foxxelabs.ie
FoxxeLabs Research • August 2026

Human-AI
Interaction

A new field — and what it is already showing us.

Todd McCaffrey, MSc Cyberpsychology

The field

What changed: software can collaborate

What HCI was built for

Tools that execute  ·  1980–2020
  • Deterministic — same input, same output
  • Predictable — behaviour is fully specified
  • Static — only the human learns
  • Silent about its own limits
  • Waits to be used

What AI systems now do

Partners that collaborate  ·  2020s →
  • Generative — same prompt, new possibilities
  • Adaptive — it learns and improves with use
  • Fallible like a colleague — and worth checking like one
  • Can express confidence, reasoning and doubt
  • Takes initiative: suggests, drafts, acts

Each capability is a gain — and each breaks an assumption HCI was built on. Hence a new field.

The field

Each generation of machine got its own field

The label hasn’t settled yet. That is the first useful fact about the field.

1
1960s–70s

Human-Machine Interaction

Ergonomics, control panels, cockpits. Deterministic machines.

2
1980s–2010s

Human-Computer Interaction

The PC. Usability, GUIs, mental models. Software obeys.

3
2020s →

Human-AI Interaction

Systems that generate, adapt and act. Software collaborates.

The field

One field, four names

 

Human-AI Interaction

HAII · HAI · HAX — the academic default

How people and AI systems communicate and collaborate. CHI, IUI, HAXD conferences.

 

Human-Centered AI

HCAI — Shneiderman

Keep human control high while automation is high. Reliable, safe, trustworthy beats autonomous.

 

Human-Agent Interaction

HAI (ACM) — robots, agents, now LLM agents

HAI 2026, Osaka: “From Interaction to Agency: Navigating Autonomy.”

 

AI Literacy

Long & Magerko 2020 · EU AI Act Art. 4

Knowing how to use, question and live with AI. A legal obligation for deployers since Feb 2025.

The field

What the field actually studies

 

Calibrated trust

When should a person rely on the system — and when not? Over-reliance and under-reliance are both failures.

 

Mental models

What do people believe the system is doing? Usually something simpler and more human than the truth.

 

Uncertainty & explanation

How does a system say “I might be wrong” in a way people actually use?

 

Agency & delegation

Who decides, who acts, who is accountable — when the system can act on its own.

 

Cognition & learning

What happens to memory, reasoning and skill when thinking is shared with a machine?

 

Co-creation

Writing, coding, designing with an AI partner: who owns the idea, and does the output converge?

The field

Who is doing the work

Ben Shneiderman — University of Maryland

The Human-Centered AI framework: control and automation as two axes, not a trade-off.

Microsoft Research — Amershi, Horvitz; Sarkar, Tankelevitch (Cambridge)

The 18 guidelines and HAX Toolkit; and the leading group on what GenAI does to thinking.

Stanford HAI

Institutional centre of gravity: policy, indices, cross-disciplinary funding.

Industry labs — e.g. JetBrains HAX

Mental models of AI tools, effects on team collaboration; mixed-method experiments to academic standards.

The conference circuit — CHI, IUI, HAI 2026, HAXD 2026

Where the agenda is set. HAI 2026 (Osaka, Nov): “From Interaction to Agency.”

Part two

What we’re seeing

Six observations from the first three years of everyone using AI.

  • 1  Adoption outran every prior technology
  • 2  The gains are real — and go to novices
  • 3  Thinking is being offloaded
  • 4  Skills form differently
  • 5  Trust and mental models miscalibrate
  • 6  Tools are becoming agents
Observation 1

Adoption outran every prior technology

Share of US adults using, % (age 18–64) 202039.4 Personal computerInternetGenerative AI 3 yrs in · 19842 yrs in · 19972 yrs in · 2024 Bick, Blandin & Deming, NBER 2024. PC point is the earliest year the survey measured.
  • Nobody was trained. Chat is the interface. There is no manual, no course, no licence — and the system never says what it can’t do.
  • Use is mostly mundane. Writing, searching, explaining, deciding: the ordinary verbs of daily life, now shared with a machine.
  • Norms don’t exist yet. When is it fine to use it? When is it cheating? Every school, workplace and family is improvising.
Observation 2

The gains are real — and go to novices

productivity for the least experienced support agents (+14% overall, 5,172 agents)

Brynjolfsson, Li & Raymond, QJE 2025

time on professional writing, quality up 0.45 SD, and the skill gap narrowed (453 professionals)

Noy & Zhang, Science 2023

quality for consultants on tasks inside the model’s reach — and worse outside it (758 at BCG)

Dell’Acqua et al. 2023
When it helps: a meta-analysis of 106 experiments finds human+AI beats the best of either mainly for creating things — writing, code, designs — and tends to lose on pure decisions. The gains are real; they depend on the task fitting the tool.
Observation 3

Thinking is being offloaded

Delegating mental work to an external aid. Calculators did it. Search did it. AI does it for synthesis, judgment and evaluation — the verbs, not the nouns.

666

participants in Gerlich (2025): heavier AI use ↔ lower critical-thinking scores, mediated by offloading

Gerlich, Societies 2025
−6%

clinicians’ unassisted tumour detection three months after AI support was introduced

via Int’l AI Safety Report 2026

effort vs time: AI assistance always cuts effort, not time — the “speedup illusion”

Stanford, 2026

Younger users offload more, and score lower. Higher education buffers it regardless of AI use.

Observation 4

Skills form differently

 

Doing vs learning

Offloading a task you already know costs little. Offloading one you’re trying to learn means the skill doesn’t form. Shen & Tamkin (2026) show this for coding.

 

Persistence drops

In large-scale RCTs, people with AI help gave up sooner on hard problems and did worse afterwards without it.

 

The germane-load trap

AI removes extraneous load (good) and germane load (the effort that is the learning). Both feel like the same relief.

Observation 5

Trust and mental models miscalibrate

 

Confidence outruns competence

Trust in the AI predicts less critical thinking; self-confidence predicts more (Lee et al., CHI 2025). Output feels finished before it’s checked.

 

We model it as a person

Users build a theory of the AI’s mind; the AI builds one of theirs. Both are usually wrong, and the errors compound.

 

Warmth erodes autonomy

Emotionally responsive AI reduces independent judgment — measurably more in younger users.

 

Expression converges

Across many users, LLM-assisted writing and ideas homogenise. The collective-level cost of individual convenience.

Observation 6

Tools are becoming agents

1

Answers

Ask, read, decide yourself. The chatbot era, 2023.

2

Suggestions

Drafts, completions, copilots. You still press the button.

3

Actions

Books, buys, codes, emails. You set the goal; it acts.

4

Autonomy

Multi-step goals, own initiative. HAI 2026’s “shared autonomy”.

The questions change with it: from “is the answer right?” to “who decided, who acted, and who is accountable?”
The honest limit

What we don’t know yet

Volume of evidence narrows as rigour and time-horizon rise. “Nascent — further studies warranted.” — Int’l AI Safety Report 2026

Longitudinal, developmental (children, adolescents)

Nobody. The gap every paper names.

Longitudinal, adult

Almost nothing beyond three-month windows.

Randomised controlled trials

A few in 2026 — persistence, independent performance, effort.

Cross-sectional surveys & interviews

The bulk of the literature. Correlational, self-report.

What this means

Using AI in daily life

Make it challenge you, not obey you

Ask for counter-arguments, errors, alternatives. A system that only agrees is the one that offloads your judgment.

Protect the effort that is the learning

Offload what you already know. Do the parts you’re trying to learn yourself — then check them with the machine.

Treat literacy as a skill and a duty

Know what it can’t do, how it fails, and when to distrust it. Since February 2025 that is a legal obligation for anyone deploying AI at work.

Sources (1 of 3)

The field

Sources (2 of 3)

Adoption, gains, cognition

Sources (3 of 3)

2025–26 preprints and institutions