H-Farm&Maize

We conduct research, intervention, and outreach on the interaction between people and technology.

What was

Critical view and technological criteria

Difusión

Exposición del aprendizaje alcanzado

Research

with empirical experiments

Outreach

(articles, talks, scientific papers)

Exploration

of technology trends

Intervention

by prototyping and stories

RESEARCH

& OUTREACH

PROCESSES & BIASES

How to improve the decision-making process when AI & humans have to collaborate.

Humans show bias to AI-artwork.

Ingredients to research on a topic:

the topic will be relevant soon.
the topic is starting to be mentioned in the media.
we can use our knowledge of human behavior and digital interfaces to formulate the hypotheses of the experiment.

human-in-the-loop decision processes

European Commission

default

Human-in-the-loop decision process (e.g. RisCanvi)

Does this process favor excessive compliance (automation bias)?

Government officials using RisCanvi disagree with the algorithm

that is, only two inmates out of ten end up confirming the system’s prediction and reoffending after being classified as high risk.

only 3.2% of the time.

positive predictive value of 18%

This is so, even though, RisCanvi has a

Saura, G., & Aragó, L. (2021). Un algoritmo impreciso condiciona la libertad de los presos. La Vanguardia

Hypothesis:

Manipulating the time at which AI assessment is presented and forcing humans to explicitly emit a judgment can help to…

(a) increase the accuracy when the system makes errors.

(b) reduce excessive compliance.

Design of the two experiments

Group

Step 0

Step 1

Step 2

AISupport→Judgment

Judgment→AISupport

Description of a criminal case and witness testimonies

Confirm or modify AI’s assessment

Judgment without AI Support

Confirm or modify AI’s assessment

Procedure

Step 0

Description of a criminal case and witness testimonies.

Procedure

Step 1.

Confirm or modify AI’s assessment.

AIsupport → Judgment Group

Procedure

Step 1.

Confirm or modify AI’s assessment.

AIsupport → Judgment Group

Procedure

Step 2.

Judgment without AI Support.

AIsupport → Judgment Group

Order reversed

Group

Step 0

Step 1

Step 2

AISupport→Judgment

Judgment→AISupport

Description of a criminal case and witness testimonies

Confirm or modify AI’s assessment

Judgment without AI Support

Confirm or modify AI’s assessment

Group

Step 0

Step 1

Step 2

AISupport→Judgment

Judgment→AISupport

Description of a criminal case and witness testimonies

Confirm or modify AI’s assessment

Judgment without AI Support

Confirm or modify AI’s assessment

Were they more accurate judging at the beginning of the process than at the end?

Judgment→AISupport Group were more accurate

Group

Step 0

Step 1

Step 2

AISupport→Judgment

Judgment→AISupport

Description of a criminal case and witness testimonies

Confirm or modify AI’s assessment

Judgment without AI Support

Confirm or modify AI’s assessment

Were there difference between groups in compliance with the incorrect AI assessment?

Low compliance... but reduced accuracy

We did not find a statistically significant difference in compliance between groups.
Erroneous AI support negatively affected participants’ accuracy decision when they modified the AI assessment.

STORIES

https://riscanvi.bikolabs.io/

STORY WITH 3Ds

Cables.gl to explain complex concepts such as the different ways to measure the accuracy of an algorithm.

https://www.bikolabs.io/empatizandoconias

SCROLLYTELLING

An interactive story with irony to better understand how AIs "see" different everyday objects.

Previous exploration/informal research

https://www.bikolabs.io/collections/carta

INTERVENTION

We wrote and promoted a letter asking the Spanish Government to a moratorium on the use and commercialization of facial recognition and analysis systems.