The role of standards and evidence
Professor Chris Marsden,
Monash Digital Law Group
29 November 2023
AI Co-regulation –
Lessons from the EU?
Victorian Heart Institute Innovation Conference
Associate Director,
Monash Data Futures Institute
Director, Monash Digital Law Group
Professor of AI,
Technology and the Law
Visiting Professor UKRI TAS Regulation & Governance Hub
A lesson from Babylon’s Tower of Babel
Don’t believe the hype!
AI chatbots
1927
REGULATORY
CONCERNS
ABOUT AI
CHATBOTS
CO-REGULATION OF AI
UNDER EU LAW
• Some medical AI/chatbot disasters
• EU proposal for a Regulation on AI, known as the ‘AI Act’,
• Use of co-regulation of standards in the EU AI Act
• At the end of the talk, Chris will offer some (very premature)
observations on what Australia may learn from the process.
Marsden and Tambini
“There is a danger that some aspects of internet self-
regulation fail to conform to accepted standards.
We recommend co-regulatory audit as the best balance of
fundamental rights and responsive regulation.”
Directorate-General for Communications Networks, Content and Technology (European Commission) , Programme
in Comparative Law and Policy (2004) Self-Regulation of Digital Media Converging on the Internet: Industry Codes of
Conduct in Sectoral Analysis, Final Report of IAPCODE Project for European Commission DG Information Society
Safer Internet Action Plan, 30 April, Section 12.7 at https://publications.europa.eu/en/publication-detail/-
/publication/b7c998d9-75d6-464d-9d91-d59aa90a543c/language-en
2004 EC selfregulation.info
We still collaborate – see Regulating Big Tech (OUP, 2021)
The Story of a Morally Bankrupt Unicorn & a Medical AI Disaster
Babylon – the $4billion disaster
Dr David Watkins: Babylon Health (2013 - 2023)
• Fraudulent claims
• Dangerously flawed technology
• Regulatory failings
• 21 October 2021, Babylon valued at $4.2B
• May 2023, Babylon bankrupt.
BABYLON HEALTH AI?
 Why did Babylon fail? Babylon business model was always based on lies.
 “There’s an element of ‘fake it until you make it’ in Silicon Valley”
 Venture Capitalist who didn't invest in @babylonhealth.
 “I don’t think I’ve ever had a meeting when he has not said something provably untrue”
 Blitzscaling Babylon: Ali Parsa’s break-neck scramble for success (thetimes.co.uk)
MEDICINES & HEALTHCARE PRODUCTS
REGULATORY AGENCY: FAILED
1. Babylon was an unethical corporation that knowingly put patients at risk of harm.
2. UK regulators ignored serious corporate failings in 2017-18.
3. Babylon business model was based on AI tech that didn’t exist.
 UK Health Secretary Matt Hancock supported it all the way
 Despite existing safety concerns of regulators
 Care Quality Commission & MHRA
 May 2018 Babylon ‘upgraded’ their AI Chatbot
 managed to make it even more dangerous
BABYLON AI “CAPABLE OF AUTOMATING DIAGNOSIS &
CONSULTATION TRANSCRIPTION WITH COST SAVING OF 70%”
Ali Parsa made AI technology claims including;
- Alexa Integration
- AI Emotion recognition
- AI Voice transcription
- AI Diagnosis
all AI vapourware
THEIR NEW AI CHATBOT CLAIMED TO BE AS ACCURATE AS A
DOCTOR, WHICH WAS A REAL PAIN IN THE CHEST
The patient was a 67 yr old obese 20/day male smoker –
with typical symptoms of a #HeartAttack.
This negligent advice was a known flaw in #AskBabylon #Chatbot
@MHRAdevices formally raised it as incident with them in May 2018
CQC CLAIMED THEY COULD
NOT STOP AI CHATBOT HYPE
 Complaints against individual unethical doctors go
through General Medical Council
 Not renowned for its lightning speed
 Medical Act 1858 charity co-regulator
 Council for Healthcare Regulatory Excellence
December 2011 paper 'right-touch regulation'
 Advertising ‘Standards’ (sic) Authority is a self
regulator for ad sales
HERE’S WHAT HAPPENED FROM ITS 2021
SHARE SALE ONWARDS
THERANOS: SMARTEST WOMAN
IN THE ROOM
 Estafadora, absolutemente!
 Stanford dropout 2004 as with later AI luminaries @FTX & OpenAI
 Faked a low voice and Steve Jobs look from 2007
 2014: Theranos valued at $9 billion; raised $400 million
 By the end of 2014, her name on 18 U.S. & 66 foreign patents
 2015 appointed to Harvard Medical School Board of Fellows
 2016 –regulator Centers for Medicare/Medicaid Services intervened
 Conviction after case 2018-22: Wire fraud (3 counts) Conspiracy to
commit wire fraud (1 count). Serving 135 months) in prison
 Holmes’ father, Christian Rasmus Holmes IV, VP at Enron,
 bankrupt energy company after a giant accounting fraud
2023: OPENAI DEVELOPS
CHATGPT –CLAIMS IT CAN
REPLACE DOCTORS
 Babylon 5?
 Save us from more idiot tech bros
 Yes, it can help doctors write notes…
UK NHS USE OF PATIENT DATA FOR AI
• Ben Goldacre Review (2022)
• https://www.gov.uk/government/publications/better-broader-safer-using-health-
data-for-research-and-analysis
• Health Secretary Hancock in 2019 – AI will save NHS
• Replacing doctors’ unions…
• PM advisor Cummings & Palantir
• Faculty AI (start-up) and COVID tracking app
• UK app too centralised – spyware & cybersecurity trap
• rejected by Google & Apple on privacy grounds
$BILLION PALANTIR NHS
FEDERATED DATA CONTRACT
• Agreed this week
• Isaac Levito – NSW lobbyist extraordinaire
TRUST & VERIFY: BIG DATA CAN HELP
BUT…
EU AI ACT: MY VIEW WITH GEORGE CHRISTOU
CO-REGULATION INVOLVES DELEGATING THE
CREATION OF STANDARDS TO COMPLY WITH
LEGISLATION TO STANDARD-SETTING
ORGANIZATIONS
• Can be technical, normative or both
• (STS scholars will insist all are both)
• Note formal requirement in legislation –
• this is NOT self-regulation
• Potential advantages:
• flexibility and adaptability to rapidly-changing technology,
• as well as opportunities for expert input and industry participation.
• Risks: conflicts of interest, lack of transparency/accountability,
• potential for under- or over-regulation.
Is this a rise from self-regulation to ethics washing acceptance?
Enhanced and Meta Regulation
Co-regulation vs enhanced?
Meta regulation
• Regulated self-regulation
• Beaufort Scale of Co-regulation (Marsden 2011)
• Grabowsky (2016) identifies 1983 as the first use of ‘meta regulation’
• ch09.pdf (anu.edu.au)
IS THIS AN ACT WORTH EMULATING?
• Neither Artificial –
• definition may be too broad for AI
• Nor Intelligent
• Devolves to existing standards bodies
• Nor an Act
• Still 3 years from enforcement even after Trilogue
• which may end in September for a plenary vote
• So 2027 enforcement? AI Board proposal?
Council of Europe
Strasbourg effect?
As with data protection – Convention 108+
‘
• CAI - Committee on Artificial Intelligence - Artificial
Intelligence (coe.int)
• Convention on Regulation of Artificial Intelligence
• Agreed November 2023?
• Treats human rights as integral to regulation – unlike EU
• EU negotiating as a party:
• https://futurium.ec.europa.eu/en/european-ai-
alliance/blog/eu-participates-actively-council-europe-
negotiations-development-new-convention-artificial
‘
AUSTRALIA MAIN LESSON: REPENT AT
LEISURE
#BrusselsEffect has been #PortarlingtonEffect (Ireland not VIC)
• EU laws are unenforceable which suits tech companies
• Not only GAFAM from US/Ireland/Lux – also SAP and other EU based
• But undermines trust in tech in society more generally
MINISTER HUSIC SPOKE AT UTS NEW
INSTITUTE LAUNCH OCTOBER 2022
• “I want Australia to become the world leader in responsible AI.
• This includes setting reasonable regulations and standards...
• providing citizens with confidence that technologies like AI are being developed
in ways that's trusted, secure and to their benefit.
• Our existing legal frameworks need to be fit for purpose”
So he wants to be judged on comparative regulation
• Suggests co-regulation with standards – perhaps incorporating CSIRO (and
other) local audit standards
• Husic, Ed (2022) Remarks of the Minister at Launch of the Human Technologies Institute, UTS, 20
October, at https://uts.edu.au/sites/default/files/2022-
10/Transcript%20Ed%20Husic%20UTS%20Human%20Technologies%20Institute%20launch.pdf
WHAT WOULD BE BETTER AI
REGULATION?
• UK and Australia using sectoral regulation and sand boxes?
• UK regulatory strategy for trustworthy AI, 29 March 2023, relies on national
sectoral regulators and the Competition Markets Authority: AI regulation: a
pro-innovation approach - GOV.UK (www.gov.uk)
• Australia released its own National AI Strategy, in June 2021
https://wp.oecd.ai/app/uploads/2021/12/Australia_AI_Action_Plan_2021.pd
f
• Petition launched, supported by more than 1,000 AI luminaries, that called
for a six month moratorium on the training of “AI systems more powerful
than GPT-4” (https://futureoflife.org/, March 2023).
• United Nations: 193 member states agreed to a Global Agreement on
Ethics of AI 25 November 2021
• Big summits this summer – AIforGood, Global Digital Compact
Evidence base for AI regulation.pptx

Evidence base for AI regulation.pptx

  • 1.
    The role ofstandards and evidence Professor Chris Marsden, Monash Digital Law Group 29 November 2023 AI Co-regulation – Lessons from the EU? Victorian Heart Institute Innovation Conference
  • 2.
    Associate Director, Monash DataFutures Institute Director, Monash Digital Law Group Professor of AI, Technology and the Law Visiting Professor UKRI TAS Regulation & Governance Hub
  • 3.
    A lesson fromBabylon’s Tower of Babel Don’t believe the hype! AI chatbots
  • 4.
  • 5.
    CO-REGULATION OF AI UNDEREU LAW • Some medical AI/chatbot disasters • EU proposal for a Regulation on AI, known as the ‘AI Act’, • Use of co-regulation of standards in the EU AI Act • At the end of the talk, Chris will offer some (very premature) observations on what Australia may learn from the process.
  • 6.
    Marsden and Tambini “Thereis a danger that some aspects of internet self- regulation fail to conform to accepted standards. We recommend co-regulatory audit as the best balance of fundamental rights and responsive regulation.” Directorate-General for Communications Networks, Content and Technology (European Commission) , Programme in Comparative Law and Policy (2004) Self-Regulation of Digital Media Converging on the Internet: Industry Codes of Conduct in Sectoral Analysis, Final Report of IAPCODE Project for European Commission DG Information Society Safer Internet Action Plan, 30 April, Section 12.7 at https://publications.europa.eu/en/publication-detail/- /publication/b7c998d9-75d6-464d-9d91-d59aa90a543c/language-en 2004 EC selfregulation.info We still collaborate – see Regulating Big Tech (OUP, 2021)
  • 7.
    The Story ofa Morally Bankrupt Unicorn & a Medical AI Disaster Babylon – the $4billion disaster Dr David Watkins: Babylon Health (2013 - 2023) • Fraudulent claims • Dangerously flawed technology • Regulatory failings • 21 October 2021, Babylon valued at $4.2B • May 2023, Babylon bankrupt.
  • 8.
    BABYLON HEALTH AI? Why did Babylon fail? Babylon business model was always based on lies.  “There’s an element of ‘fake it until you make it’ in Silicon Valley”  Venture Capitalist who didn't invest in @babylonhealth.  “I don’t think I’ve ever had a meeting when he has not said something provably untrue”  Blitzscaling Babylon: Ali Parsa’s break-neck scramble for success (thetimes.co.uk)
  • 9.
    MEDICINES & HEALTHCAREPRODUCTS REGULATORY AGENCY: FAILED 1. Babylon was an unethical corporation that knowingly put patients at risk of harm. 2. UK regulators ignored serious corporate failings in 2017-18. 3. Babylon business model was based on AI tech that didn’t exist.  UK Health Secretary Matt Hancock supported it all the way  Despite existing safety concerns of regulators  Care Quality Commission & MHRA  May 2018 Babylon ‘upgraded’ their AI Chatbot  managed to make it even more dangerous
  • 10.
    BABYLON AI “CAPABLEOF AUTOMATING DIAGNOSIS & CONSULTATION TRANSCRIPTION WITH COST SAVING OF 70%” Ali Parsa made AI technology claims including; - Alexa Integration - AI Emotion recognition - AI Voice transcription - AI Diagnosis all AI vapourware
  • 11.
    THEIR NEW AICHATBOT CLAIMED TO BE AS ACCURATE AS A DOCTOR, WHICH WAS A REAL PAIN IN THE CHEST The patient was a 67 yr old obese 20/day male smoker – with typical symptoms of a #HeartAttack. This negligent advice was a known flaw in #AskBabylon #Chatbot @MHRAdevices formally raised it as incident with them in May 2018
  • 13.
    CQC CLAIMED THEYCOULD NOT STOP AI CHATBOT HYPE  Complaints against individual unethical doctors go through General Medical Council  Not renowned for its lightning speed  Medical Act 1858 charity co-regulator  Council for Healthcare Regulatory Excellence December 2011 paper 'right-touch regulation'  Advertising ‘Standards’ (sic) Authority is a self regulator for ad sales
  • 14.
    HERE’S WHAT HAPPENEDFROM ITS 2021 SHARE SALE ONWARDS
  • 15.
    THERANOS: SMARTEST WOMAN INTHE ROOM  Estafadora, absolutemente!  Stanford dropout 2004 as with later AI luminaries @FTX & OpenAI  Faked a low voice and Steve Jobs look from 2007  2014: Theranos valued at $9 billion; raised $400 million  By the end of 2014, her name on 18 U.S. & 66 foreign patents  2015 appointed to Harvard Medical School Board of Fellows  2016 –regulator Centers for Medicare/Medicaid Services intervened  Conviction after case 2018-22: Wire fraud (3 counts) Conspiracy to commit wire fraud (1 count). Serving 135 months) in prison  Holmes’ father, Christian Rasmus Holmes IV, VP at Enron,  bankrupt energy company after a giant accounting fraud
  • 16.
    2023: OPENAI DEVELOPS CHATGPT–CLAIMS IT CAN REPLACE DOCTORS  Babylon 5?  Save us from more idiot tech bros  Yes, it can help doctors write notes…
  • 18.
    UK NHS USEOF PATIENT DATA FOR AI • Ben Goldacre Review (2022) • https://www.gov.uk/government/publications/better-broader-safer-using-health- data-for-research-and-analysis • Health Secretary Hancock in 2019 – AI will save NHS • Replacing doctors’ unions… • PM advisor Cummings & Palantir • Faculty AI (start-up) and COVID tracking app • UK app too centralised – spyware & cybersecurity trap • rejected by Google & Apple on privacy grounds
  • 19.
    $BILLION PALANTIR NHS FEDERATEDDATA CONTRACT • Agreed this week • Isaac Levito – NSW lobbyist extraordinaire
  • 20.
    TRUST & VERIFY:BIG DATA CAN HELP BUT…
  • 21.
    EU AI ACT:MY VIEW WITH GEORGE CHRISTOU
  • 22.
    CO-REGULATION INVOLVES DELEGATINGTHE CREATION OF STANDARDS TO COMPLY WITH LEGISLATION TO STANDARD-SETTING ORGANIZATIONS • Can be technical, normative or both • (STS scholars will insist all are both) • Note formal requirement in legislation – • this is NOT self-regulation • Potential advantages: • flexibility and adaptability to rapidly-changing technology, • as well as opportunities for expert input and industry participation. • Risks: conflicts of interest, lack of transparency/accountability, • potential for under- or over-regulation.
  • 23.
    Is this arise from self-regulation to ethics washing acceptance? Enhanced and Meta Regulation Co-regulation vs enhanced? Meta regulation • Regulated self-regulation • Beaufort Scale of Co-regulation (Marsden 2011) • Grabowsky (2016) identifies 1983 as the first use of ‘meta regulation’ • ch09.pdf (anu.edu.au)
  • 24.
    IS THIS ANACT WORTH EMULATING? • Neither Artificial – • definition may be too broad for AI • Nor Intelligent • Devolves to existing standards bodies • Nor an Act • Still 3 years from enforcement even after Trilogue • which may end in September for a plenary vote • So 2027 enforcement? AI Board proposal?
  • 25.
    Council of Europe Strasbourgeffect? As with data protection – Convention 108+ ‘ • CAI - Committee on Artificial Intelligence - Artificial Intelligence (coe.int) • Convention on Regulation of Artificial Intelligence • Agreed November 2023? • Treats human rights as integral to regulation – unlike EU • EU negotiating as a party: • https://futurium.ec.europa.eu/en/european-ai- alliance/blog/eu-participates-actively-council-europe- negotiations-development-new-convention-artificial ‘
  • 26.
    AUSTRALIA MAIN LESSON:REPENT AT LEISURE #BrusselsEffect has been #PortarlingtonEffect (Ireland not VIC) • EU laws are unenforceable which suits tech companies • Not only GAFAM from US/Ireland/Lux – also SAP and other EU based • But undermines trust in tech in society more generally
  • 27.
    MINISTER HUSIC SPOKEAT UTS NEW INSTITUTE LAUNCH OCTOBER 2022 • “I want Australia to become the world leader in responsible AI. • This includes setting reasonable regulations and standards... • providing citizens with confidence that technologies like AI are being developed in ways that's trusted, secure and to their benefit. • Our existing legal frameworks need to be fit for purpose” So he wants to be judged on comparative regulation • Suggests co-regulation with standards – perhaps incorporating CSIRO (and other) local audit standards • Husic, Ed (2022) Remarks of the Minister at Launch of the Human Technologies Institute, UTS, 20 October, at https://uts.edu.au/sites/default/files/2022- 10/Transcript%20Ed%20Husic%20UTS%20Human%20Technologies%20Institute%20launch.pdf
  • 28.
    WHAT WOULD BEBETTER AI REGULATION? • UK and Australia using sectoral regulation and sand boxes? • UK regulatory strategy for trustworthy AI, 29 March 2023, relies on national sectoral regulators and the Competition Markets Authority: AI regulation: a pro-innovation approach - GOV.UK (www.gov.uk) • Australia released its own National AI Strategy, in June 2021 https://wp.oecd.ai/app/uploads/2021/12/Australia_AI_Action_Plan_2021.pd f • Petition launched, supported by more than 1,000 AI luminaries, that called for a six month moratorium on the training of “AI systems more powerful than GPT-4” (https://futureoflife.org/, March 2023). • United Nations: 193 member states agreed to a Global Agreement on Ethics of AI 25 November 2021 • Big summits this summer – AIforGood, Global Digital Compact