![[Graphic 2]](https://image.slidesharecdn.com/notredamepresentationsep2018-aiethicsfull-181005005528/85/Managing-the-Risks-of-AI-A-Planning-Guide-for-Executives-14-320.jpg)
![[Graphic 3]](https://image.slidesharecdn.com/notredamepresentationsep2018-aiethicsfull-181005005528/85/Managing-the-Risks-of-AI-A-Planning-Guide-for-Executives-18-320.jpg)
![[Graphic 5]](https://image.slidesharecdn.com/notredamepresentationsep2018-aiethicsfull-181005005528/85/Managing-the-Risks-of-AI-A-Planning-Guide-for-Executives-24-320.jpg)
Managing the Risks of AI - A Planning Guide for Executives
- 1. Managing the Risks of AI - A Planning Guide for Executives With Daniel Faggella CEO at TechEmergence (Emerj)
- 2. Overview • Framing AI and ethics in the context of business strategy and business value • Highlight a specific ethical issue (transparency, accountability, job loss, etc) – Provide a simple framework for beginning to tackle this ethical issue, examples included • End
- 4. Promises • By the end of this talk you’ll be able to: – (a) Determine what ethical concerns are relevant for your business – (b) Work though many of the potentially relevant ethical concerns with simple frameworks • My job is to train your antennae to pick up on what matters, and feel just fine ignoring what doesn’t – and you’ll have a better sense of what to pay attention to by the end of this talk
- 5. About TechEmergence • Largest audience of AI-focused executives on earth: – 250,000 AI-focused business readers per month, large email list and podcast listenership – We cover the most important sectors of AI in industry (banking, pharma, etc), and make 95% of our research free – Focus: Possibilities / Probabilities of AI in Industry • Our work: – Commissioned market research for business and government clients (World Bank, others) – AI strategy development with a focus on aligning company goals with existing, powerful AI trends in the market • Covering the ethical concerns of AI since 2012
- 6. • “AI needs to have human values” • “Needs to be free from bias” • “AI should be used for good”
- 7. Tropes • “AI needs to have human values” • “Needs to be free from bias” • “AI should be used for good”
- 8. Tropes
- 9. Business Context Comes First • 1 - Strategic direction of the company: What thriving looks like in the future (growth, profit, positioning / place in the market) • 2 - What are the critical initiatives to get us to those goals, what would need to happen / change in the years ahead? • 3 - What can AI do… where can it fit into this mix? (this involves informing 2, but also adding new ideas to it… occasionally even adding to 1) – 3a - Determine the possible applications for “1” and “2” – 3b - Determine future of job roles / department structures – 3c - Determine potential ethical considerations that might matter
- 10. Business Context Comes First • The “ethics” conversation - like the “AI” conversation - can’t be addressed on its own because it’s a “hot topic.” – AI applications for their own sake are shameworthy in so much as they at an attempt to appear “cool” – AI ethics conversation at an executive level are shameworthy in so much as they are an attempt to appear “good”
- 11. 1 - Transparency • Decision-making criterion, when is transparency (interpretability) needed? – Not Needed: Content / product recommendation – Needed: Mutual fund allocations. Cancer diagnoses – ??: Loan or financing risk assessment • What are the decisions of algorithms based on? Are those the criterion we’d choose? Will there be legal or PR trouble if people realized what these systems were training on? – Income levels. Location of residence. Race. Gender. Age. Religion.
- 12. 2 - Accuracy • If you need 100% probably aren’t looking for machine learning as a solution. Machine learning is a statistical approach. If you need a simple “yes” or “no” to be final, then just program regular software to do just that, hard-code the rules. • If it’s impossible to hard-core the rules, and it’s impossible for an expert human - on their BEST day - to be 100% correct about an issue - then indeed we can’t expect a machine learning system to be any more accurate.
- 13. 2 - Accuracy • Better than a baseline (such as better than current human performance): – Customer service email tickets (DigitalGenius) – Paperwork filing or processing (RAGE Frameworks) – Fraud detection for payments or fake user accounts (SiftScience) • Long feedback loop: – Loan risk (must “trickle in” over time… very hard to call early on) – Hiring or recruiting (Humalyze) • Must be 100% accurate: – Presenting financial data in the form of reports to an executive (Ezeop) • Not applicable: – Recommendations for products or content (measured by response, not by % correct) (LiftIgniter)
- 14. [Graphic 2]
- 15. Generally Speaking: Transparency and Accuracy Requirements Time and Resources Required for Adoption
- 16. 3 - Accountability • What kinds of issues are “owned” by who? When something goes wrong, who takes the blame and who is responsible to fix it? – Generally, the AI team who builds the model and then continually stays on top of its algorithmic “drifting” should have to own the issues that come up • If an AI application is developed by an outside vendor, there must be some ongoing maintenance, testing, and upkeep on the part of the vendor and the buyer… creating a joint team responsible for the performance of the applications, and responsible for addressing the issues with the software. – Often, this should be drawn up in the contractual agreement with a vendor
- 17. 4 – Decisions: Man or Machine? • What can be handled completely by a machine? • What must be handled completely by a human? • Middle ground: – Sometimes something must be handed to a machine IF a certain criterion is the case, such as IF: • A system is less than 90% sure about it’s customer support ticket reply • A system is less than 80% sure about it’s risk assessment for a loan – Can be informed by a machine, but decision made by a human • Informing a plan for treatment - IBM Watson Health • Stock trading • Note: Many processes involve multiple decisions, and the role of many and machine will vary for each.
- 18. [Graphic 3]
- 19. Generally Speaking: The More Human-Machine Hand-Offs in an Application Time and Resources Required for Adoption
- 20. 5 – Job Security and AI • Big firms: – Hide as much of the conversations about restructuring or layoffs – (Possibly) Seem “progressive” by talking about these topics and doing retraining – (Definitely) Think about what each of your roles look like in the next 5-10 years – (Definitely) Hire what you need to hire, not as you have in the past. Hire tech whenever you can. • The only consideration big firms have to make here is with regards to perception. Manage it well: – Seem ready and willing to embrace disruption with a strong vision for the future where you WIN, not just survive. – Have - and express confidence in - a plan to use talent well in the re-oriented job functions you’ll have in the future. – Don’t avoid any planning about about potential layoffs and re-orgs, but obviously keep them under wraps. – If you have to do layoffs or re-orgs, do it with strong reference to (a) and (b)… i.e. a strong vision for thriving in the future.
- 22. 5 – Job Security and AI • Positions requiring management skills, or social connection / social tact are rather challenging to automate • Positions of the highest risk for automation are positions that require minimal context on anything other than the inputs and outputs of the role – High context vs low context – Plumber vs welder in a manufacturing line – Financial manager in procurement vs auditor
- 23. 5 – Job Security and AI https://www.youtube.com/watch?v=4NEIwKooOoI
- 24. [Graphic 5]
- 25. Generally Speaking: New Talent or Talent Re-Skilling Requirements Time and Resources Required for Adoption
- 26. 5 – Job Security and AI • Determine what the future of the various departments and job roles might look like, including: – Which roles might require more employees, which roles might require less – Critical talent acquisition needs in the next 2-5 years – How to re-train current employees to meet future demand – Initial pilot programs for re-skilling • Hypothetical examples: – Certain developers learning basic Python skills – Customer support agents learning to handle more edge-cases and high-touch troubleshooting
- 27. Take-Aways for Leaders • Keep these conversations housed squarely within your existing strategic planning process - as an extension and enhancement thereto, not as a distraction from it • Ethics can be framed as a consideration for decision-making, for: – Imbuing company values into our tech and products – Determining our resource requirements to bring AI’s value to life
- 28. That’s All, Folks Be in touch: dan@emerj.com @danfaggella