Creating Effective Explanations

AI explanations fall into 2 categories serving different needs.

• General system explanations describe how the entire AI behaves, regardless of individual inputs. General explanations might state that a recommendation engine uses viewing history and ratings to suggest movies. This helps users form mental models of the system's approach. These explanations remain consistent across all users, providing stable understanding. General explanations work in onboarding and documentation.
• Specific output explanations clarify why the system produced a particular result for a particular user. Specific explanations dig into individual cases. They might reveal that a movie was recommended because you watched 3 similar thrillers last week. These change with each interaction, connecting user actions to system responses. Specific explanations belong in the moment of interaction when users question results.

Context determines the choice. New users benefit from general explanations that establish baseline understanding. Experienced users troubleshooting unexpected results need specific explanations. High-stakes decisions often require both. Many successful AI products layer both types, starting broad and allowing drill-down into specifics.

Partial explanations clarify key elements of how AI works while leaving out parts that are unknown, highly complex, or not useful for the immediate task. They focus on factors that matter for user decisions.

Consider a news app showing "Choose topics to see stories curated just for you." This explains that personalization happens through topic selection without detailing recommendation algorithms. Users understand how to influence their feed without needing technical knowledge.

Similarly, a weather chatbot might say "I need your location to provide accurate weather." This reveals a data requirement without explaining how it processes geographic and meteorological data. These explanations reveal just enough to be helpful. Too much detail overwhelms. Too little leaves users guessing. The key is identifying what helps users make decisions.^[1]

Influential feature explanations describe which key factors or data inputs most significantly influence an AI outcome. Simple models like regression can often reveal which data features had the greatest influence on the system's output.

A resume screening AI might indicate that work experience and relevant skills were the primary factors in its recommendation, while education level had moderate influence. This helps job seekers understand which parts of their application carried the most weight. Feature explanations can also show chains of influence. A travel recommendation system might reveal that your selected month and activity preferences directly shaped its suggestions. Choosing "December" plus "outdoor adventures" leads to ski resort recommendations, while "December" plus "beach relaxation" suggests tropical destinations.

For complex models like language generators, the influential features often involve patterns from training data. A writing assistant might explain that formal business documents in its training influenced its professional tone suggestions, making the connection clear without technical details.

The key is showing users which inputs matter most for their specific results. This transparency helps them understand the AI's reasoning and make better use of the system.^[2]

Contrastive explanations clarify why AI made one decision over another by comparing what was used versus what wasn't. This comparison approach feels more natural than absolute statements. A job matching system might explain: "This position was recommended because of your Python expertise, not your management experience." Users immediately understand both what mattered and what didn't for this specific match. Counterfactual explanations show how different inputs would change AI outcomes. A writing assistant might demonstrate: "Selecting 'formal' tone generates a polished, professional definition while 'friendly' produces a conversational explanation with metaphors and humor." Users see direct cause-and-effect relationships between their style choices and the AI's responses.

The power lies in revealing alternatives. Users understand not just what is, but what could be.

Every AI prediction is based on data, making data sources a crucial part of explanations. 3 key aspects matter:

• Scope: overview of data collected
• Reach: whether personalized or aggregated
• Removal: whether users can delete or reset data

Consider a music app suggesting "Time to wind down" playlists at 9pm. Users might wonder how it knows their schedule. Explaining "Based on your typical listening patterns showing calmer music after 9pm" clarifies the data connection. Without this transparency, personalization feels creepy.

Data source explanations help users know system limits. If a fitness app only tracks steps and heart rate, users understand why it misses their weightlifting when calculating calories. This prevents over-trusting incomplete assessments.

Feature explanations show which data mattered most. A job matching AI might explain: "Match score based on skills overlap (45%), experience level (30%), location preference (25%)." This helps candidates understand why certain jobs ranked higher.

Privacy requires proper infrastructure beyond basic policies. Be transparent about data use while respecting boundaries.

Model confidence indicates how certain the system is in its prediction. 4 main approaches exist:

• Categorical displays group confidence into buckets. A weather app might show sunny icons for high-confidence predictions but cloudy icons with question marks for uncertain forecasts. This simplifies decisions without requiring probability understanding.
• N-best alternatives show multiple possibilities. A bird identification app might display "Likely: Robin, Cardinal, or Sparrow" rather than picking one. This excels when confidence is low, prompting users to apply their own judgment while revealing system thinking.
• Numeric displays show percentages but assume probability comprehension. AI rarely predicts with 100% confidence, potentially confusing users expecting certainty. Context matters: 85% confidence in a game recommendation differs from 85% in medical screening.
• Visual approaches use graphics intuitively. Stock predictions might show expanding forecast ranges over time, with narrow bands for tomorrow but wide spreads for next month. This communicates uncertainty naturally.

Testing different displays early reveals what resonates with your specific users and use cases.

Example-based explanations show users similar cases to help them understand AI decisions. Instead of explaining how the AI works, they show examples that make the logic clear. These explanations work because humans are naturally good at spotting patterns and making comparisons.

Real estate apps demonstrate this approach perfectly. When estimating a home's value at $1M, the app displays actual comparable homes nearby. Users see a 5-bedroom house at 1115 Berkeley Ave that sold for $1,180,000 and a 3-bedroom at 1216 Windermere that sold for $1,650,007. This makes the valuation understandable without any technical explanation.

The power comes from selection. Showing random similar homes won't help. The examples need to highlight what drives the price difference. Is it the number of bedrooms? The square footage? The neighborhood? Good examples make these factors obvious through comparison.

Users trust what they can verify themselves. Seeing real sold prices from actual addresses feels more convincing than any algorithm explanation.

Interactive explanations let users experiment with AI to understand how it works. Instead of reading about the system, users try different inputs and see results change instantly. This hands-on approach builds deeper understanding.

Photo editing apps show this well. Users can test AI filters on sample images before using their own photos. Trying "Vintage Film" versus "Professional Portrait" reveals each filter's effects without risk. They learn by doing, not reading.

Financial tools excel here too. Users move sliders for retirement age or savings rate and watch projections update. Changing retirement from 65 to 70 might add $200,000 to savings. These instant updates make abstract concepts concrete.

Speed matters. Results must appear immediately after changes. A recipe app that slowly recalculates serving sizes breaks the learning connection. Users need that instant feedback to understand relationships.

Good design provides enough options to explore meaningfully without overwhelming users. A language app might offer speed and accent controls while hiding complex phonetic settings. This focused approach keeps experimentation productive and enjoyable.

Model metrics can explain AI outcomes, but technical numbers confuse users. The solution is translating these measurements into everyday language that helps people make decisions.

A document search showing "0.73 cosine similarity" means nothing to most users. But "73% relevant to your search" makes immediate sense. Movie apps do this well by showing "85% match" instead of complex filtering scores.

Generative AI needs different approaches since single accuracy numbers don't capture quality. A coding assistant might create "helpfulness scores" based on developer feedback. Seeing "92% helpful" gives users meaningful context about whether to trust a suggestion.

Choose metrics that connect to user goals. News apps benefit from "freshness scores." Dating apps use "compatibility percentages." But showing "perplexity values" for text generation helps no one. Visual elements improve understanding. Stars, progress bars, and color coding communicate better than numbers alone. Keep these consistent across your product so users learn what to expect.

Not every AI feature needs explanation. When systems work exactly as users expect, explanations become annoying interruptions rather than helpful information.

Spell checkers demonstrate this perfectly. When "teh" becomes "the," nobody needs to know why. The correction matches what users want and expect. An explanation would waste time on something already understood.

Some explanations could harm more than help. They might reveal company secrets or user data that should stay private. A small "AI-enhanced" label might be enough without details. Consider interface complexity too. A camera app with plant identification and video stabilization shouldn't explain everything at once. Wait until users actually use a feature before explaining it. Context determines when silence works better than words.

The best AI explanations focus on what users need for decisions, not technical completeness. Different situations require different detail levels, from brief hints to full documentation.

Most cases need focused explanations. A social media feed might explain it shows "posts from friends you interact with most" without listing every ranking factor. Users get enough to understand their feed without overwhelming detail. The detailed explanation would distract from the main task.

High-stakes situations demand transparency. AI systems helping judges with sentencing must explain everything. This context requires full documentation including data sources and decision thresholds.

Testing reveals the right balance. Watch real users to see where they get confused or overwhelmed. Different audiences often need different explanation depths. Build flexibility into your approach rather than forcing one size on everyone.