AUTOMATION AND A.I. Technology & Sourcing Webinar - 8 September 2016 Giangicomo Olivi and Gareth Stokes - DLA Piper
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
AUTOMATION AND A.I.
Technology & Sourcing Webinar
Giangicomo Olivi and Gareth Stokes
8 September 2016
www.dlapiper.com 8 September 2016 0
Introductions
Giangiacomo Olivi Gareth Stokes
Partner Partner
Technology & Sourcing Group Technology & Sourcing Group
Italy UK
T: +39 02 80 618 515 T +44 (0)121 262 5831
M +39 335 53 26 994 M +44 (0)7968 559 210
giangiacomo.olivi@dlapiper.com gareth.stokes@dlapiper.com
www.dlapiper.com 8 September 2016 1
Agenda Introduction Some definitions - Science fiction, or everyday fact? Catalyst Sector-specific examples – Business Processes – Manufacturing – FinTech – Consumer web – Digital Assistants – Smart Products Legal Considerations – Contracting – Data privacy and security – Discrimination – Public liability & ethics www.dlapiper.com 8 September 2016 2
Back to Asimov's Law… 1. A robot may not injure a human being or, through inaction, allow a human being to come to harm 2. A robot must obey the orders given to it by human beings except where such order would conflict with the First Law 3. A robot must protect its own existence as long as such protection does not conflict with the First or Second laws www.dlapiper.com 8 September 2016 3
Robot or not?
No particularly widely accepted definition of "Intelligence", so "Artificial Intelligence" is
even more problematic.
Need to distinguish 'General AI' from specialist artificially intelligent systems
1996
2016
"Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a
computation.' "
Rodney Brooks, Director of MIT's Artificial Intelligence Laboratory, 2008
www.dlapiper.com 8 September 2016 4
Specialist AIs – Learning the hard(ware) way
Most current specialist AI systems are distinguished by a learning component
Often need to be 'trained' initially, or 'evolved'
Positive feedback loops mean that the system improves over time, enhancing its
ability to handle new data more accurately.
"Does this
picture have a
cat in it?"
Training
Testing Results of Testing
Initiation – no with
on new testing fed on new
data initial
data into system data
data
82%
87%
90%
50% 63% 74% 93%
www.dlapiper.com 8 September 2016 5
If you can't define, rename…
Exactly what constitutes AI can still be hotly debated
Turing Test, proposed by mathematician Alan Turing in 1950
A B
Other forms of special AI go by different names:
Virtual digital assistant
Cognitive systems Deep learning Digital life
Voice recognition
Zero-UI systems Machine learning Evolved algorithms
Virtual neuron
Autonomous systems Image recognition Self-programming systems
Neural network Whole-brain emulation
Sufficiently tricky problem to have inspired a podcast:
https://www.theincomparable.com/robot/
www.dlapiper.com 8 September 2016 6
Genuine concern for the future?
The "singularity" – a point at which General AI rapids accelerates the rate of increase in
its own intelligence, leading to Artificial Super-Intelligence (ASI)
Reckoned by serious commentators to be somewhere between 20 and 80 years away
Humans
Intelligence
Computers
1950 2016
Time
www.dlapiper.com 8 September 2016 7
Current specialist AI - Catalyst
Specialist AIs solve specific problems. E.g.:
– Voice recognition to text
– Natural language processing
– Image recognition
– Other types of pattern recognition in datasets (fraud / credit scoring / diagnostics etc.)
Allows the creation of algorithmic systems that would previously have required some
human input – as a simple example:
Voice Human typist Text can be edited, searched,
Previously recording transcribes manipulated in a computer
Brave new Voice AI speech-to- Text can be edited, searched,
world recording text engine manipulated in a computer
Result – specialist AI enables, enhances and accelerates current systems
www.dlapiper.com 8 September 2016 8
AI in use – Business Process (Outsourcing)
Many business processes involve human operators using computer systems in
reasonably predictable ways in response to a given task list:
– Helpdesk
– Claims processing
– Application processing
– Finance systems etc.
Building compatible systems to work with legacy IT can be expensive.
A machine learning AI can 'watch and learn' using hardware that records the signal sent
to a human operator's VDU, and mouse/keyboard input.
Over time, the AI's 'prediction' of what the human will do next to process a given task on
the task list will approach 100% accuracy
At a given prediction accuracy – 99% perhaps – the AI can take over tasks of that type
www.dlapiper.com 8 September 2016 9AI in use – Business Process (Outsourcing)
Pre-AI operating model
1
AI training phase - human operator undertakes the work; AI to watch, learn, predict
2
AI able to take over relevant tasks at 99%+ prediction accuracy
3
www.dlapiper.com 8 September 2016 10AI in use – Business Process (Outsourcing)
Phased reduction in headcount as tasks transition to AI provision
Labour costs form less of Total Cost of Ownership for the system; hardware and
software costs are the main components
Labour arbitrage less of a commercial driver
Re-locate services from off-shore provision to local provision
– Geopolitical risk reduction / jurisdiction and tax complexity reduction
– Easier supervision and audit
– Manage confidentiality and data protection risks better
AI becomes a potential single point of failure – the "all eggs in one basket" problem
www.dlapiper.com 8 September 2016 11AI in use - Manufacturing
Industry 4.0 – Data driven manufacturing
Agile production lines
IoT and sensors
Customisation
Digital models
Industry 4.0
AI / Cognitive systems
Data
1.0 - Mechanisation 2.0 – Assembly lines 3.0 – Robotics
www.dlapiper.com 8 September 2016 12AI in use - Digital modelling
Sensors Containers traditionally loaded on
ships according to stated weight
Weight often incorrect by +/-10% or
more
Sensors Ships therefore list and this leads
to fuel inefficiency
Adding sensors (IoT) to detect the
real weight of containers / listing of
the ship connected to an AI that
plans the loading leads to greater
flexibility and efficiency gains
~15% fuel savings
www.dlapiper.com 8 September 2016 13AI in use - FinTech
7653 4535 5544 1234 7653 4535 5544 1234
Automated Investment AI-powered fraud detection
& Wealth Management and credit checking
Transfer £100
to my Mum
Mint GoodBudget Wally+
Customer spend analysis and advice 'Zero UI' financial transactions
www.dlapiper.com 8 September 2016 14AI in use – Consumer web
Trend analysis
Recommendations
"Customers also bought"
Related searches / related images / related videos
www.dlapiper.com 8 September 2016 15AI in use – Digital Assistants www.dlapiper.com 8 September 2016 16
AI in use – Smart products
Weather
service data
Internet
GPS-enabled
smartphone
app
Internet-
connected
thermostat
Temperature
sensor
A.I. based
processing
on server
Heating Aircon
www.dlapiper.com 8 September 2016 17Legal considerations - Contracting
Current contracting models assume failure modes based on human error
– Service level models incentivise suppliers to avoid 'low grade' issues that might arise
if staff don't follow proper processes
– Liability limits and exclusions
– Confidentiality, data protection, security and audit provisions all assume human
fallibility can be avoided by oversight
AI delivered services have different failure modes, and contracts need to
reflect this
– Lower probability of 'low grade' issues that SLAs could correct
– Greater risks associated with 'catastrophic failure' = higher liability cap?
– Ownership of the 'trained' AI? Risk of 'pollution' of the AI with bad data?
– Oversight of how the AI 'mind' works is more difficult
www.dlapiper.com 8 September 2016 18Legal considerations – Contracting & HR
Normal transfer-in / transfer out TUPE model for outsourced services
Customer Customer's Customer's (or replacement
1 employees supplier's) employees
Supplier
Supplier's employees
Service start date Service end date
AI-based service provision – transfer in, gradual redundancy
Customer
Customer's What IP / knowledge does the
employees customer get?
2
Supplier
Supplier's employees
Service start date Service end date
Labour arbitrage commercial justifications diminished
Cost of redundancy spread throughout the term – how is that managed?
"Corporate memory" is held by staff (Exit management / TUPE etc.)
No staff = no TUPE on exit, but how is exit knowledge transfer managed?
www.dlapiper.com 8 September 2016 19Legal considerations – Data privacy and security
AI-based systems especially useful in data-rich environments
– Rapid processing of large volumes of data
– Pattern matching in 'Big Data' datasets
– Data from new sources – IoT sensors etc.
– New data types – photo/video/social network feeds
Much of this is 'personal data' within the meaning of Directive 95/46/EC
Possible to derive / generate additional data via processing
Geolocation data, photos, social media – where you are, what you do, who you know,
what you like and dislike, what you think?
Tightening legal framework – General Data Protection Regulation
Compliance mindset – design in compliance from the start
AI represents a huge concentration of data – hacking target
www.dlapiper.com 8 September 2016 20Legal considerations - Discrimination
EU Charter of Fundamental Rights
Equality Act 2010 protected characteristics:
– age
– disability
– gender reassignment
– marriage or civil partnership
– race
– religion or belief
– sex
– sexual orientation
Need to consider indirect discrimination as well – in many cities addresses in particular
areas, postcodes or streets may effectively 'encode' for race, religion etc. Certain job
types tend to 'encode' for gender etc.
www.dlapiper.com 8 September 2016 21Legal considerations – Liability and ethics
Who is liable for the acts of an AI?
The owner/operator? The vendor?
The manufacturer?
On what basis is liability to be
decided? Vicarious? Strict?
Switch Applying old cases to new problems.
Is case law about liability for a bolting
horse applicable to a haywire self-
driving car?
Directive 85/374/EEC on liability for
defective products, covering
A damages caused by a robot's
manufacturing defects / Strict liability
B
As AIs become more sophisticated,
The Trolley Problem – In its classic formulation, a trolley is how do they weigh and 'value' one
rolling out of control down hill toward five people on the line person against another? In extremis,
(A). If it hits them they will be killed. You can switch the could your social media profile decide
trolley onto a second line where only one person would be
whether your car 'saves' you?
killed (B). Is switching the track the correct action or not?
www.dlapiper.com 8 September 2016 22New regulations?
European Parliament - Draft report with recommendations to the Commission on
Civil Law rules on Robotics - 31 May 2016
Definition of classification of "smart robots"
– Interconnectivity and data analysis / learning through experience
– Physical support / adapting to environment
Registration of "smart robots"
Civil law liability
– Possible restrictions only for damages to property
– "Strict liability" rule
– Compulsory insurance ("producer")
Interoperability and harmonisation
Disclosure of use of robots and artificial intelligence by undertakings
www.dlapiper.com 8 September 2016 23New regulations?
"Charter on robotics" - ethical code to foster responsible innovation
– Main principles ("Beneficence" and "Non-maleficence", etc.) / Fundamental rights
Precautions / Inclusiveness / Accountability / Safety / Reversibility / Privacy / Harm
minimisation
– Research Ethics Committee (REC)
– Licence for designers
– Licence for users
Compensation fund
New legal status for robots? - "electronic personality" when they interact autonomously
with third parties / make autonomous decisions
www.dlapiper.com 8 September 2016 24Any questions?
Giangiacomo Olivi Gareth Stokes
Partner Partner
Technology & Sourcing Group Technology & Sourcing Group
Italy UK
T: +39 02 80 618 515 T +44 (0)121 262 5831
M +39 335 53 26 994 M +44 (0)7968 559 210
giangiacomo.olivi@dlapiper.com gareth.stokes@dlapiper.com
www.dlapiper.com 8 September 2016 25You can also read
