Sunday, October 25, 2020

RoboTED: a case study in Ethical Risk Assessment

A few weeks ago I gave a short paper at the excellent International Conference on Robot Ethics and Standards (ICRES 2020), outlining a case study in Ethical Risk Assessment - see our paper here. Our chosen case study is a robot teddy bear, inspired by one of my favourite movie robots: Teddy, in A. I. Artificial Intelligence.

Although Ethical Risk Assessment (ERA) is not new - it is after all what research ethics committees do - the idea of extending traditional risk assessment, as practised by safety engineers, to cover ethical risks is new. ERA is I believe one of the most powerful tools available to the responsible roboticist, and happily we already have a published standard setting out a guideline on ERA for robotics in BS 8611, published in 2016.

Before looking at the ERA, we need to summarise the specification of our fictional robot teddy bear: RoboTed. First, RoboTed is based on the following technology:

  • RoboTed is an Internet (WiFi) connected device, 
  • RoboTed has cloud-based speech recognition and conversational AI (chatbot) and local speech synthesis,
  • RoboTed’s eyes are functional cameras allowing RoboTed to recognise faces,
  • RoboTed has motorised arms and legs to provide it with limited baby-like movement and locomotion.
And second RoboTed is designed to:

  • Recognise its owner, learning their face and name and turning its face toward the child.
  • Respond to physical play such as hugs and tickles.
  • Tell stories, while allowing a child to interrupt the story to ask questions or ask for sections to be repeated.
  • Sing songs, while encouraging the child to sing along and learn the song.
  • Act as a child minder, allowing parents to both remotely listen, watch and speak via RoboTed.
The tables below summarise the ERA of RoboTED for (1) psychological, (2) privacy & transparency and (3) environmental risks. Each table has 4 columns, for the hazard, risk, level of risk (high, medium or low) and actions to mitigate the risk. BS8611 defines an ethical risk as the “probability of ethical harm occurring from the frequency and severity of exposure to a hazard”; an ethical hazard as “a potential source of ethical harm”, and an ethical harm as “anything likely to compromise psychological and/or societal and environmental well-being".

(1) Psychological Risks

(2) Security and Transparency Risks

(3) Environmental Risks

For a more detailed commentary on each of these tables see our full paper - which also, for completeness, covers physical (safety) risks.

And here are the slides from my short ICRES 2020 presentation:

Through this fictional case study we argue we have demonstrated the value of ethical risk assessment. Our RoboTed ERA has shown that attention to ethical risks can
  • suggest new functions, such as “RoboTed needs to sleep now”,
  • draw attention to how designs can be modified to mitigate some risks, 
  • highlight the need for user engagement, and
  • reject some product functionality as too risky.
But ERA is not guaranteed to expose all ethical risks. It is a subjective process which will only be successful if the risk assessment team are prepared to think both critically and creatively about the question: what could go wrong? As Shannon Vallor and her colleagues write in their excellent Ethics in Tech Practice toolkit design teams must develop the “habit of exercising the skill of moral imagination to see how an ethical failure of the project might easily happen, and to understand the preventable causes so that they can be mitigated or avoided”.

Thursday, August 20, 2020

"Why Did You Just Do That?" Explainability and Artificial Theory of Mind for Social Robots

This week I have been attending (virtually) the excellent RoboPhilosophy conference, and this morning gave a plenary talk "Why did you just do that?" Here is the abstract:
An important aspect of transparency is enabling a user to understand what a robot might do in different circumstances. An elderly person might be very unsure about robots, so it is important that her assisted living robot is helpful, predictable – never does anything that puzzles or frightens her – and above all safe. It should be easy for her to learn what the robot does and why, in different circumstances, so that she can build a mental model of her robot. An intuitive approach would be for the robot to be able to explain itself, in natural language, in response to spoken requests such as “Robot, why did you just do that?” or “Robot, what would you do if I fell down?” In this talk I will outline current work, within project RoboTIPS, to apply recent research on artificial theory of mind to the challenge of providing social robots with the ability to explain themselves. 
And here are the slides:

Here are links to the movies:

And here are the papers referenced in the talk, with links:
  1. Jobin, A., Ienca, M. & Vayena, E. (2019) The global landscape of AI ethics guidelines. Nat Mach Intell 1, 389–399
  2. Winfield, A. Ethical standards in robotics and AI. Nature Electronics 2, 46–48 (2019).  Pre-print here.
  3. Winfield, A. F. (2018) Experiments in Artificial Theory of Mind: from safety to story telling. Front. Robot. AI 5:75.
  4. Blum, C., Winfield, A. F. and Hafner, V. V. (2018) Simulation-based internal models for safer robots. Frontiers in Robotics and AI, 4 (74). pp. 1-17.
  5. Vanderelst, D. and Winfield, A. F. (2018) An architecture for ethical robots inspired by the simulation theory of cognition. Cognitive Systems Research, 48. pp. 56-66.
  6. Winfield AFT (2018) When Robots Tell Each Other Stories: The Emergence of Artificial Fiction. In: Walsh R., Stepney S. (eds) Narrating Complexity. Springer, Cham. Preprint here.
  7. Winfield, AF and Jirotka, M. (2017) The case for an ethical black box. In: Gao, Y. et al, eds. (2017) Towards Autonomous Robot Systems. LNCS 10454, pp. 262-273, Springer. Preprint here.
  8. Winfield AFT, Katie Winkle, Helena Webb, Ulrik Lyngs, Marina Jirotka and Carl Macrae, Robot Accident Investigation: a case study in Responsible Robotics, chapter submitted to RoboSoft.
and mentioned in the Q&A:
  1. Winfield, AF, K. Michael, J. Pitt and V. Evers (2019) Machine Ethics: The Design and Governance of Ethical AI and Autonomous Systems [Scanning the Issue], in Proceedings of the IEEE, vol. 107, no. 3, pp. 509-517.
  2. Vanderelst, D. and Winfield, A. (2018), The Dark Side of Ethical Robots, AIES '18: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society Dec 2018 Pages 317–322. 

Monday, August 10, 2020

Back to robot coding part 1: hello world

One of the many things I promised myself when I retired nearly two years ago was to get back to some coding. Why? Two reasons: one is that writing and debugging code is hugely satisfying - for those like me not smart enough to do pure maths or theoretical physics - it's the closest you can get to working with pure mind stuff. But the second is that I want to prototype a number of ideas in cognitive robots which tie together work in artificial theory of mind and the ethical black box, with old ideas on how robots telling each other stories and new ideas on how social robots might be able to explain themselves in response to questions like "Robot: what would you do if I forget to take my medicine?"

But before starting to work on the robot (a NAO) I first needed to learn Python, so completed most of the Codecadamy's excellent Learn Python 2 course over the last few weeks. I have to admit that I started learning Python with big misgivings over the language. I especially don't like the way Python plays fast and loose with variable types, allowing you to arbitrarily assign a thing (integer, float, string, etc) to a variable and then assign a different kind of thing to the same variable; very different to the strongly typed languages I have used since student days: Algol 60, Algol 68, Pascal and C. However, there are things I do like: the use of indentation as part of the syntax for instance, and lots of nice built in functions like range(), so x = range(0,10) puts a list ('array' in old money) of integers from 0 to 9 in x. 

So, having got my head around Python I finally made a start with the robot on Thursday last week. I didn't get far and it was *very* frustrating. 

Act 1: setting up on my Mac

Attempting to set things up on my elderly Mac air was a bad mistake which sent me spiralling down a rabbit hole of problems. The first thing you have to do is download and unzip the NAO API, called naoqi, from Aldebaran. The same web page then suggests you simply try to import naoqi from within Python, and if there are no errors all's well.

As soon as I got the export path commands right,  import naoqi resulted in the following error

Reason: unsafe use of relative rpath libboost_python.dylib in /Users/alansair/Desktop/naoqi/pynaoqi-python2.7- with restricted binary

According to stack overflow this problem is caused by Mac OSX system integrity protection (SIP)

Then (somewhat nervously) I tried turning SIP off, as instructed here.

But import naoqi still gives a different error. Perhaps its because my Python is in the wrong place, the Aldebaran page says it must be at /usr/local/bin/python (the default on the mac is /usr/bin. Ok so I So, reinstall python 2.7 from  so that it is in /usr/local/bin/python. But now I get another error message:

>> import naoqi
Fatal Python error: PyThreadState_Get: no current thread
Abort trap: 6

A quick search and I read: "this error shows up when a module tries to use a python library that is different than the one the interpreter uses, that is, when you mix two different pythons. I would run otool -L <dyld> on each of the dynamic libraries in the list of Binary Images, and see which ones is linked to the system Python."

At which point I admitted defeat.

Act 2: setting up on my Linux machine

Once I had established that the Python on my Linux machine was also the required version 2.7, I then downloaded and unzipped the NAO API, this time for Linux.

This time I was able to import naoqi with no errors, and within just a few minutes ran my first NAO program: hello world

from naoqi import ALProxy
tts = ALProxy("ALTextToSpeech", "", 9559)
tts.say("Hello, world!")

whereupon my NAO robot spoke the words "Hello world". Success!

Friday, June 05, 2020

Robot Accident Investigation

Yesterday I gave an talk at the ICRA 2020 workshop Against Robot Dystopias. The workshop should have been in Paris but - like most academic meetings during the lockdown - was held online. In the zoom chat window toward the end of the workshop many of us were wistfully imagining continued discussions in a Parisian bar over a few glasses of wine. Next year I hope. The workshop was excellent and all of the talks should be online soon.

My talk was an extended version of last year's talk for AI@Oxford What could possibly go wrong. With results from our new paper Robot Accident Investigation, the talk outlines a fictional investigation of a fictional robot accident. We had hoped to stage the mock accident, in the lab, with human volunteers and report a real investigation (of a mock accident) but the lockdown put paid to that too. So we have had to use our imagination and construct - I hope plausibly - the process and findings of the accident investigation.

Here is the abstract of our paper.
Robot accidents are inevitable. Although rare, they have been happening since assembly-line robots were first introduced in the 1960s. But a new generation of social robots are now becoming commonplace. Often with sophisticated embedded artificial intelligence (AI) social robots might be deployed as care robots to assist elderly or disabled people to live independently. Smart robot toys offer a compelling interactive play experience for children and increasingly capable autonomous vehicles (AVs) the promise of hands-free personal transport and fully autonomous taxis. Unlike industrial robots which are deployed in safety cages, social robots are designed to operate in human environments and interact closely with humans; the likelihood of robot accidents is therefore much greater for social robots than industrial robots. This paper sets out a draft framework for social robot accident investigation; a framework which proposes both the technology and processes that would allow social robot accidents to be investigated with no less rigour than we expect of air or rail accident investigations. The paper also places accident investigation within the practice of responsible robotics, and makes the case that social robotics without accident investigation would be no less irresponsible than aviation without air accident investigation.
And the slides from yesterday's talk:

Special thanks to project colleagues and co-authors: Prof Marina Jirotka, Prof Carl Macrae, Dr Helena Webb, Dr Ulrik Lyngs and Katie Winkle.

Monday, April 20, 2020

Autonomous Robot Evolution: an update

It's been over a year since my last progress report from the Autonomous Robot Evolution (ARE) project, so an update on the ARE Robot Fabricator (RoboFab) is long overdue. There have been several significant advances. First is integration of each of the elements of RoboFab. Second is the design and implementation of an assembly fixture, and third significantly improved wiring. Here is a CAD drawing of the integrated RoboFab.

The ARE RoboFab has four major subsystems: up to three 3D printer(s), an organ bank, an assembly fixture and a centrally positioned robot arm (multi-axis manipulator). The purpose of each of these subsystems is outlined as follows:
  • The 3D printers are used to print the evolved robot’s skeleton, which might be a single part, or several. With more than one 3D printer we can speed up the process by 3D printing skeletons for several different evolved robots in parallel, or – for robots with multi-part skeletons – each part can be printed in parallel.
  • The organ bank contains a set of pre-fabricated organs, organised so that the robot arm can pick organs ready for placing within the part-built robot. For more on the organs see previous blog post(s).
  • The assembly fixture is designed to hold (and if necessary rotate) the robot’s core skeleton while organs are placed and wired up.
  • The robot arm is the engine of RoboFab. Fitted with special gripper the robot arm is responsible for assembling the complete robot.
And here is the Bristol RoboFab (there is a second identical RoboFab in York):

Note that the assembly fixture is mounted upside down at the top front of the RoboFab. This has the advantage that there is a reasonable volume of clear space for assembly of the robot under the fixture, which is reachable by the robot arm.

The fabrication and assembly sequence has six stages:
  1. RoboFab receives the required coordinates of the organs and one or more mesh file(s) of the shape of the skeleton.
  2. The skeleton is 3D printed.
  3. The robot arm fetches the core ‘brain’ organ from the organ bank and clips it into the skeleton on the print bed. This is a strong locking clip.
  4. The robot arm then lifts the core organ and skeleton assemblage off the print bed, and attaches it to the assembly fixture. The core organ has metal disks on its underside which are used to secure the assemblage to the fixture with electromagnets.
  5. The robot arm then picks and places the required organs from the organ bank, clipping them into place on the skeleton.
  6. Finally the robot arm wires each organ to the core organ, to complete the robot.

Here is a complete robot, fabricated, assembled and wired by the RoboFab. This evolved robot has a total of three organs: the core ‘brain’ organ, and two wheel organs.
Note especially the wires connecting the wheel organs to the core organ. My colleague Matt has come up with an ingenious design in which a coiled cable is contained within the organ. After the organs have been attached to the skeleton (stage 5), the robot arm in turn grabs each organ's jack plug and pulls the cable to plug into the core organ (stage 6). This design minimises the previously encountered problem of the robot gripper getting tangled in dangling loose wires during stage 6.

And here is a video clip of the complete process:


The work described here has been led by my brilliant colleague Matt Hale, very ably supported by York colleagues Edgar Buchanan and Mike Angus. The only credit I can take is that I came up with some of the ideas and co-wrote the bid that secured the EPSRC funding for the project.


For a much more detailed account of the RoboFab see this paper, which was presented at ALife 2019 last summer in Newcastle: The ARE Robot Fabricator: How to (Re)produce Robots that Can Evolve in the Real World.

Related blog posts

First automated robot assembly (February 2019)
Autonomous Robot Evolution: from cradle to grave (July 2018)
Autonomous Robot Evolution: first challenges (Oct 2018)

Tuesday, September 17, 2019

What's the worst that could happen? Why we need robot/AI accident investigation.

Robots. What could possibly go wrong?

Imagine that your elderly mother, or grandmother, has an assisted living robot to help her live independently at home. The robot is capable of fetching her drinks, reminding her to take her medicine and keeping in touch with family. Then one afternoon you get a call from a neighbour who has called round and sees your grandmother collapsed on the floor. When the paramedics arrive they find the robot wandering around apparently aimlessly. One of its functions is to call for help if your grandmother stops moving, but it seems that the robot failed to do this. 

Fortunately your grandmother recovers but the doctors find bruising on her legs, consistent with the robot running into them. Not surprisingly you want to know what happened: did the robot cause the accident? Or maybe it didn't but made matters worse, and why did it fail to raise the alarm? 

Although this is a fictional scenario it could happen today. If it did you would be totally reliant on the goodwill of the robot manufacturer to discover what went wrong. Even then you might not get the answers you seek; it's entirely possible the robot and the company that made it are just not equipped with the tools and processes to undertake an investigation.

Right now there are no established processes for robot accident investigation. 

Of course accidents happen, and that just as true for robots as any other machinery [1].

Finding statistics is tough. But this web page shows serious accidents with industrial robots in the US since the mid 1980s. Driverless car fatalities of course make the headlines. There have been five (that we know about) since 2016. But we have next to no data on accidents in human robot interaction (HRI); that is for robots designed to interact directly with humans. Here is one - a security robot - that happened to be reported.

But a Responsible Roboticist must be interested in *all* accidents, whether serious or not. We should also be very interested in near misses; these are taken *very* seriously in aviation [2], and there is good evidence that reporting near misses improves safety.

So I am very excited to introduce our 5-year EPSRC funded project RoboTIPS – responsible robots for the digital economy. Led by Professor Marina Jirotka at the University of Oxford, we believe RoboTIPS to be the first project with the aim of systematically studying the question of how to investigate accidents with social robots.

So what are we doing in RoboTIPS..?

First we will look at the technology needed to support accident investigation.

In a paper published 2 years ago Marina and I argued the case for an Ethical Black Box (EBB) [3]. Our proposition is very simple: that all robots (and some AIs) should be equipped by law with a standard device which continuously records a time stamped log of the internal state of the system, key decisions, and sampled input or sensor data (in effect the robot equivalent of an aircraft flight data recorder). Without such a device finding out what the robot was doing, and why, in the moments leading up to an accident is more or less impossible. In RoboTIPS we will be developing and testing a model EBB for social robots.

But accident investigation is a human process of discovery and reconstruction. So in this project we will be designing and running three staged (mock) accidents, each covering a different application domain: 
  • assisted living robots, 
  • educational (toy) robots, and 
  • driverless cars.
In these scenarios we will be using real robots and will be seeking human volunteers to act in three roles, as the 
  • subject(s) of the accident, 
  • witnesses to the accident, and as 
  • members of the accident investigation team.
Thus we aim to develop and demonstrate both technologies and processes (and ultimately policy recommendations) for robot accident investigation. And the whole project will be conducted within the framework of Responsible Research and Innovation; it will, in effect, be a case study in Responsible Robotics.

The text above is the script for a very short (10 minute) TED-style talk I gave at the conference AI@Oxford today in the Impact of Trust in AI session, and here below are the slides.


[1] Dhillon BS (1991) Robot Accidents. In: Robot Reliability and Safety. Springer, New York, NY
[2] Macrae C (2014) Close Calls: Managing risk and resilience in Airline flight safety, Palgrave macmillan.
[3] Winfield AFT and Jirotka M (2017) The Case for an Ethical Black Box. In: Gao Y, Fallah S, Jin Y, Lekakou C (eds) Towards Autonomous Robotic Systems. TAROS 2017. Lecture Notes in Computer Science, vol 10454. Springer, Cham.

Wednesday, July 31, 2019

On the simulation (and energy costs) of human intelligence, the singularity and simulationism

For many researchers the Holy Grail of robotics and AI is the creation of artificial persons: artefacts with equivalent general competencies as humans. Such artefacts would literally be simulations of humans. Some researchers are motivated by the utility of AGI; others have an almost religious faith in the transhumanist promise of the technological singularity. Others, like myself, are driven only by scientific curiosity. Simulations of intelligence provide us with working models of (elements of) natural intelligence. As Richard Feynman famously said ‘What I cannot create, I do not understand’. Used in this way simulations are like microscopes for the study of intelligence; they are scientific instruments.

Like all scientific instruments simulation needs to be used with great care; simulations need to be calibrated, validated and – most importantly – their limitations understood. Without that understanding any claims to new insights into the nature of intelligence – or for the quality and fidelity of an artificial intelligence as a model of some aspect of natural intelligence – should be regarded with suspicion.

In this essay I have critically reflected on some of the predictions for human-equivalent AI (AGI); the paths to AGI (and especially via artificial evolution); the technological singularity, and the idea that we are ourselves simulations in a simulated universe (simulationism). The quest for human-equivalent AI clearly faces many challenges. One (perhaps stating the obvious) is that it is a very hard problem. Another, as I have argued in this essay, is that the energy costs are likely to limit progress.

However, I believe that the task is made even more difficult for two further reasons. The first is – as hinted above – that we have failed to recognize simulations of intelligence (which all AIs and robots are) as scientific instruments, which need to be designed, operated and results interpreted, with no less care than we would a particle collider or the Hubble telescope.

The second, and more general observation, is that we lack a general (mathematical) theory of intelligence. This lack of theory means that a significant proportion of AI research is not hypothesis  driven, but incrementalist and ad-hoc. Of course such an approach can and is leading to interesting  and (commercially) valuable advances in narrow AI. But without strong theoretical foundations, the grand challenge of human-equivalent AI seems rather like trying to build particle accelerators to understand the nature of matter, without the Standard Model of particle physics.

The text above is the concluding discussion of my essay On the simulation (and energy costs) of human intelligence, the singularity and simulationism, which appears in an edited collection of essays in a book called From Astrophysics to Unconventional Computation. Published in April 2019, the book marks the 60th birthday of astrophysicist, computer scientist and all round genius, Susan Stepney.

Note: regular visitors to the blog will recognise themes covered in several previous blog posts, brought together in I hope a coherent and interesting way.

Monday, July 29, 2019

Ethical Standards in Robotics and AI: what they are and why they matter

Here are the slides for my keynote, presented this morning at the International Conference on Robot Ethics and Standards (ICRES 2019). The talk is based on my paper Ethical Standards in Robotics and AI published in Nature Electronics a few months ago (here is a pre-print).

To see the speaker notes click on the options button on the google slides toolbar above.

Friday, June 28, 2019

Energy and Exploitation: AIs dirty secrets

A couple of days ago I gave a short 15 minute talk at an excellent 5x15 event in Bristol. The talk I actually gave was different to the one I'd originally suggested. Two things prompted the switch: one was seeing the amazing line up of speakers on the programme - all covering more or less controversial topics - and the other was my increasing anger in recent months over the energy and human costs of AI. So it was that I wrote a completely new talk the day before this event.

But before I get to my talk I must mention the amazing other speakers: we heard Phillipa Perry speaking on child parent relationships, Hallie Rubenhold on the truth about Jack the Ripper's victims,  Jenny Riley speaking very movingly about One25's support for Bristol's (often homeless) sex workers, and Amy Sinclair introducing her activism with Extinction Rebellion.

Here is the script for my talk (for the slides go to the end of this blog post).

Artificial Intelligence and Machine Learning are often presented as bright clean new technologies with the potential to solve many of humanity's most pressing problems.

We already enjoy the benefit of truly remarkable AI technology, like machine translation and smart maps. Driverless cars might help us get around before too long, and DeepMind's diagnostic AI can detect eye diseases from retinal scans as accurately as a doctor.

Before getting into the ethics of AI I need to give you a quick tutorial on machine learning. The most powerful and exciting AI today is based on Artificial Neural Networks. Here [slide 3] is a simplified diagram of a Deep Learning network for recognizing images. Each small circle is a *very* simplified mathematical model of biological neurons, and the outputs of each layer of artificial neurons feed the inputs of the next layer. In order to be able to recognise images the network must first be trained with images that are already labelled - in this case my dog Lola.

But in order to reliably recognise Lola the network needs to be trained not with one picture of Lola but many. This set of images is called the training data set and without a good data set the network will not work at all or will be biased. (In reality there will need to be not 4 but hundreds of images of Lola).

So what does an AI ethicist do? Well, the short answer is worry. I worry about the ethical and societal impact of AI on individuals, society and the environment. Here are some keywords on ethics [slide 4], reflecting that we must work toward AI that respects Human Rights, diversity and dignity, is unbiased and sustainable, transparent, accountable and socially responsible.

But I do more than just worry. I also take practical steps like drafting ethical principles, and helping to write ethical standards for the British Standards Institute and the IEEE Standards Association. I lead P7001: a new standard on transparency in of autonomous systems based on the simple ethical principle that it should always be possible to find out why an AI made a particular decision. I have given evidence in parliament several times, and recently took part in a study of AI and robotics in healthcare and what this means for the workforce of the NHS.

Now I want to share two serious new worries with you.

The first is about the energy cost of AI. In 2016 Go champion Lee Sedol was famously defeated by DeepMind's AlphaGo. It was a remarkable achievement for AI. But consider the energy cost. In a single two hour match Sedol burned around 170 kcals: roughly the amount of energy you would get from an egg sandwich.  Or about the power of an LED night light -  1 Watt. In the same two hours the AlphaGo machine reportedly consumed 50,000 times more energy than Sedol. Equivalent to a 50 kW generator for industrial lighting. And that's not taking account of the energy used to train AlphaGo.

Now some people think we can make human equivalent AI by simulating the human brain. But the most complex animal brain so far simulated is that of c-elegans – the nematode worm. It has 302 neurons and about 5000 synapses - these are the connections between neurons. A couple of years ago I worked out that simulating a neural network for a simple robot with only a 10th the number of neurons of c-elegans costs 2000 times more energy than the whole worm.

In a new paper that came out just a few days ago we have for the first time estimates of the carbon cost of training large AI models for natural language processing such as machine translation [1]. The carbon cost of simple models is quite modest, but with tuning and experimentation the carbon cost leaps to 7 times the carbon footprint of an average human in one year (or 2 times if you're an American).

And the energy cost of optimising the biggest model is a staggering 5 times the carbon cost of a car over its whole lifetime, including manufacturing it in the first place. The dollar cost of that amount of energy is estimated at between one and 3 million US$. (Something that only companies with very deep pockets can afford.)

These energy costs seem completely at odds with the urgent need to halve carbon dioxide emissions by 2030. At the very least AI companies need to be honest about the huge energy costs of machine learning.

Now I want to turn to the human cost of AI. It is often said that one of the biggest fears around AI is the loss of jobs. In fact the opposite is happening. Many new jobs are being created, but the tragedy is that they are not great jobs, to say the least. Let me introduce you to two of these new kinds of jobs.

The first is AI tagging. This is manually labelling objects in images to, for instance, generate training data sets for driverless car AIs. Better (and safer) AI needs huge training data sets and a whole new outsourced industry has sprung up all over the world to meet this need. Here [slide 9] is an AI tagging factory in China.

Conversational AI or chat bots also need human help. Amazon for instance employs thousands of both full-time employees and contract workers to listen to and annotate speech. The tagged speech is then fed back to Alexa to improve its comprehension. And last month the Guardian reported that Google employs around 100,000 temps, vendors and contractors: literally an army of linguists working in "white collar sweatshops" to create the handcrafted data sets required for Google translate to learn dozens of languages. Not surprisingly there is a huge disparity between the wages and working conditions of these workers and Google's full time employees.

AI tagging jobs are dull, repetitive and in the case of the linguists highly skilled. But by far the worst kind of new white collar job in the AI industry is content moderation.

These tens of thousands of people, employed by third-party contractors, are required to watch and vet offensive content: hate speech, violent pornography, cruelty and sometimes murder of both animals and humans for Facebook, YouTube and other media platforms [2]. These jobs are not just dull and repetitive they are positively dangerous. Harrowing reports tell of PTSD-like trauma symptoms, panic attacks and burnout after one year, alongside micromanagement, poor working conditions and ineffective counselling. And very poor pay - typically $28,800 a year. Compare this with average annual salaries at Facebook of ~$240,000.

The big revelation to me over the past few months is the extent to which AI has a human supply chain, and I am an AI insider! The genius designers of this amazing tech rely on both huge amounts of energy and a hidden army of what Mary Gray and Siddhartha Suri call Ghost Workers.

I would like to leave you with a question: how can we, as ethical consumers, justify continuing to make use of unsustainable and unethical AI technologies?


[1] Emma Strubell, Ananya Ganesh, Andrew McCallum (2019) Energy and Policy Considerations for Deep Learning in NLP, arXiv:1906.02243
[2] Sarah Roberts (2016) Digital Refuse: Canadian Garbage, Commercial Content Moderation and the Global Circulation of Social Media’s Waste, Media Studies Publications. 14.

Thursday, May 30, 2019

My top three policy and governance issues in AI/ML

In preparation for a meeting of the WEF global AI council today, we were asked the question:

What do you think are the top three policy and governance issues that face AI/ML currently? 

Here are my answers.

1.     For me the biggest governance issue facing AI/ML ethics is the gap between principles and practice. The hard problem the industry faces is turning good intentions into demonstrably good behaviour. In the last 2.5 years there has been a gold rush of new ethical principles in AI. Since Jan 2017 at least 22 sets of ethical principles have been published, including principles from Google, IBM, Microsoft and Intel. Yet any evidence that these principles are making a difference within those companies is hard to find – leading to a justifiable accusation of ethics-washing – and if anything the reputations of some leading AI companies are looking increasingly tarnished.

2.     Like others I am deeply concerned by the acute gender imbalance in AI (estimates of the proportion of women in AI vary between ~12% and ~22%). This is not just unfair, I believe it too be positively dangerous, since it is resulting in AI products and services that reflect the values and ambitions of (young, predominantly white) men. This makes it a governance issue. I cannot help wondering if the deeply troubling rise of surveillance capitalism is not, at least in part, a consequence of male values.

3.     A major policy concern is the apparently very poor quality of many of the jobs created by the large AI/ML companies. Of course the AI/ML engineers are paid exceptionally well, but it seems that there is a very large number of very poorly paid workers who, in effect, compensate for the fact that AI is not (yet) capable of identifying offensive content, nor is it able to learn without training data generated from large quantities of manually tagged objects in images, nor can conversational AI manage all queries that might be presented to it. This hidden army of piece workers, employed in developing countries by third party sub contractors and paid very poorly, are undertaking work that is at best extremely tedious (you might say robotic) and at worst psychologically very harmful; this has been called AI’s dirty little secret and should not – in my view – go unaddressed.