Bias and discrimination:
Humans are biased and discriminatory, and AI uses human data to train. This means the AI
itself gets biased. There is little to nothing to do about it.
There is widespread discrimination in society, and accurate, representative data will reflect
this in the AI.
Accurate data is biased, unbiased data is inaccurate.
Self-fulfilling prophecies: you investigate the people that are most likely to commit a crime,
so that group gets targeted and indeed commits the most crime (because you only look for
them). This is the problem of predictive policing.
Privacy
The AI systems use loads of data to create a training set on which decisions are being made.
This also includes personal data.
The GDPR is in place to protect personal data (any information that relates to identifying an
individual).
Processing data is only legal when consent is given, or to protect the general public.
Data minimization principle: collect no more data then necessary. (Difficult to enforce,
ambiguity surrounding what is necessary)
Proposals: ban of data use for targeted advertising, limiting use of sensitive data for all
secondary purposes, banning biometric collection of children or in specific contexts (work,
school)
Limits: new tools that don’t require personal data
Improvements: Restrict availability of consent to situations where consequences solely
affect that user. Render legal basis of consent insufficient for uses where data will be linked
to other people’s data. Establish collectivist rights, such that groups affected can also assert
rights.
Mass surveillance is not illegal for national security reasons.
Affect Recognition: technology that analyses facial expressions from both static images and
videos to reveal information on one's emotional state. Not developed in a way that satisfies
data protection yet.
Predictive privacy: information that can be guessed about people by matching it to other
people’s information.
This results in bad decisions being made by AI: credit ratings, job applications, nudging. The
GDPR does not combat this.
Anonymization: the process of removing personal identifiers, both direct and indirect, that
may lead to an individual being identified.
Does not fall within the scope of the GDPR.
Workplace surveillance: any form of employee monitoring undertaken by an employer. Low
wage workers affected most radically.
Surveillance Interoperability: combination of govt data collection, corporate surveillance
and third-party harvesting.
Systems are riddled with flaws yet make decisions about workers.
, EU law to protect them: Worker Access to Data, Algorithmic Transparency, Contestability.
Improvements: not only be for platform workers, include whistleblower protections, should
have collective, not just individual, rights
Human in the loop: proposals focus on requirement of human review. But:
AI is rarely fully automated, more a spectrum of human operated and machine.
Automation bias: humans likely trust the system.
Responsibility goes to human who has little to do with the decision.
Policy suggestions:
- Focus less on how data is collected and what happens to that data, but rather on the
actual outcome of the data usage.
- Collective rights, not just individual rights: As such, algorithms can be considered for
their effect on collective control.
Defining AI
How we define AI fundamentally impacts how we consider it / use it societally
AI Act definition:
a) Machine learning approaches, including supervised, unsupervised and reinforcement
learning, using a wide variety of methods including deep learning;
(b) Logic- and knowledge-based approaches, including knowledge representation, inductive
(logic) programming, knowledge bases, inference and deductive engines, (symbolic)
reasoning and expert systems;
(c) Statistical approaches, Bayesian estimation, search and optimization methods.
Very broad!
Regulating AI:
Law divides things into:
- Sources of obligation
Commodities – objects of the law. Therefore, can have legal binding effects. E.g., I
broke my neighbor’s TV (the TV is a commodity).
If not a commodity, not recognized by the law.
- Actors: Subjects whose actions have legal consequences
- Persons: Hold rights and duties
Natural persons like humans
Legal persons like companies
What is AI? Not an actor or legal person
On what level: International, Regional (Europe, Americas), National, Local
Self-regulation: companies that create AI regulate themselves, so innovation is not impeded.
Pros: they know the software and systems the best. Cons: they do not have the best interest
in mind for the general public, only for making money. No incentive to think about privacy,
data protection, etc.
Humans are biased and discriminatory, and AI uses human data to train. This means the AI
itself gets biased. There is little to nothing to do about it.
There is widespread discrimination in society, and accurate, representative data will reflect
this in the AI.
Accurate data is biased, unbiased data is inaccurate.
Self-fulfilling prophecies: you investigate the people that are most likely to commit a crime,
so that group gets targeted and indeed commits the most crime (because you only look for
them). This is the problem of predictive policing.
Privacy
The AI systems use loads of data to create a training set on which decisions are being made.
This also includes personal data.
The GDPR is in place to protect personal data (any information that relates to identifying an
individual).
Processing data is only legal when consent is given, or to protect the general public.
Data minimization principle: collect no more data then necessary. (Difficult to enforce,
ambiguity surrounding what is necessary)
Proposals: ban of data use for targeted advertising, limiting use of sensitive data for all
secondary purposes, banning biometric collection of children or in specific contexts (work,
school)
Limits: new tools that don’t require personal data
Improvements: Restrict availability of consent to situations where consequences solely
affect that user. Render legal basis of consent insufficient for uses where data will be linked
to other people’s data. Establish collectivist rights, such that groups affected can also assert
rights.
Mass surveillance is not illegal for national security reasons.
Affect Recognition: technology that analyses facial expressions from both static images and
videos to reveal information on one's emotional state. Not developed in a way that satisfies
data protection yet.
Predictive privacy: information that can be guessed about people by matching it to other
people’s information.
This results in bad decisions being made by AI: credit ratings, job applications, nudging. The
GDPR does not combat this.
Anonymization: the process of removing personal identifiers, both direct and indirect, that
may lead to an individual being identified.
Does not fall within the scope of the GDPR.
Workplace surveillance: any form of employee monitoring undertaken by an employer. Low
wage workers affected most radically.
Surveillance Interoperability: combination of govt data collection, corporate surveillance
and third-party harvesting.
Systems are riddled with flaws yet make decisions about workers.
, EU law to protect them: Worker Access to Data, Algorithmic Transparency, Contestability.
Improvements: not only be for platform workers, include whistleblower protections, should
have collective, not just individual, rights
Human in the loop: proposals focus on requirement of human review. But:
AI is rarely fully automated, more a spectrum of human operated and machine.
Automation bias: humans likely trust the system.
Responsibility goes to human who has little to do with the decision.
Policy suggestions:
- Focus less on how data is collected and what happens to that data, but rather on the
actual outcome of the data usage.
- Collective rights, not just individual rights: As such, algorithms can be considered for
their effect on collective control.
Defining AI
How we define AI fundamentally impacts how we consider it / use it societally
AI Act definition:
a) Machine learning approaches, including supervised, unsupervised and reinforcement
learning, using a wide variety of methods including deep learning;
(b) Logic- and knowledge-based approaches, including knowledge representation, inductive
(logic) programming, knowledge bases, inference and deductive engines, (symbolic)
reasoning and expert systems;
(c) Statistical approaches, Bayesian estimation, search and optimization methods.
Very broad!
Regulating AI:
Law divides things into:
- Sources of obligation
Commodities – objects of the law. Therefore, can have legal binding effects. E.g., I
broke my neighbor’s TV (the TV is a commodity).
If not a commodity, not recognized by the law.
- Actors: Subjects whose actions have legal consequences
- Persons: Hold rights and duties
Natural persons like humans
Legal persons like companies
What is AI? Not an actor or legal person
On what level: International, Regional (Europe, Americas), National, Local
Self-regulation: companies that create AI regulate themselves, so innovation is not impeded.
Pros: they know the software and systems the best. Cons: they do not have the best interest
in mind for the general public, only for making money. No incentive to think about privacy,
data protection, etc.
