MAIO

• Algorithmic management shapes how managers perceive their professional identity
• Managers want to deepen their autonomy, set professional boundary against new practices and other groups, and enhance their status by increasing their specialization
• Professional identities affect the way managers integrate knowledge claims generated by algorithms

 -> Managers will support or reject algorithmically-generated decisions depending on whether they view it as enhancing or undermining their professional identities and status

• Shifting roles: workers and managers not passive recipients of algorithmic results as they align algorithmic systems to their needs and interests
• Demand for algorithmic competencies: skills supporting workers in developing symbiotic relationships with algorithms
• Risk of upskilling only a fragment of organizational members (deskilling others)
• Emerging role of “algorithmists”: data scientists, human translators, mediators, and agents of algorithmic logic (Gal et al., 2020)
• Algorithmic competencies limited by resistance against algorithms: algorithmic aversion and cognitive complacency

• Technical opacity: “black box” character of AI systems (design principles and operational complexity)
• Organizational opacity: lack of information due to strategic interests and intellectual property (professional gatekeeping and externalization of algorithmic development to third parties)
• Solutions:
    • Explainable AI: making complex models more understandable by humans through various technical methods (labor-specific and organization-specific policies)
    • Algorithmic audit: idea that algorithms may produce a record that can be read, understood, and altered in future iterations
    • Disclosing more information on AI systems: human involvement, data, models, and algorithmic inferences
    • Stakeholder involvement in AI design: communication and deliberation 

Innovation is the development and implementation of new ideas by people who over time engage in transactions with others within an institutional order

Invention is the conversion of cash into ideas.

Innovation is the conversion of ideas into cash.

Novel

Useful

Implemented

Create new knowledge and provide significant technological breakthroughs in products and processes

Build upon existing knowledge base and provide small improvements in current product lines and processes

image

•(1)opportunity identification and idea generation – e.g. identifying user needs, scouting promising technologies, generating ideas;

•(2)idea evaluation and selection – e.g. idea assessment, evaluation;

•(3)concept and solution development – e.g. prototyping, concept testing; and

•(4)commercialization launch phase – e.g. marketing, sales, pricing.

•Use generative AI to free up the time for innovation

•Use generative AI to create.

•An enemy for innovation and creativity?

Text summarization

Exploring solution spaces with AI (idea generation)

Creativity is the source of innovation

Innovation is the process of transforming creative ideas to create value

•Application

•Value creation

Robustness

•Originality

•Imagination

Flexibility

image

AI is increasingly being used in decision-making in sensitive industries e.g. banking

1. Contemporary AI (e.g. LLMs) have access to internet
2. Self-replication capability: We have taught AI how to code
3. AI knows about humans: we have taught AI about human behaviours

image

Vulnerability -> Robustness

Hazard exposure -> Monitoring

Hazard -> Alignment 

Adversarial Attacks and Privacy Attacks

An adversarial attack is designed to fool an ML model into causing mispredictions by means of injecting deceitful data meant to deceive classifiers. This type of corrupted input goes by the name adversarial example

An adversarial example is a corrupted instance characterized by a perturbation of small magnitude, virtually imperceptible,which determines the ML model to make a mistake. To human eyes, adversarial examples seem identical to the original. To machines, however, they work almost as an optical illusion, causing them to misclassify data and make false predictions.

In privacy-related attacks, the goal of an attacker is to gain knowledge that was not intended to be shared. Such knowledge can be about the training data or information about the model or even extracting information about properties of the data

Targeted Attack : aims to misclassify the input (e.g., image) to a specific label (e.g. panda to gibbon)

Untargeted Attack : aims to misclassify the input to any wrong label (e.g. panda to any other animal)

Many works have tried to, but follow-up works showed that all fail

The main successful defenses in practice now incorporate adversarial examples during training

kimage

Model Inversion

Data Extraction 

Membership Inference

image

image

Yes, Preventing memorization hurts accuracy

image

Goal: Attacker uses the model to infer if a particular data point is present.

Defend using differential privacy!