Hi, I’m Chris Flathmann
CF

I study what happens when people and AI have to work together.

I’m an assistant professor at Clemson, where I direct the BIG CAT Research Group and co-direct CU-CHAI. My work focuses on human-AI teams, trust, training, and performance.

TeamworkCommunicationTrustResilience
HCC 3030 · Week 1

What Do We Mean by
Human-Centered AI?

A way to design and evaluate AI that starts with people, work, and consequences, not just model performance.

Start with the hard question

When an AI-influenced decision hurts someone, who owns the mistake?

The developer? The interface designer? The person using it? The organization? Someone else?

A distinction we will use all semester

An accurate model can still fail in the real world.

How well the
model performs
Accuracy, calibration, latency, and robustness.
How well the
whole system works
Outcomes, workflow fit, access, workload, recovery, and legitimacy.
A practical definition

We can recognize AI by what it infers and what that output can change.

AI
system
A machine-based system that takes inputs and infers predictions, content, recommendations, or decisions. It may operate with more or less autonomy and may change after deployment.
Three levels that are easy to blur together

The model, the product, and the system are not the same thing.

01 · MODEL

Produces an output

Turns inputs into a score, label, prediction, or generated response.

Example: risk score
02 · PRODUCT

Puts the output to work

Adds an interface, thresholds, defaults, explanations, and controls.

Example: triage dashboard
03 · SYSTEM

Creates real consequences

Includes the people, policies, incentives, infrastructure, and communities involved.

Example: city service process
Follow the output

A model output matters only when the rest of the system does something with it.

Inputs

Reports, records, sensors.

Inference

Score, classify, generate.

Interface

Display, alert, rank.

Decision

Accept, edit, defer.

Action

Allocate, deny, dispatch.

Outcome

Benefit, burden, harm.

Why we call these sociotechnical systems

The model never works alone.

Context
historycultureinfrastructureinequality
Institution
goalspolicyauthorityincentives
Work
rolescoordinationtrainingtime pressure
Interface
defaultsexplanationscontrolsrecovery
Model
dataobjectivearchitectureevaluation
Our working definition
HUMAN
CENTERED
Start with people’s goals, abilities, values, and lived consequences, then keep those concerns present from problem framing through retirement.
Three shortcuts that do not hold up

A friendly interface does not make a system human-centered.

×
Friendly interface
Usability cannot repair a harmful objective or illegitimate use.
×
Ethics checklist
Principles need mechanisms, owners, evidence, and stop rules.
×
Human approval
A click is not oversight when the person lacks time, knowledge, or authority.
Automation and control are different questions

Automation and human control can rise together.

What role have we given the AI?

The closer AI gets to acting, the more authority we give it.

Sense

Detect a pattern in text, image, audio, or behavior.

Predict

Estimate a future state, risk, demand, or outcome.

Recommend

Rank options and direct human attention.

Generate

Create content, plans, code, or candidate designs.

Act

Call tools, transact, allocate, communicate, or control.

Less direct authorityMore direct authority →
The action surface

Risk changes when a suggestion becomes an action.

Display

A person sees an output.

Recommend

The system orders attention.

Default

Inaction becomes a decision.

Execute

The system changes the world.

“Human in the loop” tells us very little

Putting a person in the loop does not guarantee meaningful oversight.

The question is whether that person can notice, judge, and act.

What can the reviewer actually know, decide, and change?
Four practical tests for oversight

Can the person spot a problem in time and do something useful?

01
Notice
Will a problem become visible before harm is locked in?
02
Understand
Can the person interpret the output, context, and uncertainty?
03
Intervene
Do they have time, alternatives, authority, and a safe control?
04
Learn
Will the incident change the model, workflow, or policy?
Agency
Can people meaningfully shape what the system does?

Agency requires real options, enough understanding to choose, and controls that make the person’s intention matter.

directionchoicecorrectionrefusal
Legibility

People need information that helps them act, not every internal detail.

Capability
What can it do reliably?
State
What is it doing now?
Basis
What evidence shaped this output?
Limits
Where should it not be trusted?
Recovery
How can a person correct or reverse it?
The right view depends on the role.

A dispatcher, auditor, affected resident, and model engineer need different evidence about the same system.

Uncertainty

“The model is 82% confident” is not the whole uncertainty story.

Output
How stable or ambiguous is this specific inference?
Could another answer fit?
Model
Does evaluation support use for this population and condition?
Did we test here?
Data
What is missing, delayed, measured poorly, or produced selectively?
What is unseen?
World
Could the environment, behavior, or objective change?
Will tomorrow match?
Appropriate reliance

The goal is knowing when to rely on the system, and when not to.

Under-reliance

Reject useful support despite evidence that it fits the task.

Appropriate reliance

Use or reject based on capability, context, evidence, and stakes.

Over-reliance

Accept output beyond the system’s competence or authority.

Ignore everythingAccept everything
Reliance depends on a relationship

Good judgment depends on the AI, the person, and the stakes.

System capability

Can it perform here?Population · condition · task · time
Evidence defines the competence boundary.

Human capability

Can the person judge here?Expertise · attention · workload · incentives
Expertise and context shape verification.

Decision stakes

What if we are wrong?Severity · reversibility · distribution · recourse
Higher stakes demand stronger evidence and control.
Four ways automation shows up in practice

Automation can be used well, misused, ignored, or imposed.

Use

Voluntary engagement when automation is suitable.

Misuse

Over-reliance or use outside the competence boundary.

Disuse

Under-reliance that rejects beneficial automation.

Abuse

Automation imposed without adequate regard for human consequences.

Responsibility

Someone must own each responsibility.

Problem owner
Defines the legitimate purpose and acceptable harm.
Should AI be used here?
Builder
Documents data, objectives, limits, and evaluation.
What evidence supports this use?
Deployer
Fits the system to workflow, policy, and population.
What changed at deployment?
Operator
Uses judgment within a defined authority boundary.
When must I intervene?
Governor
Audits, responds, repairs, and can pause or stop use.
Who can halt the system?
Contestability

If AI shapes a consequential decision, people need a way to challenge it.

01

Detect

Know that AI influenced the outcome.

02

Question

Access the reason and relevant evidence.

03

Correct

Repair data, context, or assumptions.

04

Appeal

Reach an empowered, independent reviewer.

05

Repair

Reverse harm and improve the system.

Governance belongs inside the system

Rules become real through permissions, logs, thresholds, review, and stop controls.

Governance decides who may do what, what evidence is required, and who reviews the result.

It limits action before an incident and makes recovery possible afterward.

01
Purpose and prohibited uses
02
Authority and access controls
03
Evaluation and deployment conditions
04
Monitoring, audit, and incident response
05
Appeal, override, pause, and shutdown
Human-centered design still involves tradeoffs

The important move is to make the tradeoff visible.

Speed
AND
Deliberation
Consistency
AND
Contextual discretion
Personalization
AND
Privacy
Automation
AND
Human expertise
The same lens travels across domains

The questions stay useful, but the evidence changes with the setting.

Human-Centered
AI

Healthcare

Thresholds, workflow, alert burden.

Education

Learning evidence, access, integrity.

Fairness

Metrics, proxies, policy choices.

Privacy

Purpose, inference, correction.

Agriculture

Ecology, allocation, resilience.

Development

Context, language, dependency.

Planetary systems

Boundaries, rebound, lifecycle.

Work and power

Authority, audit, shutdown.

A four-question check for any AI application
SYSTEM
OUTCOMES
agencyresponsibilitycontestabilitylegibility

Before calling an AI system “good,” ask four questions.

Purpose: What human goal justifies the system?

Evidence: What supports this use in this context?

Power: Who can direct, question, correct, and stop it?

Consequences: Who benefits, who carries risk, and who can recover?