22 spatialized frameworks to explore the biggest question of human-centered AI

• Interactive machine learning — Interactive Machine Learning
• Users are people, not oracles — Interactive Machine Learning
• People give more positive than negative feedback — Interactive Machine Learning
• People want to give more than labels — Interactive Machine Learning
• Transparency helps people teach — Interactive Machine Learning
• Support assessment of model quality — Interactive Machine Learning
• Time queries to the user appropriately — Interactive Machine Learning
• Enable richer ways to steer the learner — Interactive Machine Learning
• Make clear what the system can do — Microsoft HAX Guidelines
• Make clear how well the system can do what it can do — Microsoft HAX Guidelines
• Time services based on context — Microsoft HAX Guidelines
• Show contextually relevant information — Microsoft HAX Guidelines
• Match relevant social norms — Microsoft HAX Guidelines
• Mitigate social biases — Microsoft HAX Guidelines
• Support efficient invocation — Microsoft HAX Guidelines
• Support efficient dismissal — Microsoft HAX Guidelines
• Support efficient correction — Microsoft HAX Guidelines
• Scope services when in doubt — Microsoft HAX Guidelines
• Make clear why the system did what it did — Microsoft HAX Guidelines
• Remember recent interactions — Microsoft HAX Guidelines
• Learn from user behavior — Microsoft HAX Guidelines
• Update and adapt cautiously — Microsoft HAX Guidelines
• Encourage granular feedback — Microsoft HAX Guidelines
• Convey the consequences of user actions — Microsoft HAX Guidelines
• Provide global controls — Microsoft HAX Guidelines
• Notify users about changes — Microsoft HAX Guidelines
• CrowdTruth: disagreement is signal — CrowdTruth / Seven Myths of Annotation
• Myth One: One Truth — CrowdTruth / Seven Myths of Annotation
• Myth Two: Disagreement Is Bad — CrowdTruth / Seven Myths of Annotation
• Myth Three: Detailed Guidelines Help — CrowdTruth / Seven Myths of Annotation
• Myth Four: One Is Enough — CrowdTruth / Seven Myths of Annotation
• Myth Five: Experts Are Better — CrowdTruth / Seven Myths of Annotation
• Myth Six: All Examples Are Created Equal — CrowdTruth / Seven Myths of Annotation
• Myth Seven: Once Done, Forever Valid — CrowdTruth / Seven Myths of Annotation
• A constitution of explicit principles — Constitutional AI
• Critique and revision — Constitutional AI
• Reinforcement learning from AI feedback (RLAIF) — Constitutional AI
• Harmless but not evasive — Constitutional AI
• Chain-of-thought for legible decisions — Constitutional AI
• Principles chosen ad hoc need wider input — Constitutional AI
• Over-optimization and reduced human oversight — Constitutional AI
• Selecting representative humans and getting quality feedback is difficult — Open Problems of RLHF
• Evaluators can hold harmful biases or poison data — Open Problems of RLHF
• Limited time, attention, care and observability — Open Problems of RLHF
• Humans cannot evaluate performance on difficult tasks well — Open Problems of RLHF
• Humans can be misled, so their evaluations can be gamed — Open Problems of RLHF
• Data collection can introduce harmful biases — Open Problems of RLHF
• Cost/quality and richness/efficiency tradeoffs in feedback — Open Problems of RLHF
• Human values resist a single reward function — Open Problems of RLHF
• Static benchmarks fall short — Chatbot Arena
• Anonymous pairwise votes on users' own prompts — Chatbot Arena
• Statistical ranking with active sampling — Chatbot Arena
• Live prompts have a long tail — Chatbot Arena
• Crowd votes track expert judgments — Chatbot Arena
• Flag anomalous voters — Chatbot Arena
• Whose preferences the leaderboard reflects — Chatbot Arena
• Feedback — Co-Creative Loops
• Co-Creative Loops Framework — Co-Creative Loops
• Generation — Co-Creative Loops
• Iteration — Co-Creative Loops
• Co-Creative Loop — Co-Creative Loops
• Begin with the AI Last Approach — Co-Creative Loops
• Personalize your Scale of an Iterative Loop — Co-Creative Loops
• If you don't know how it works, you're in a bad place — Co-Creative Loops
• Become an Expert in Cross-Skilling and Agent Management — Co-Creative Loops
• Intentionally Decide what the Dirty Work is — Co-Creative Loops
• Your Wrong Answer is Better than AI’s Right Answer — Co-Creative Loops
• Treat yourself like a generative algorithm — Co-Creative Loops
• Always cite the Models you use — Co-Creative Loops
• Learn the tools of the craft, but try building your own tools — Co-Creative Loops
• Human-AI Overlapping Signals are Critical — Co-Creative Loops
• Use both Coordinate-Based & Coordinate-Less AI Feedback — Co-Creative Loops
• Novelty is not the metric. Use AI to appreciate the obvious — Co-Creative Loops
• AI Emotional Support isn't Strange — Co-Creative Loops
• Reward functions are hard to specify — RL from Human Preferences
• Pairwise comparison of short clips — RL from Human Preferences
• Learn the reward model online — RL from Human Preferences
• Ask about the most uncertain comparisons — RL from Human Preferences
• Model labeler error explicitly — RL from Human Preferences
• Human feedback carries implicit shaping — RL from Human Preferences
• Small amounts of feedback go far — RL from Human Preferences
• A datasheet for every dataset — Datasheets for Datasets
• Serve creators' reflection and consumers' decisions — Datasheets for Datasets
• Collection Process — Datasheets for Datasets
• Composition — Datasheets for Datasets
• Distribution and Maintenance — Datasheets for Datasets
• Motivation — Datasheets for Datasets
• Preprocessing/cleaning/labeling — Datasheets for Datasets
• Uses — Datasheets for Datasets
• Developing significant value-added automation — Mixed-Initiative Principles
• Employing socially appropriate behaviors for agent-user interaction — Mixed-Initiative Principles
• Maintaining working memory of recent interactions — Mixed-Initiative Principles
• Continuing to learn by observing — Mixed-Initiative Principles
• Considering uncertainty about a user's goals — Mixed-Initiative Principles
• Considering the status of a user's attention in the timing of services — Mixed-Initiative Principles
• Inferring ideal action in light of costs, benefits, and uncertainties — Mixed-Initiative Principles
• Employing dialog to resolve key uncertainties — Mixed-Initiative Principles
• Allowing efficient direct invocation and termination — Mixed-Initiative Principles
• Minimizing the cost of poor guesses about action and timing — Mixed-Initiative Principles
• Scoping precision of service to match uncertainty, variation in goals — Mixed-Initiative Principles
• Providing mechanisms for efficient agent-user collaboration to refine results — Mixed-Initiative Principles
• Gulf of execution — Direct Manipulation Interfaces
• Gulf of evaluation — Direct Manipulation Interfaces
• Directness as low cognitive effort — Direct Manipulation Interfaces
• Semantic distance — Direct Manipulation Interfaces
• Articulatory distance — Direct Manipulation Interfaces
• Direct engagement — Direct Manipulation Interfaces
• Bridge the gulfs from both sides — Direct Manipulation Interfaces
• Limits of direct manipulation — Direct Manipulation Interfaces
• Broad coverage and recognition of incompleteness — HELM
• Multi-metric measurement — HELM
• Standardization — HELM
• Scenario taxonomy — HELM
• Automated metrics can miss real quality differences — HELM
• A living benchmark — HELM
• Transparency means appropriate understanding — Human-Centered LLM Transparency
• Complex and Uncertain Model Capabilities and Behaviors — Human-Centered LLM Transparency
• New applications and expanded stakeholders — Human-Centered LLM Transparency
• Rapidly Evolving and Often Flawed Public Perception — Human-Centered LLM Transparency
• Model reporting and publishing evaluation results — Human-Centered LLM Transparency
• Providing explanations — Human-Centered LLM Transparency
• Communicating uncertainty — Human-Centered LLM Transparency
• Pair transparency with control — Human-Centered LLM Transparency
• Model cards — Model Cards
• Disaggregated and intersectional evaluation — Model Cards
• Intended Use — Model Cards
• Factors — Model Cards
• Metrics — Model Cards
• Evaluation Data and Training Data — Model Cards
• Ethical Considerations and Caveats — Model Cards
• Scale is not alignment — InstructGPT
• Helpful, honest, harmless — InstructGPT
• Three-step RLHF pipeline — InstructGPT
• Rank many outputs per prompt — InstructGPT
• Screen labelers for sensitivity — InstructGPT
• Training and evaluation criteria can diverge — InstructGPT
• Alignment tax — InstructGPT
• Whose preferences are being aligned to — InstructGPT
• Data Collection + Evaluation — Google PAIR Guidebook
• Errors + Graceful Failure — Google PAIR Guidebook
• Explainability + Trust — Google PAIR Guidebook
• Feedback + Control — Google PAIR Guidebook
• Mental Models — Google PAIR Guidebook
• User Needs + Defining Success — Google PAIR Guidebook
• Validators need validating — EvalGen (Who Validates the Validators?)
• Criteria drift — EvalGen (Who Validates the Validators?)
• Criteria depend on outputs — EvalGen (Who Validates the Validators?)
• EvalGen mixed-initiative workflow — EvalGen (Who Validates the Validators?)
• Support iteration on criteria and implementations together — EvalGen (Who Validates the Validators?)
• Code assertions are trusted more than LLM graders — EvalGen (Who Validates the Validators?)
• Grade per criterion with adjustable tolerance — EvalGen (Who Validates the Validators?)
• Two-dimensional HCAI framework — Human-Centered AI (HCAI)
• Consistent interfaces to form, express and revise intent — Human-Centered AI (HCAI)
• Continuous display of objects and actions of interest — Human-Centered AI (HCAI)
• Rapid, incremental, reversible actions — Human-Centered AI (HCAI)
• Informative feedback for every action — Human-Centered AI (HCAI)
• Progress indicators that show status — Human-Centered AI (HCAI)
• Completion reports that confirm accomplishment — Human-Centered AI (HCAI)
• Reliable systems — Human-Centered AI (HCAI)
• Safety culture — Human-Centered AI (HCAI)
• Trustworthy certification — Human-Centered AI (HCAI)
• Gulf of envisioning — Gulf of Envisioning
• Capability gap — Gulf of Envisioning
• Instruction gap — Gulf of Envisioning
• Intentionality gap — Gulf of Envisioning
• Visually track prompts and outputs — Gulf of Envisioning
• Suggest ideas for prompting — Gulf of Envisioning
• Provide multiple outputs — Gulf of Envisioning
• Make output explainable and directly editable — Gulf of Envisioning
• Design Responsibly — Generative AI Design Principles
• Use a human-centered approach — Generative AI Design Principles
• Identify & resolve value tensions — Generative AI Design Principles
• Expose or limit emergent behaviors — Generative AI Design Principles
• Test & monitor for user harms — Generative AI Design Principles
• Design for Mental Models — Generative AI Design Principles
• Orient the user to generative variability — Generative AI Design Principles
• Teach effective use — Generative AI Design Principles
• Understand the user's mental model — Generative AI Design Principles
• Teach the AI system about the user — Generative AI Design Principles
• Design for Appropriate Trust & Reliance — Generative AI Design Principles
• Calibrate trust using explanations — Generative AI Design Principles
• Provide rationales for outputs — Generative AI Design Principles
• Use friction to avoid overreliance — Generative AI Design Principles
• Signify the role of the AI — Generative AI Design Principles
• Design for Generative Variability — Generative AI Design Principles
• Leverage multiple outputs — Generative AI Design Principles
• Visualize the user's journey — Generative AI Design Principles
• Enable curation & annotation — Generative AI Design Principles
• Draw attention to differences or variations across outputs — Generative AI Design Principles
• Design for Co-Creation — Generative AI Design Principles
• Help the user craft effective outcome specifications — Generative AI Design Principles
• Provide generic input parameters — Generative AI Design Principles
• Provide controls relevant to the use case & technology — Generative AI Design Principles
• Support co-editing of generated outputs — Generative AI Design Principles
• Design for Imperfection — Generative AI Design Principles
• Make uncertainty visible — Generative AI Design Principles
• Evaluate outputs using domain-specific metrics — Generative AI Design Principles
• Offer ways to improve outputs — Generative AI Design Principles
• Provide feedback mechanisms — Generative AI Design Principles
• Understanding what AI can and cannot do — Human-AI Design Difficulty
• Envisioning new uses of AI — Human-AI Design Difficulty
• Iterative prototyping and testing — Human-AI Design Difficulty
• Crafting interactions and addressing ethics — Human-AI Design Difficulty
• Collaborating with AI engineers — Human-AI Design Difficulty
• Capability uncertainty — Human-AI Design Difficulty
• Output complexity — Human-AI Design Difficulty
• Superficial Alignment Hypothesis — LIMA
• Quality and diversity beat quantity — LIMA
• 1,000 curated examples from mixed sources — LIMA
• A consistent response style is part of quality — LIMA
• Head-to-head human preference evaluation — LIMA
• Hand-crafted quality is costly and brittle — LIMA