AI | Andrea M. Matwyshyn

AI-SOP'S FABLES

ANDREA M. MATWYSHYN

MIRANDA K. MOWBRAY

Introduction

CHAPTER 1: GRIM(M) TALES OF AI

What is AI?
1. Margaret Boden taxonomy
2. The Current Models
  1. SuperAGI and AGI
  2. GAI
  3. Narrow AI
Three Core Problems

1. old – security generally and tech debt

2. aggravated – transparency and data quality, attack automation

3. new(ish) – speed, degree of credible impersonation, AI cluster(ducks)

“Hallucination”
“Intelligence”
1. Theories of intelligence
2. Instrumentalist limited readings

3.Three Core Questions

1. Functions

1. is the particular system seeking to replace human labor?

2. is the particular system seeking to replace some human thinking?

3. is the particular system seeking to replace humans?

2. Supervision

1. Supervised

2. Unsupervised

3. Reinforcement

4. Combination

3. Applications

CHAPTER 2: ONCE UPON A(I) TIME...

The Terrible Toads: Context Specificity
1. Recursive Hamsters
2. Rabbit Fence
3. Data quality
The Digesting Duck: Harm
1. The Great Emu War
2. Defining “Autonomy”
The Argle-Bargle Bears: Intent
1. Goldilocks from three perspectives
2. Bench bear v. Lasagna bear v. roof bear v. revenge bear v. cocaine bear
3. Berenstain Bears
4. Degree of automation and reversal/ unlearning ability

CHAPTER 3:

Lessons for Future AI
1. Context
  1. Suitability is fact-specific, e.g. comedy
  2. Feedback problems
  3. Mushroom murder books
  4. Spy Whales and dolphins, predictive octopi, bomb/mine sniffing rats
  5. Beavers and inconvenient dams
2. Harm
  1. Autonomy issues in cars – hardware, training data (incorrect, inadequate), wrong generalization, faulty distribution, fear, screaming, loss of trust. Disconnect with normies, ignoring DOT
  2. Tay – prompt engineering: abusing systems but as built – a problematic repurposing, not a criminal act of computer intrusion even though the poor build can be a security issue
  3. Security issues –
    1. Training data manipulation by attackers or insiders
    2. Lack of model archiving and logging
    3. Lack of inspection ability
    4. Lack of capabilities for “unlearning”
    5. Lack of transparency on which training data (and which version) is included in proprietary products
    6. Data quality deficits leading to “spicy” prompt injection opportunities
      1. Sampling
      2. Labeling
      3. Proxy variable issues
      4. Bias magnification
      5. Secret whitelists/blacklists
      6. NLP problems
      7. Sensor data problems from hardware and/or design
      8. Dependency/Relying on third parties: There was recently, a particularly problematic discovery of a rogue insider in an open source application in the last couple of weeks that set the computer security world ablaze with discussion and concern because of its impact on Linux systems. And of course, some machine learning algorithms use linux in some component of their design operation or deployment. So we are living in a very interconnected world, where the public sector and the private sector interrelate, and we have flaws both into the technologies as built, but also in the governance that controls them.
    7. Same problems of internal controls and supervision that previously existed but even more complex, e.g. supply chain inclusion
    8. Legally problematic use
    9. And more…
  4. Evolutionary/emergent nature of harm and diagnosing it
  5. Dogs and orangutans driving etc. – self-preservation instinct overrides
3. Intent
  1. Goal specification (can be hybrid): money versus psych (ego/human replacement/ curiosity) versus malice
  2. Anthropomorphism of animals, e.g. love
  3. Ambiguity in line between malice and incompetence
  4. Corvids – disabling anticorvid tech – messy intent
  5. Squirrels and categories of adversarial attacks on AI, e.g. perturbation – don’t require high “intelligence” just persistence

CHAPTER 4:

The Path Forward
1. Whose viewpoint? Animal reality versus our imagined versions of them wo understanding the reality of their limitation and skills
  1. Cats – interaction effects re: anthropomorphism and obsession but see cat that brought phone to man who fell on bathroom floor. Exceptional cases happen so needs to be fact-specific analysis
  2. Bomb/land mine rats - Actual superiority – versus humans versus any current tech
  3. Turing v. Lovelace
  4. Humility
  5. Avoiding shallow definitions uninformed by humanities and all fields of psychology - not just neuropsych.
  6. Humans beat AI in human-machine symbiosis condition – DARPA grand challenge, Go experiment
  7. Human-machine symbiosis requires trustworthiness backstops in law
Context Sensitivity
1. Dogs
2. Capabilities that can exceed human
3. Special purpose v. general risks/ benefits
4. Spectrum of kinds
Harm
1. Look to existing categories of human harms
2. Failure of a duty to warn/ disclosure
  1. Failure to monitor
3. Failure to mitigate ongoing harm/ auditor-identified problems ß bolster clarity of this duty
  1. Replicability and replication as core to security correction / stochastic choice
Intent/Knowledge
1. Mental state attributions are questions of fact – don’t need one to apply to the machine
2. Was it under your control?
3. Audit requirements – willful blindness not ok
4. Don’t need a machine theory of mind
5. Don’t need a machine theory of consciousness
6. Definitely don’t need machine personhood legal theory

CHAPTER 5:

Remedies: a set of human persons must always be accountable/responsible
1. Tags – labeling of AI to signal categories of errors and set baseline expectations of fallibility
  1. Mowbray’s “Shevchenko Machine” – interrogated code must not lie about itself
  2. VW defeat devices fraud enforcement and additional mandatory disclosure
2. Pedigrees – data provenance and data quality disclosures
3. Training and testing – audit; threat modeling
4. Leashes – don’t use humans as guinea pigs; lab control
5. Muzzles – oversight and restraint
6. Neutering and capture – unlearning; board and individual consequences
7. Animal Control – a new Bureau of Technology Safety to address emergent problems etc.

Conclusion