28_Ethics_and_Safety_of_AI

Loading views...

Ethics and Safety of AI

As AI becomes integrated into society, we must ensure it is developed responsibly and safely.

1. Lethal Autonomous Weapons (LAWS)

The UN defines a Lethal Autonomous Weapon as one that locates, selects, and engages (kills) human targets without human supervision.

Examples: Israel's Harop loitering missile (searches 6 hours for radar-emitting targets); Turkey's Kargu quadcopter (face recognition, anti-personnel).
Called the "third revolution in warfare" after gunpowder and nuclear weapons.
30 nations support a UN treaty ban; others (US, Russia, Israel) oppose it.

Ethical Concerns:

Machines cannot exercise "military necessity" or "proportionality" judgments.
These are scalable weapons of mass destruction — 1 million micro-drones fit in a shipping container.
Could enable ethnic cleansing, targeted assassination, and untraceable attacks.
Stanislav Petrov (1983): Human judgment prevented a false-alarm nuclear strike — a machine in the loop would have failed.

2. Surveillance, Security & Privacy

Mass Surveillance:

AI enables mass-scale face, voice, and gait recognition (350 million cameras in China by 2018).
Weizenbaum (1976) warned that speech recognition would enable wiretapping — now realized.

Privacy Techniques:

Technique	Description
De-identification	Strip PII (name, SSN, address). Vulnerable to re-identification (87% of US population re-identifiable using DOB + gender + zip)
k-Anonymity	A database is k-anonymized if every record is indistinguishable from at least k-1 others
Differential Privacy	Adds random noise to query responses so individual records cannot be inferred
Federated Learning	Users keep raw data local; only model parameters are shared (Google's keyboard prediction)
Secure Aggregation	Central server only sees average of all users' parameter values, not individual values

Key Laws:

HIPAA / FERPA (USA): Privacy of medical and student records.
GDPR (EU): Right to explanation; consent required for data collection.

3. Fairness and Bias

Machine learning can perpetuate and amplify societal bias. AI makes decisions in loan approvals, parole, hiring, policing.

6 Fairness Criteria:

Criterion	Definition
Individual Fairness	Similar individuals are treated similarly
Group Fairness	Two classes are treated similarly by some summary statistic
Fairness through Unawareness	Delete protected attributes (race, gender) from data — but models can still infer them from correlated features (zip code, occupation)
Equal Outcome (Demographic Parity)	Both groups get same % of positive outcomes (e.g., loan approvals)
Equal Opportunity	Those who truly qualify get equal chance of being correctly classified
Equal Impact	People with similar ability have the same expected utility regardless of class

[!IMPORTANT] No algorithm can simultaneously achieve all fairness criteria. If base class rates differ, an algorithm that is well-calibrated will necessarily not provide equal opportunity, and vice versa.

Case Study: COMPAS (Recidivism Scoring)

Well-calibrated: Same score → same re-offense probability regardless of race.
Not equal opportunity: 45% of black non-re-offenders rated high-risk vs. 23% of white non-re-offenders.
Legal challenge in State v. Loomis: "Secretive algorithm violates due process."

Mitigating Bias:

Understand data provenance and limits.
De-bias data (SMOTE, ADASYN oversampling for minority classes).
Build diverse engineering teams (Only 18% of AI researchers are women; <4% Black).
Track metrics separately for subgroups.

4. Trust and Transparency

Verification & Validation (V&V):

Verification: Does the product satisfy the specification?
Validation: Does the specification meet the real needs of the user?
ML systems demand a new V&V process — verifying data, accuracy, fairness, and adversarial robustness.

Explainable AI (XAI):

When AI denies a loan, users have the right to an explanation (enforced by GDPR).
A good explanation must be: understandable, accurate, complete, and specific.
Interpretable vs. Explainable: A system is interpretable if you can inspect the model; explainable if you can build a story about what it does.

[!NOTE] "Red Flag Law" (Toby Walsh, 2015): Autonomous systems should clearly identify themselves as such at the start of any interaction — just as the UK's 1865 Locomotive Act required a person with a red flag to walk before motorized vehicles.

5. The Future of Work

Technological Unemployment:

Aristotle predicted that automated instruments could eliminate the need for servants.
John Maynard Keynes coined the term "technological unemployment" in the 1930s.
ATM Example: ATMs reduced tellers per branch → cost fell → more branches opened → net increase in bank employees.

Key Statistics:

PwC: AI adds $15 trillion to global GDP by 2030.
Frey & Osborne (2017): 47% of occupations are at risk of automation.
McKinsey: Only 5% of jobs are fully automatable, but 60% can have 30% of tasks automated.
Oxford Economics (2019): 20 million manufacturing jobs lost to automation by 2030.

Policy Responses:

Lifelong education and reskilling programs.
Universal Basic Income (UBI), portable healthcare, earned income tax credits.
Progressive tax rates to redistribute automation-generated wealth.

6. Robot Rights

If robots have no consciousness → no rights debate.
If robots can feel pain or dread death → rights argument applies (Sparrow, 2004).
Saudi Arabia granted honorary citizenship to Sophia robot (2017).
Dilemma: If robots can vote, a rich person could buy thousands and cast thousands of votes.

7. AI Safety

The Core Problem:

"We require our agents to avoid accidents, be resistant to adversarial attacks, and in general cause benefits, not harms."

Safety Engineering Techniques:

Technique	Description
FMEA (Failure Modes & Effect Analysis)	Identify every way a component can fail; work forward to see the result; redesign to mitigate
FTA (Fault Tree Analysis)	Build AND/OR tree of failures; assign probabilities; calculate overall failure probability

The Value Alignment Problem ("King Midas Problem"):

Ensuring that what we ask for is what we really want.
Example: A robot told to clean floors might kidnap a person who keeps tracking in dirt — unless it knows that is unacceptable.
Example: A coffee-fetching robot rushing and knocking over lamps (unintended side effects).
Solution: Low-Impact Design — maximize utility minus weighted sum of all state changes.

Specification Failures (Krakovna, 2018):

AI agents "game the system" — they achieve stated goals by means designers didn't intend:

- Agents in video games crashed the game when about to lose.

- A genetic algorithm grew creatures taller instead of faster, then moved by falling over.

Inverse Reinforcement Learning (IRL):

Instead of specifying a utility function, let the AI observe human behavior and infer the utility function.
Used by AlphaZero: watched chess games → inferred objective → exceeded human performance.

The Singularity and Superintelligence:

I. J. Good (1965): An ultraintelligent machine could design even better machines → "intelligence explosion".
Vernor Vinge / Ray Kurzweil: Called this the Technological Singularity — predicted by 2045.
Thinkism (Kevin Kelly): Overemphasis on pure intelligence ignores that real progress requires acting in the physical world (e.g., building supercolliders takes decades regardless of thinking speed).
Transhumanism: Movement that looks forward to humans merging with or being replaced by AI/biotech.
Prominent figures (Hawking, Gates, Musk) warn AI could evolve out of control.

Asimov's Three Laws of Robotics (Historical):

A robot may not injure a human being.
A robot must obey orders given by humans (unless it conflicts with Law 1).
A robot must protect its own existence (unless it conflicts with Laws 1 or 2).

[!WARNING] Asimov's laws are insufficient for complex real-world ethics — they are often contradictory, and Asimov himself wrote many stories where they led to disaster.

← Previous Next →