AI reporting frequently applies simplistic tropes for AI systems.

• “white robot” (technoutopian)
• “terminator” (technodystopian)

These tropes mirror people, not computer systems. White robots are naive but noble people (can be turned to evil), terminators are evil people (can be redeemed)

Supertool

• Shneiderman’s “supertool” enables us to consider how AI fits into existing tool use.
• Not a complete solution, still lots of questions and unknowns!

Three big ideas:

• HCAI framework: guide human-centric thinking for creative design
• design metaphors: practical ideas for designing HCAI systems (not white robots)
• governance structures: practice steps to realise ethical principles

Support:

• design aspirations
• individual goals
• human values (most important!)

Shneiderman (2022) argues against AI goals that equate computers with people.

• people are already good at lots of things (celebrate that)
• argues to change from simulating human behaviour to enhancing it.
• As an example:
  • people are good at emotions, and sensitive to them.
  • people can be disturbed by emotional simulations in robots

Rather than anthropomorphic systems: argues for Creativity Support Tools (CSTs) (see Expressive Interaction), direct manipulation interfaces (see Interfaces)

• Shneiderman (2022) critiques standard models of autonomy/automation by computers.
• e.g., Parasuraman et al. (2000) model of levels of autonomy.
• typically, autonomy is seen as 1D spectrum
• so if the computer is autonomous, the human does nothing.

What if both computer and human can have autonomy?

• reliable: produce expected responses when needed
  • comes from good engineering, verification, validation
  • technical audit trails when things go wrong
• safe: social/cultural commitment
  • reporting failures and near misses
  • review of problems and solutions
• trustworthy: what people want from a system
  • independent oversight (certification, regulation, insurance)
  • (communication and understandable context)

• existing 1D framework poses an unhelpful zero-sum tradeoff in automation
• problem: full automation seems like the only goal
• 2D framework: computer and human control can both be useful in different contexts
• helps understand existing systems, and find dangerous areas

1. Overview first, zoom and filter, then details-on-demand
2. Preview first, select and initiate, then manage execution
3. Steer by way of interactive control panels
4. Capture history and audit trails from powerful sensors

5. People thrive on human-to-human communication
6. Be cautious when outcomes are consequential
7. Prevent adversarial attacks
8. Incident reporting websites accelerate refinement

Science and innovation goals can come into conflict. Significant amounts of design sometimes needed to turn a scientific output into a product.

Autonomous social robots

• Science goals: a general purpose robot elders, parcel delivery, etc.
• Innovation goals: tune solution for each context of use.

Online meeting services

• Science: devices / software support collaboration.
• Innovation: Microsoft Teams, Zoom, Google meet.

• early computers were labeled “thinking machines” and “electronic brains”.
• Martin (1993) argues that such myths slowed workplace acceptance and created unrealistic expectations
• Turing (1950) asked “Can Machines Think?”
  • Envisioned machines competing with humans in intellectual fields
• debate: a deceptive metaphor rather than a meaningful test (Natale, 2021)
• pop culture influence of powerful thinking machines and robots

• Teammates: modelling human-robot interaction after human-human interaction
• Telebots: remote-controlled systems with high degrees of automation but ultimately human control

Teammate approach has many challenges (Klien et al., 2004):

• Unmet expectation to disappointment.
• False beliefs about robot autonomy/responsibility.
• Emotional attachment can mislead usage.

• Human operators remain accountable (e.g., remote surgery)
• Leverage what machines do better, don’t copy humans
• Responsive controls (e.g., DaVinci surgical system)
• Enable users to fix, personalise, provide feedback for future design iterations
• incorporating human creativity
  • tools should support human innovation, not replace it.
  • enable users to fix, personalise, provide feedback for future design iterations.

Question: is agentic coding AI (e.g., CoPilot) a telebot or teammate?

Autonomy: delegation to an authorised entity to take action within specific boundaries (David & Nielsen, 2016)

• autonomous systems are hard and interesting to design
• but when autonomy falls short, frustration is enhanced
• problems can include: reduction in human attention and awareness
• ironically, workload can increase, more vigilance is required

Assured Autonomy

• “assured” with HCAI, formal methods for proving correctness, testing and certification
• so far, a human operator is still ultimately responsible

• Control panels and remote control centers.
• Visual monitoring, audit trails, and feedback.
• Retrospective analysis of failure data.

Examples:

• Aviation: pilots, co-pilots, TRACON, ARTCC, FAA certification, NASA Rovers.
• Healthcare: ICU monitoring.
• Social media / e-commerce: alerts, feedback, interlocks.

What kind of AI systems should we live with?

• Human-like figures from stories and myths (e.g., golem, Frankenstein’s creation)
• Repeated failures: gimmicky, unrealistic addressing real needs.
• Animal robots (e.g., Sony AIBO) offer emotional connection without promising intelligence
• Simpler, clearer use cases (e.g., therapy, companionship).
• Avoids ethical concerns tied to humanoid robots.

Good example of a successful household robot: washing machine.

• dishwashers, pool cleaners, security systems
• integration of sensors, automation, and machine learning.
• improved user interfaces (e.g., smartphone apps, touchscreens)
• increasingly robot-like without anthropomorphic form

Challenges:

• lack of standardised controls across devices.
• interface consistency, especially for: setting timers, health monitoring devices.

• tensions between the scientific goals of AI/ML and innovation goals of HCI community
• possible to bring together scientific and innovation goals to develop new designs
• autonomy does not have to be a zero-sum 1D spectrum
• pop-culture reflects us, not always a good predictor for best ideas

Alan Kay (designed first windowed UI at Xerox PARC) said:

the best way to predict the future is to invent it

so go do that!

• Effective to use (effectiveness)
• Efficient to use (efficiency)
• Safe to use (safety)
• Having good utility (utility)
• Easy to learn (learnability)
• Easy to remember how to use (memorability)

This version from: (Rogers et al., 2023)

1. Discover: understand the problem and the people affected
2. Define: define the problem clearly so that it can be addressed
3. Develop: create ideas, prototypes, sketches, etc, that might address the problem
4. Deliver: test potential solutions to find promising directions, and iterate

• quick
• timely
• inexpensive
• disposable
• plentiful
• clear vocabulary
• distinct gesture
• minimal detail
• appropriate degree of refinement
• suggest and explore rather than confirm

(Buxton (2007), p.111-113)

• “primitive form”
• the form that comes before… something.
• in this context:
  • a testable form
  • a form we can experience
• enables evaluation and iteration
• primitive: should be somehow rough or limited

• Interview techniques: structured, semi-structured, open
• Questionnaires: closed, open, rating scale questions
• Established questionnaires: Software Usability Survey (SUS), NASA Task Load Index (TLX), Creativity Support Index (CSI)
• DIY questionnaires can be tricky to do well!
• All useful, but need to be justified
• Require different types of analysis, both can involve quantitative and qualitative.

• descriptive statistics
  • minimum, maximum
  • lower and upper quartile
  • median and mean
  • number of data points (count)
• plot distribution
  • scatter plot: see all the data! good for checking outliers and comparing aspects of data
  • boxplot: useful to compare distributions clearly charles approved plot!

These approaches may be enough to make clear research findings!

Lots of qualitative techniques but our focus is (Reflexive) Thematic Analysis (RTA) (Braun & Clarke, 2022), a well-known and accessible methodology.

1. familiarise with the data
2. coding (short labels, multiple rounds)
3. generating initial themes (higher level than codes)
4. developing, reviewing, and refining themes

Your themes become the findings of your qualitative analysis.

There are different types of themes, and a common distinction:

• Themes that categorise groups of codes: bucket themes, semantic themes, thin themes
• Themes that interpret the codes, revealing hidden information: latent themes, thick themes

Charles (2025; i.e., these slides!) suggests that 4 is a key heuristic for assessing theme thickness. (Disclaimer: may be revised in future!)

Number of words heuristic:

If your theme is <4 words, it might be a bit thin.

Number of themes heuristic:

If you are proposing >4 themes, they might be a bit thin.

Source: Charles, 2025. 😬

• Command Line
• Graphical
• Multimedia
• Virtual reality
• Web
• Mobile
• Appliance
• Voice
• Pen
• Touch
• Touchless

• Haptic
• Multimodal
• Shareable
• Tangible
• Augmented reality
• Wearables
• Robots and drones
• Brain-computer
• Smart
• Shape-changing
• Holographic

Cognitive Processes (Eysenck & Brysbaert, 2023):

1. Attention
2. Perception
3. Memory
4. Learning
5. Reading, speaking, listening
6. Problem solving, planning, reasoning, decision making

Social and Emotional aspects

• conversation (face-to-face vs remote)
• co-presence
• Emotions and behaviour relate
• Models of emotional design
• Affective Computing and Emotional AI
• Persuasive Technologies
• Anthropomorphism

• Evaluation Goal/Aims
• Participants
• Setting
• Data to collect
• Methods
• Ethical Considerations and Consent
• Data capture, recording, storage
• Analysis method
• Output(s) of evaluation process

• going beyond descriptive statistics…
• significance testing: quantifying differences in mean
  • t-tests: for comparing two means
  • ANOVA: for comparing 3+ means, incl. repeated measures
  • non-parametric alternatives: Mann-Whitney U, Wilcoxon signed ranks
  • χ² test: comparing categorical data
• correlation analysis
• regression

One-way ANOVA:

from scipy.stats import f_oneway
import statsmodels.api as sm
from statsmodels.formula.api import ols

# group by 'independent' column and compare dependent column
groups = [group['dependent'].values for _, group in df.groupby('independent')]
f_stat, p_value = f_oneway(*groups)

# create a Model from a formula and dataframe and run anova on that
model = ols('dependent ~ C(independent)', data=df).fit()
anova_table = sm.stats.anova_lm(model, typ=2)

Factorial ANOVA:

# factorial anova: example effects of two independent variables and their interaction
# model: tempo ~ key + mode + key:mode
model = ols('dep ~ C(ind_1) + C(ind_2) + C(ind_1):C(ind_2)', data=df).fit()
anova_table = sm.stats.anova_lm(model, typ=2)

Mapping sensed gestures to an expressive output that is fed back to the user.

• gestures: the use of motions by the limbs or body as a means of expression
• can be unintentional, control, or ancillary gestures
• from non-human actors (e.g.,the movement of a leaves on a branch of a tree)
• “any sort of motion, that may be understood as an expression of something”

The interaction itself is expressive, and the output is an expression as well. We consult Composing Interactions (Baalman, 2022) as a resource.

We got through a lot this semester!

Thanks for coming on this journey with me!

Good luck with your final projects and your other assessments this semester!