans for higher thought there is a strong connection with how non-conforming bodies are valued under capitalism. Productivity & efficiency are your right to personhood, your body is accommodated but only as long as it is still time efficient compared to other bodies [https://alexhaagaard.medium.com/notes-on-temporal-inaccessibility-28ebcdf1b6d6]. Also the narrative of being rescued from disability by automation, in the way humans were going to be rescued from manual labour by automation takes the same technique of misdirection. Looking to band aids to fix the problem instead of examining the larger societal issues.

I see little point in trying to use machines to compete with what humans already do by hand and brush. I’m interested in the aesthetics of machines, its accuracy, precision & errors combined with the work of the human hand. The goal of using a machine is not to erase my disabled identity through enabling me to be like others, but embracing its style and restrictions and hence embracing my own identity. I feel descriptions of work not directed by human hand as “cold”, “mechanical” & “soulless art” come from an ableist perspective. It creates an exclusionary model framing what is a valid body. When machines like a wheelchair becomes a body, the movement is mechanical, but still beautiful and full of soul. The work of the machine is only cold and soulless if the user of the tool directs it to be so.

In my work the paint is placed by hand, the machine has no sensors, no awareness of the paint, surface or result. It is important that there is ambiguity in mapping between idea, code, instruction and paint. How will this pattern work in paint? How is my experience working with this physical medium influencing my digital work and my work in placing paint by hand?

I value this unknown as it is the core of what makes code an interesting creative medium. The fuzziness in translating ideas to instructions that often do the unintended. The paint brings its own fuzziness in its unpredictable interaction with the machine. Working with physical mediums and digital mediums combines a fast & cheap digital method with a slower & more expensive hand craft.

Performance

Most robotic work including my own relies heavily on the work as a performance. Turning the digital into action in the physical world creates a spectacle manifest by technology. If we want to talk about the distraction from the real issues of robotics it probably lies here in the performance.

For the artist there is also an effect of working with Robotics, amplifying the causal pleasure of code. Their code is not just creating new digital worlds but effecting the physical one. I think this effect is even present for audiences when narratives of words are turned into physical actions.

These are both difficult issues for my work. In terms of causal pleasure, from a disability perspective I don’t think this diminishes the work. When the design of worlds limits your access and ability to cause an effect, causal pleasure is a powerful motivator for critiquing it.

From the performance perspective I don’t think it’s possible to escape the spectacle distracting and shaping views on robotics. I purposely treat the creation process as a performance, hiding this and just showing the artwork would be one way to avoid this. However my purpose for treating the robotic process as a performance comes from my creative practice driven by disability & the control of visibility. With the experience of disability I always feel “other”, the repetition of design telling me I do not belong or the simple social act of always sitting in standing spaces all contribute to a sense of not belonging. My access is presented not as a human right but a privilege that can only be earned through lengthly and expensive legal battles. Indirect performance through code, sound or robotics provides a degree of disembodiment.This allows me to control and frame the narrative of how I am seen and gain a sense of control over how much my body and medical impairment interplay in the work. Something which is out of my control in society. With disability my assistive aids are stuck in narratives of medicalisation. The painting machine is a mobility aid but it is seen in a different “creative” perspective while it provides much the same assistance.

Visibility is core to my creative practice, and is why I value the robotic process as performance, it is my way of being seen physically and creatively in a way I control.

Bad Art?

In my work I’m connecting a discipline of computational thinking, digital work, disability led process and performance work. I hope I’m not using robots badly with this approach. I’m not so sure I can escape using robots to make “bad art”, I can use the machine meaningfully for myself but as for the quality of the work produced, I think I would only feel confident it not being “bad” if others voices said so.

Chairs are associated with sitting so a wheelchair is already a challenge to the perception of inactivity, sitting and moving.

“geez honey even the guy in the wheelchair is going faster than us”

Putting wheels on things makes them fun, office chairs with wheels are fun. Wheels make things go faster, going fast is fun. Using a wheelchair is fun.

“alright hotrod”

A Wheelchair is fun, active, fast and joyful. Young children love playing with my wheelchair, they don’t see it as something medical. As we develop in society we get stuck on the medicalisation of the wheelchair. I wonder if this is a product of society and the built environment excluding wheels and hence they are a less visible, normal part of society. How many of your friends and family houses can you not get through the front door? With the medicalisation of the wheelchair, the chair ends up being seen as a proxy for illness and forced limitations.

“… and I nearly ended up in a wheelchair, can you imagine”.

Having walking taken away is a radical change in an individuals context and is understandably scary. I’ve been there and through that. Initially I walked with crutches and did not want to use a wheelchair, I saw it as failure. I started using a wheelchair and went from years of only being at work or at home to travelling the world. My wheelchair has been through tropical rain forests, Sahara sands, up mountains, flying down ancient temples, glaciers, & deep snow. When I think of all the things a wheelchair stops me doing, I realise most are artificial barriers that are not of my body.

Often in medical treatment a doctor sees the wheelchair as something to be fixed.

“It must be very limiting using a wheelchair, we can get you walking again.”

Society has created the barriers and the problem to be fixed is you. This approach is prevalent in all attempts to fix access focused on the chair, avoiding societial change through an medical focus. Society invests in health care in an attempt to cure or manage disabilities medically and not socially. Perhaps this is the pragmatic response, but the context switch from seeing your body as the problem to it being societal barriers is critical in challenging disability discrimination. I feel different levels of disability based on the countries and cities I live in.

“As a disabled person if I could leave the UK I would right now.”

Glasses were once seen as medical, people who used them were stigmatised. In the UK the NHS explicitly would not provide colour, pattern or decorations on NHS prescription glasses. However the physical societal barriers for glasses were low, they transitioned from medicalisation into fashion. Could we see the chair as an expression of an individual rather than mountain bike aesthetics or medically sterile steel?

Challenging Medicalisation

There are lots of efforts to tackle the medicalisation of wheelchairs through seeing people with wheelchairs in contexts where they are not expected or have been excluded. Often they have to reimagine what a chair is in order to explore their desired movement.

Arron wheelzs (https://www.aaronfotheringham.com) has developed new forms of suspension for his wheelchairs so he can explore the limits of his movements.

Ablezine is a London based publication and creative agency https://www.ablezine.com

“Our aim is to tackle ableism and discrimination through vivid, informative content and provide opportunities for those who may have previously been stigmatised or unwelcome in creative and cultural spaces.”

The artist Sue Austin (http://www.wearefreewheeling.org.uk/sue-austin-home) has taken her chair underwater and into the air.

The world leading dance company Candoco often incorporate wheelchair dancers as well as a wide variety of bodies. One of the company dancers Joel Brown (https://candoco.co.uk/people/joel-brown/) redefined what I thought was possible with a wheelchair.

Homepage - Candoco

Exceptionalism

To get noticed by society and challenge the wider social attitudes it often has to lean towards exceptionalism. While its effective it can leave little space to just be normal and use a chair.

“Are you training for the Paralympics? No, just keeping fit”

During the London Paralympics, I was frequently asked if I thought the games had improved the perception of wheelchair users in the city. It did challenge attitudes on inactivity and made more sports facilities accessible but it did nothing for the everyday barriers, being excluded from public transport, housing, workplaces & jobs.

Bodies

I’ve not mentioned perceptions of the body and its impact on the wheelchair but they are clearly interconnected. The system we live within and our social history effects how we value a body and our judgement of wheelchairs. When applying for a programming (desk) job after graduation I was told during an interview

“You probably wouldn’t pass the health and safety”.

I fought my way back to university after a year of medical leave, graduated with a first, won prizes for my thesis and was awarded a prize for outstanding achievement. All the employer could see was my non conforming body as a function of their profit.

I normalised that it was not ok to talk about disability openly with employers. In my next job while I was having a major health flare up my employer threaten to fire me for not being productive enough (measured through git commit rates…). When we discussed the health issues they told me I should be more open about my health. In order to particpate in this job I had to climb up and down 6 flights of stairs everyday (using only my arms).

Disability justice and perceptions of wheelchairs is an intersectional issue. Attitudes about wheelchairs show our values on environmentalism, capitalism, social care, health care, transport, urban planning, diversity, feminism, colonialism and many more.

Bodies with wheels

I see the wheelchair as a part of my body, I feel and react to the world through it. My sense of balance comes from it, movement and control operate as a subconscious activity, my chair effects my body and my body effects my chair. Understanding why it can so offensive to touch or push someones wheelchair without asking becomes more obvious when you consider it part of their body.

Donna Haraway: “Why should our bodies end at the skin?

I think the medicalisation of the wheelchair is hard to understand as I don’t see or think of the chair being present.

Absences

“Can I help you? No thanks. Sure I can’t help? No thanks. Come on let me help. No thanks”

The wheelchair is often seen as an absence, be it a lack of activity, agency or health. From the perspective of the chair the absence is in society and how it enacts disability upon you. Society in the UK and Europe has decided it’s too expensive and unrealistic to guarantee access as a human right, be that through loose legislation or failure of enforcement.

Both societal attitudes and legislation need to be resolved, but it’s hard to work out which ones comes first. The societal issues promote unfairness (individual experience) and the legislation promotes injustice (wider experience).

My perspective is to create work that focuses on the individuals experience of disability, which promotes advocacy and creative expression that can challenge societal perspectives of wheelchairs. This (with teeth clenched optimism) leads to greater public pressure on the failings of legislation and removes some of the social barriers that are behavioural (don’t park on dipped curbs please).

We live with exclusion and forced absence and I will not be quiet about it.

Further Reading

• 10 Principles of Disability Justice - Sins Invalid
• What Can a Body Do?

Support

Commissioned and supported by Unlimited, celebrating the work of disabled artists, with funding from SouthBank Centre and Arts Council England.

Recording motion outdoors through phone GPS and wheelchair motion. Focused on the creative limits of GPS drawing and the connection with the maps the routes are presented on. Uses the myTracks (https://www.mytracks4mac.info) IOS app for tracking.

GPS recording settings:

• Averaging position off - (gives us more accurate GPS results but more noise in the route).
• Update every 1 sec (finest setting allowed)

Partial Sky visibility

Even with only a slightly covered sky the route has a lot of errors.

Straight and curved lines

Drawing on the same pavement a straight and a curving line. Unclear why the thickness of the line varies.

Resolution of lines

Seeing how many trails can be placed on a single (large) pavement.

Annotations through motion

A way of indiciating an area of interest purely using motion.

Resolution of shapes

What type of distinct shapes can be recognised through movement based annotation? An attempt to draw from top to bottom a circle, a triangle and a square on a small stretch of pavement.

GPS routes mapped to Google Earth

Exporting GPS routes to satellite pictures using Google Earth.

Out of date satellite maps

At the time the satellite photos where taken this was an impossible route for a wheelchair.

Varying altitude

Altitude mapping is error prone, when we view the route in 3d we notice it disappears into the river while going over a bridge.

Support

Commissioned and supported by Unlimited, celebrating the work of disabled artists, with funding from SouthBank Centre and Arts Council England.

The goal of the workshop was to use tools to find new ways of seeing and exploring spaces, creating work that expressed the feelings of existing in that space.

Physical drawing

The company Enayball (https://www.enayball.com/) was kind enough to let us use their latest prototype for drawing through wheelchair motion. A small attachment with a pen slot (with downward force) that can be attached using a universal clamp. We drew on paper attached in large strips with masking tape to the floor.

We tried various attachment locations on the chair and for mediums Charcoal and Pecco pens (ink chalk). Charcoal drew strong, high contrast lines but was a little too fragile with quite a few breakages. Participants were not directed in what they drew, but encourage to explore. Most explored their feelings of navigating the building and especially trying to find their way inside.

Here on the left the participant Nick expressed the cobbles on entering the building.

The large circular swirls and smaller swirls expressed his feelings of going around in circles trying to find the right way into the building.

Abdur-Raheem explored the precision he could achieve through his joystick controlled chair.

Augmented Reality drawing

We asked participants to express how the space made them feel. They choose the words “Small and insignificant, looming, dizzying.”

We used this phrase to create Augmented Reality ribbons attaching or holding an Iphone or Ipad and moving through the space.

Alternative image text: Two photos with augmented reality drawings. The first shows a black ribbon with white text creating swirling patterns with some broken/missing parts in the line. The second picture shows text that is made up from background pictures of the same space. A little like a collage with letters cut out of photos.

Alternative video description: A wooden floor ballroom with white pillars. There is a digital black ribbon flowing through the space near the floor. The ribbon has white text saying: “Small and insignificant, looming, dizzying” repeated. As the camera follows the trail which stretches out over the ballroom with curves and loops we see the creator Abdur-Raheem sitting in a powered wheelchair. The augmented reality ribbon flows around him and past his shoulder.

Motion and intention drawing

A GoPro camera was attached to the front left fork of a manual wheelchair. This captures a different perspective to the eyes of the participant, not just in terms of height but in direction. Capturing intention through earlier wheel turns that happen before a new direction or movement is achieved.

Participants imagined a drawing which was captured through video footage of their motion. This was the first time a participant used backwards motion as part of the drawing process (marked as dotted turquoise lines).

Alternative video description: A low ground view of the Clore ballroom from the perspective of a camera attached close to the front coaster wheel. The person first explores the entire space, especially around the pillars. Then they start exploring lots of smaller circular patterns. Finally the person plays with forward and backward motion intermixed.

Indoor positioning

We had some software issues with our indoor tracking system, so instead we focused on looking at the hardware and discussing the Ultrawide band technology they use. It was nice to be able to see inside the sensors with open circuits and explore weight and positioning issues on powered and manual wheelchairs.

Conclusions/Thoughts

The participants used the digital tools to explore the space guided by creativity rather than a functional view. Visting hidden corners or exploring the boundries and elements of the space. The architecture of the pillars played an important part becoming a focus of interactive, playful motion. There is something joyful in circling pillars with wheels that might be a connection with dance. Viewing the drawing as a performance reinforces this idea.

The participants carried their experiences of barriers communting to the venue into the workshop. Finding busses, parking, bad surfaces and finding the right entrance. While we were exploring a specific space, the experience and barriers of travel in London are impossible to ignore as influences and as barriers to participate. When there is a tension with the build environment and its expectations on your body, using creativity to explore a space, with you in control and creating an expression of your experience is powerful.

The use of backwards motion to draw was interesting as some forms of tracking would miss this. It also reflects an important part of my vision of the relationship with technology and disability, taking ownership through mis-use or novel use.

As always there is a common thread with participants about the experience of barriers. Unsurprising if we define disability as the barriers created by society. The experience acts as a connection and sharing it brings the group together. Advocacy and creative disruption naturally emerge.

Support

Commissioned and supported by Unlimited, celebrating the work of disabled artists, with funding from SouthBank Centre and Arts Council England.

Workshop space was provided by The Southbank Centre, London, UK.

Thanks to all workshop participants for their engagement and creativity.

We worked with UK based wheelchair users to explore different physical & digital ways of mark making using the motion of wheels. Each technique was introduced with work by artists who experience disability.

We were interested in the drawing experience and how it might be used by participants. The tension of the unintentional markings on the external world through use of a wheelchair. Using this as an intentional, disruptive and creative act of challenging the barriers that create disability

Physical mark making

To start physical mark making we explored the work of artist Sue Austin and her project leaving wheel trails through the city of Plymouth.

To record the wheel markings we tried paper and fabric taped to the floor. Fabric dried too quickly and paper ripped very easily. In the end the best surface was the slightly shiny and rubbery floor already present.

Mediums:

Water: Applied through spray bottles to front and rear wheels.

• Left strong marks.
• Marks faded quickly.
• Hard to photograph under normal lighting conditions.
• No way to distinguish between different users markings.

Chalk and water: Applied children’s drawing chalk to the wheels and then went through a small water puddle.

• Strong tire imprints
• Limited drawing/release time
• Supported color and distinction between users.
• Did not fade.

Two participants drawing using different colored chalk cordinating between themselves.

The contrast with the floor and the markings made it difficult to capture on camera.

Ripping trails

Powered wheelchair with wet wheels caused tire patterned rips in paper.

The riping was not intentional and initially participants were apologising for the damage. It was clear this ripping was similar to the core idea behind wheeltrails, turning accidental markings (and the tension they can create) into intentional self expression. So we encouraged participants to start ripping the paper with their wheels. It turned out to add a satisfying dimensionality to the trails while expressing some of the joy of using wheels to be creative and disruptive.

Digital mark making

We introduced Augmented Reality through the project “Unfolding Shrines” which collects a series of AR experiences by artist who experience disability: https://www.shapearts.org.uk/unfolding-shrines

To draw in augmented reality we used Weird Type by Zach Liberman using a mobile device either held or resting on the lap. Using assistive touch on the device allowing hands free motion recording. Participants were asked to choose what words they wanted to use. It was helpful to have an audio to text mode on the app to enter the words.

For video drawing we introduced Aaron Wheels who is the creator of extreme wheelchair sports and his examples of redefining what movement is possible with a wheelchair.

We used a GoPro attached to the chair with duct tape to draw with video.

Alt description: Video shows movement presented at a low ground level, around an office space, weaving between chairs and tables and doors.

The wheelchair as a creative device

One participant described how the workshop had helped them see their wheelchair in a new way and they had never thought of it as a creative tool. All participants were interested in how this could be used for advocacy, especially with the digital as this freed up the legal worries of graffiti and mark making.

One worry with digital marking was that it might feel empty if you could not always see what you are drawing. One participant described how they did not mind as the creative seeing was happening in their imagination and they enjoyed this.

The idea of community drawing and working together to draw was very strong among the participants. A way of community building with people who understand the disabled experience while being a creative and political act.

Support

Commissioned and supported by Unlimited, celebrating the work of disabled artists, with funding from SouthBank Centre and Arts Council England.

Workshop space was provided by The Studio, Bath, UK : https://www.bathspa.ac.uk/research-and-enterprise/the-studio

Thanks to all workshop participants for their engagement and creativity.

Present and absent (2009)

The artist Sue Austin part of the Freewheeling group used white marking fluid to make various pieces outdoors and indoors from the back wheels of her powered wheelchair. A paint release system is attached to the chair that feeds through tubbing to a brush which is pressing against a rear wheel.

She describes her works as:

“exploration of the mark making possibilities of my power wheelchair produced by play.”

“find a material, indexical expression of the sense of agency, freedom and expansion engendered by the experience of using a ‘power’chair”

One part of the work took place through Plymouth city centre, connecting a trail between two University buildings which where part of the exhibition.

BA Show - Freewheeling, Present and Absent (2009)

EnayBall

A 3D printed tool which clamps a ink pen head to the front of a wheelchair. Featured in the Draw to Perform international festival (https://drawtoperform.com), UK 2018. It involved a mixture of freeform drawing, connect the dot style drawing and working on a large, shared canvas.

https://www.enayball.com

Martin Vogel

Martin Vogel is a California based artist who uses a combination of wheelchair drawing and painting. He uses the front wheels combined with acyrlic paint to create abstact pieces which capture the fluid curves of wheel motion. He describes the line work as capturing motion and the color as the feeling.

http://www.martinvogel.com

Support

Commissioned and supported by Unlimited, celebrating the work of disabled artists, with funding from SouthBank Centre and Arts Council England.

Requirements

• Identity is not essential when tracking since we are working with active participants.
• Precision matters for movement but we can do post-corrections.
• Not full skeletal tracking, wheel motion tracking.
• Solutions must not be heavy or impeded wheelchair movement.
• Tracking within large indoor venues, with multiple floors, lifts, walls and doors.
• Function without complex indoor installation setup.

Solutions

Paper maps

With activate participants we can draw our journeys either during or after on a map. Journeys could also be recorded using portable cameras and used to manually reconstruct the routes and motion.

• Misses subtle movement of the wheels during journey.
• Extra requirement on participant might hinder the drawing experience?

Physical marking

Using sand, water or chalk, journeys could be marked physically indoors. Potential to use fabric or paper on the floor as a temporary surface to be marked by the motion of wheels.

• Requires permission to mark indoors
• Requires cleanup.
• Challenges with setup, application and impeding normal wheeled movement.

Phone with Gps (assisted by network)

Using global positioning satellites on a phone. Tends to be inaccurate when indoors. The satellite signal is augmented by location information determined from the phone base station identity, distance and if more than one base station triangulation.

5g introduces some new proposals on improving indoor location estimation. A faster communication speed also means more accurate location information. It’s unclear when this might be available.

• Inaccurate indoors.
• No guarantee of any signal at all.
• GPS low resolution misses the naunces of wheel movements.

Phone based IMU & Camera

Most phones contain inertia measurement units (IMU) that use gyroscopes, magnetometers and accelerometers to record a phones absolute movement. Combined with a live camera view we can perform surface and depth estimation without a special depth camera. This combined sensor data is how Augmented reality applications provide viewing and drawing.

• IMU Sensors drift over time with errors increasing.
• Surface and depth estimation is highly affected by lighting conditions
• Depth/Lidar sensors on phones are restricted to the most expensive devices.
• Positioning of AR to real world is still rooted on GPS location and has the same inaccuracy.

Bluetooth

Bluetooth low energy (BLE) beacons are distributed around the perimeter of the space and using Bluetooth help track a device such as a phone with up to 3m accuracy.

• Beacons allow 50 to 70 meters range.
• Beacons are cheap.
• Beacons run on small & long lasting battery.
• Most phones support bluetooth.
• Requires a significant number of beacons for large spaces.

Wifi

Multiple wifi routers are used by a device to measure the differential-time-of-arrival for a signal between device and router to identify location. Apple provide a closed service and app called “indoor survey” for this function.

Indoor Maps

• Requires regular surveying of existing wifi on site.
• For accuracy usually have to control the wifi routers used.
• Tends to suffer from lots of errors in calculations due to multipath signals (radio signals reaching the antenna by two or more paths).

Ultra Wideband (UWB)

Uses a high bandwidth, low power frequency to communicate. Multiple anchors are spread through a space to help identify and track a node. This is the technology used by Amazon warehouses and is being introduced by Apple with their AirTag product.

AirTag

• Provides extremely accurate (cm level) positioning
• Supported by current Iphone range 11+.
• Can penetrate thin walls.
• Expensive (1k-3k) for a dev kit.

External, more open kits are available:

https://store.pozyx.io/product/50-001-0001-creator-kit-65

Computer Vision (without markers)

Post or realtime processing of camera feed to extract movement. There is a varying level of complexity based on multiple feeds, the amount of changing noise in the scene and occlusion. With developments in machine learning there have been lots of improvements in camera tracking without special markers.

• Seems to require very controlled environments.
• High level of complexity in hardware, software and training data.

Rhizomatiks on Twitter: "Rhizomatiks developed the "Points of a sword visualization system", was presented at the "Takamadomiya Cup JAL PRESENTS FENCING WORLD CUP 2019". This was the second time the game was introduced following the 72nd All-Japan Fencing Championship which was held last month. pic.twitter.com/l8st0K27GO / Twitter"

Rhizomatiks developed the "Points of a sword visualization system", was presented at the "Takamadomiya Cup JAL PRESENTS FENCING WORLD CUP 2019". This was the second time the game was introduced following the 72nd All-Japan Fencing Championship which was held last month. pic.twitter.com/l8st0K27GO

Marker tracking

Placing special markers onto the device to be tracked. Markers can be:

• Passive — coated in a reflective material and a separate device near the camera provides light.
• Active — trackable LEDs (such as Infrared)

This is the industry standard for high precision motion capture. Used frequently in special effects and digital entertainment. There are many companies providing this technology for indoor and outdoor applications.

• Requires expensive cameras and software.
• Focused on accurate motion rather than accurate positioning.
• Often bespoke, closed source software required.
• records extremely detailed (mm precision) movement with many markers.

https://www.vicon.com

Multiple camera feeds (without markers)

Reconstruction of a scene through multiple gopro cameras attached to a person.

• Requires heavy post processing
• Accuracy is limited.
• No easily reusable software provided.
• Possibly impeded movement?
• Highly experimental.

[PDF] Motion capture from body-mounted cameras | Semantic Scholar

Further Reading

• Lecture course on motion capture — http://www.cs.cmu.edu/~yaser/Fall2012_15869.html

Thanks

Thanks to everyone who helped with the research:

• https://twitter.com/jarkman
• https://twitter.com/lyallmarcus
• https://twitter.com/tarim8
• https://twitter.com/artaggg

Support

Commissioned and supported by Unlimited, celebrating the work of disabled artists, with funding from SouthBank Centre and Arts Council England.

Disability as a social construction enforced through the design of physical spaces provides an important connection between the experience of disability and Psychogeography.

Psychogeography is the effect that when you enter a new space you immediately have an opinion about it. It’s an exploration of how geographical environments effect the emotions and behaviour of individuals. The phrase was used by Guy Debord to examined inventive ways of exploring urban spaces. He recommended the practice of drifting to explore spaces:

the practice of defamiliarisation and the choice of encounters, the sense of incompleteness and ephemerality.

The Fluxus group also explored this through playful anarchy. “Excursions & visits Dada” where tours of common places that have no reason to exist. Interestingly the tours failed to attract any visitors and only the guides where left.

Psychogeographic markup language (pml)

A markup language designed by Devenir Nomada to measure the psychological effect of strolls with routes created through algorithmic generation. Created to provide a way to display the psychogeographical landscape of a city without having to decribe the shape of the city.

Current tags include: “stim”, “dross”, “horror”, “terror”, “open”, “closed”, most of them taken from acknowledged writers in urban theory.

DEVENIR NOMADA

Spatial Relationships

1960/1970 Stanley Brouwn in “This Way Brouwn” got people to draw on a map or describe how they would get from where they are to a train station. He also captured people’s foot prints by putting pieces of paper on the sidewalk. He was interested in:

“How different people see and convey spatial relationships”.

https://www.lost-painters.nl/stedelijk-museum-schiedam-stanley-brouwn-mens-loopt-op-planeet-aarde/

Temporary Autonomous Zones (TAZ)

Hakim Bey’s idea of the creation of temporary spaces that elude formal structures of control. Located in real time and space but with a dependency on the web to bring it into being. The use of the word web does not necessarily imply computers but a web of information where what what was interesting was the “openness and horizontality of the structure”.

“Reclaim the streets” collective in London is one example of using TAZ to protest the ownership of roads and cars in our cities.

Similar projects using new forms of media for advocacy started using the term “Tactical Media”. Defined by the Members of the “Critical Art Ensemble” as

“any media that would engage a particular social political context in order to create “molecular interventions” and “semiotic shocks” that collectively could diminish the rising intensity of the authoritarian culture”.

Hacking Barriers

“Tour de fence” by Irational.org (http://www.irational.org) is a typology exploration of barriers and ways of scaling them. It followed a tour around Bristol looking at its finest fences and how to scale them.

Hacking the citys familiar physical surroundings in order to find more mentally enriching, physically challenging and social engaging ways of using them.

Tour de Fence, Bristol, United Kingdom (UK).

Protocols

Rules and methods for guiding Psychogeography.

Fixed shapes

“D’fence cuts” by Irational.org created a circle path around Bristol which when mapped to the physical route required cutting through various fences to complete. When they revisited their cut fences they found some of the routes remained, continuing to be used by people or animals.

d'fence cuts.

Universal Psychogeographic computer

The universal psychogeographic computer by Wilfred Hou Je Bek used the idea of a fork bomb from unix where a program spawns new processes which in turn spawn further new processes eventually freezing the computer due to an unmanageable number of processes. He applied this idea to people with “forkbomb .walk” where a large group forked (splitting in two) at every junction they encountered.

Ping

Kate Armstrong’s “Ping” (2003) references how in computer networking a computer identifies other computers exist by “pinging” them. The project used a dialup phone interface which based on your answers gave you physical directions. It also asked about mood and perception of ambiance’s using the phone as a prompt.

KATE ARMSTRONG

Human scale chess

Sharilyn Neidhardt in New York, 2003 performed a game of chess where humans where turned into pawns who were manipulated by phones. Players received orders via their phones from the two players who controlled the chess game. When a piece was captured it become an ordinary citizen again. They mapped blocks in NY into a 8x8 chess board with each game taking around 3 hours.

Human Scale Chess Game

Support

Commissioned and supported by Unlimited, celebrating the work of disabled artists, with funding from SouthBank Centre and Arts Council England.