1 of 8

Detect-Interpret-Respond

A framework to ground youth sense-making of AI

Eric Greenwald, Ari Krakowski, Timothy Hurt, & Ning Wang

University of California, Berkeley & USC Institute for Creative Technologies

2 of 8

DIR offers a flexible and extensible schema to build understanding of a variety of AI systems

To give a general description of the framework, we start with the idea that there is a world out there, on the one hand, and an AI response on the other.

From this basic input/output; stimulus/response model, we focus the learner on a simple explanatory account of AI as a system (or system of systems):

With DETECTION, an AI needs to take in information from the outside world and translate this analog world in computational terms: light is converted into pixels

With INTERPRET, An AI needs to make meaning of the information it detects: patterns of pixels are interpreted as faces

With RESPOND, the AI needs to take goal-directed action in relation to the meaning made: a face is identified to the user as a particular person, or exhibiting a particular expression

Let’s think of a self-driving car approaching a stop sign:

First, important bits of the world must be detected: Salient aspects of the environment are captured and computationally rendered as input data; the red octagon needs to be perceived

Next, those data, those computational abstractions are interpreted–they are assigned meanings and values; the red octagon is a stop sign; cars stop at stop signs

and, finally, based on the assigned meanings, the AI system provides a response in line with system goals; brakes are applied and the car comes to a stop

Through this framework a wide variety of AI systems can be simplified toward this basic mechanistic account

With many AI applications familiar to youth, the AI system can be understood in terms of how the external world is captured and quantified as data; those data are processed for interpretation according to system goals; that meaning is then operationalized toward a response

3 of 8

For more about the VH: https://vhtoolkit.ict.usc.edu/index.html

4 of 8

The Virtual Human as a case of AI

5 of 8

Identification verification, through the lens of DIR

digitized “face” 1

digitized “face” 2

Compare “face” 1 to “face” 2

Match/

No match

human

passport

admit/ detain

To provide another illustration of how DIR might help provide an age-appropriate explanatory account of an AI system, let’s take an example that those of you who’ve traveled to get here (or other recent international travel) may have encountered–

I’m from Berkeley, CA and while I was traveling in Europe earlier this summer with my family, we encountered a new (to me) system to verify that I was the person identified in my passport. Basically, a customs agent took a photo of me and of my passport, looked at their phone, and then told me it was ok to go through. When my children (10 and 13) asked what was happening, I leveraged DIR to offer a simple explanation of the system–as i’m sure you’ve pieced together, it was using the facial recognition software of a modern camera to match my actual face to the photo on my passport: In explaining this to my kids, I pointed out that

The system needs to detect my face from my physical body and from my passport using a camera, and then computationally render each as a data set–the two faces get converted into something the computer can understand; those computational abstractions are then compared and the match is reported to the user.

6 of 8

DETECT

RESPOND

INTERPRET

Symbols

Sensors

Models & Knowledge representation

Rules, policies, algorithms

Classifications & diagnosis

Actions & behavior

Decisions & predictions

Synthesis & genesis

Computation

quantification

Lens, microphones & keyboards

Neural Nets

ML

NLP

Robotics

Animation

Ethics

Language, Aesthetics

To step back out and imagine AI more broadly, the DIR schema offers a 3-phase breakdown to inspect an AI system, or to imagine how a human-performed task may be simulated by artificial intelligence. This is somewhat analogous to understanding human intellectual achievement as: perception/cognition/behavior: For an AI system, the external world is captured and quantified as data; those data are processed for interpretation according to system goals; and that meaning is operationalized toward a response

One thing that we find especially promising about DIR is how it is both flexible to accommodate a variety of AI systems and extensible to increasingly sophisticated understandings of AI: AI comes in many forms and DIR offers a schema to attach new ideas and deeper explanations:

That is to say, DIR can grow with youth as they advance in understanding, for example, exploring increasingly technical understanding of how “interpretation” operates,

To explore this a bit, and reveal just how interdisciplinary an enterprise AI development is, let’s start with the myriad human-intelligible aspects of artificial intelligence (yellow ovals); and we can also consider this from the perspective of all the different expertises involved in building an AI system (red ovals)

A challenge is that sometimes, especially where the goal is human simulacra, the response is often invisible/unremarkable/ordinary: one thing we ran into was that little kids just didn’t notice the nonverbal, backchannel behaviors–they were too subtle, and, as those of you who have spent time with young children know, kids aren’t fully in control of their bodies so they aren’t particularly attuned to nonverbal communication.

We’ve had some success by mirroring the breaching experiments of ethnomethodologists who reveal the invisible routines and rituals of the everyday by “misbehaving”

Another challenge are deep learning models that are essentially illegible to humans–there is input and output, but interpretation is a black box; we see this as both a limitation of DIR’s applicability, but also one that reveals to youth a broader ethical concern about the risks of AI systems that are not human-explainable–we cannot audit them, for example, to account for consequential outputs and decisions.

7 of 8

This material is based upon work supported by the National Science Foundation (AISL) under Grant No. 2116109.

8 of 8

Thank you!

eric.greenwald@berkeley.edu