Designing With Uncertainty: How AI Supercharges Probabilistic Thinking — Smashing Magazine

The stark reality of this challenge was brought into sharp focus by a notable incident in 2024 involving Air Canada. A customer, seeking information on bereavement fares, consulted the airline’s chatbot. The bot, with an unwarranted air of authority, provided a refund policy that, in fact, did not exist within the company’s official terms. When the customer attempted to claim the refund based on the chatbot’s assertion, Air Canada refused to honor it, leading to a legal dispute. A tribunal ultimately ruled in the customer’s favor, underscoring a critical vulnerability: the chatbot had not made a definitive decision; it had merely predicted an answer based on patterns in its training data. The airline, however, had effectively treated this prediction as an unassailable policy.

The Air Canada Chatbot Precedent: A Wake-Up Call for Corporate Liability

The Air Canada case, adjudicated by British Columbia’s Civil Resolution Tribunal, stands as a landmark decision highlighting the burgeoning legal and ethical complexities surrounding AI-powered customer service. The tribunal’s ruling established that companies are accountable for the information disseminated by their chatbots, even if that information is erroneous and contradicts official policy. This verdict sent ripples through industries heavily reliant on AI for customer interaction, signaling a need for rigorous oversight and transparent communication regarding AI-generated responses.

The incident unfolded when the customer, Jake Moffatt, was advised by Air Canada’s support chatbot that he could claim a bereavement fare refund retroactively. Following the bot’s instructions, Moffatt submitted a claim, only to have it denied by human customer service representatives who cited the actual policy, which required such claims to be made before travel. Air Canada argued the chatbot was a separate legal entity, and its advice was akin to a third-party opinion. However, the tribunal rejected this defense, asserting that it was "reasonable for a customer to expect the chatbot to provide accurate information" and that the airline could not "distance itself from the information on its own website."

This legal precedent implies significant implications for corporate responsibility. Companies can no longer simply deploy AI interfaces and assume a reduced liability shield. Instead, they must implement robust validation mechanisms, clear disclaimers, and accessible human escalation paths to prevent AI "hallucinations" from becoming legally binding commitments. The incident also served as a stark reminder of the fundamental disjunction in contemporary AI design: probabilistic systems wrapped in deterministic interfaces. AI offers a probability, a best guess, but the user interface often presents it as an undisputed truth, leading users and organizations alike to act upon these predictions as certainties.

The Paradigm Shift: From Certainty to Likelihood

Human cognition is inherently predisposed towards deterministic thinking. We seek cause-and-effect relationships, preferring to believe that specific past actions dictate precise future outcomes. This inclination makes it challenging to embrace the ambiguity of probabilistic thinking. For instance, if a coin lands on heads 999 consecutive times, a deterministic mind might conclude the coin is rigged, expecting the 1000th flip to also be heads. A probabilistic mind, conversely, understands that each flip remains an independent event, with the 1000th flip still having an equal chance of landing on heads or tails. This latter mindset, though more challenging to maintain, is precisely what designers require in the age of AI.

Products today operate within complex, often nonlinear environments, and the accelerating integration of AI further intensifies this complexity. When designers and product teams mistakenly treat AI outputs as definitive answers rather than merely one of many plausible outcomes, they risk building fragile, brittle experiences. In high-stakes sectors such as medical diagnostics, financial forecasting, or autonomous systems, such a misinterpretation can lead to genuinely dangerous consequences, jeopardizing safety, financial stability, or even lives.

AI as a Strategic Partner, Not a Definitive Oracle

This new paradigm advocates for designing probabilistically with AI as an invaluable partner. It’s about leveraging AI to refine and sharpen human thinking, rather than outsourcing critical judgment. This involves conscientiously accounting for inherent model biases, understanding nuanced human sentiment, and appropriately assessing perceived risks throughout the design process.

Most questions posed to AI do not yield binary "yes" or "no" answers. Instead, they produce probabilities derived from patterns within vast datasets. Asking an AI, "Do aliens exist?" will likely generate a response that frames the existence of extraterrestrial life as plausible but uncertain, based on scientific consensus and lack of definitive evidence. The answer doesn’t resolve the question but frames it within a spectrum of probabilities.

Designers must adopt this same interpretive lens when evaluating AI outputs. These outputs are signals, not final conclusions. They represent possible outcomes that demand careful interpretation within the broader context of product goals, anticipated user behavior, and existing business constraints. Many successful digital products already operate on this principle. Netflix, for example, doesn’t know you’ll enjoy a specific show like Superstore just because you watched The Office. Instead, it estimates the probability of your enjoyment based on sophisticated recommendation algorithms and surfaces titles accordingly. The interface, in this scenario, is intelligently responding to a prediction, not a certainty.

This logic can be directly applied to design decisions. AI models can integrate behavioral analytics with user research insights to estimate the likelihood of various outcomes, with these probabilities serving as a crucial yardstick for design strategy. Consider a retail scenario: analytics might indicate a 60% confidence that a user will complete a purchase versus a 90% confidence. At 60% confidence, the design strategy needs to be more persuasive, incorporating elements like prominent testimonials, detailed explanations, product comparisons, and reassuring trust signals to guide the user towards a decision. Conversely, at 90% confidence, the user is already highly motivated; the design’s primary objective should shift to removing any friction, enabling a swift and effortless completion of the action. The same screen, but two vastly different design challenges dictated by probability.

Furthermore, AI can simulate outcomes based on historical data and behavioral models before committing to a particular design direction. The utility of these simulations, however, heavily depends on the structure of prompts, the defined context, the hypotheses being tested, user motivations, and the specific edge cases to be stressed. For instance, evaluating early designs through structured prompts can be invaluable, especially when direct access to the target user group is limited. A template prompt might evaluate a design for usability, accessibility, and content relevance from the perspective of neurodivergent users, providing a SWOT analysis and a probability score for successful use.

However, it is crucial to understand that simulations do not replace real-world experimentation. Because AI models are trained on historical data, they often reflect past behavior more strongly than they predict future innovation or emergent trends. Imagine designing a voice interface for elderly users who struggle with touchscreens. A model trained predominantly on mobile interaction data might predict low engagement, not because the idea lacks value, but because its dataset reflects different user behaviors. Simulations should always be used to surface assumptions and explore possibilities, never to stifle innovation.

Addressing AI’s Blind Spots: Bias and the Data Imperative

A significant cautionary tale in designing with AI is the pervasive issue of bias embedded within its training data. AI systems are fundamentally built upon historical data, and that foundation invariably shapes their outputs. India’s Prime Minister Narendra Modi once highlighted this during an AI Summit in France: if an AI model is asked to generate an image of a person writing with their left hand, the output may still frequently depict a right-handed person. This is due to a statistical reality: the vast majority of people are right-handed, and this demographic skew is reflected in the massive datasets used to train image-generating AI. While these models are continually improving, the core point remains valid: AI outputs represent the most statistically likely outcome given the available data, not an objective truth.

This means designers must constantly question whether past data meaningfully predicts future behavior. If additional context can refine a prediction, it must be included. Without sufficient context, an AI output is merely one possible answer presented as the definitive one.

The notorious case of Amazon’s experimental AI recruitment tool provides an even more vivid illustration of data bias. The company reportedly scrapped the project after discovering that the model had learned to downgrade resumes from women. The underlying issue was the training data itself—approximately a decade of historical hiring decisions, which were heavily skewed towards male candidates. Consequently, the AI began penalizing resumes that included terms like "women’s," as in "women’s chess club captain," and favored language more commonly found on men’s resumes. The system was not intentionally designed to be biased; the bias was inherited from the historical data it was fed. Despite Amazon’s attempts to adjust the algorithm, they reportedly could not guarantee the elimination of other discriminatory patterns, leading to the project’s eventual termination. This example underscores why critical interpretation of AI output is paramount. Designers must understand the provenance and characteristics of the data behind a prediction and rigorously evaluate the reliability of the models they employ. A recommendation is only as sound as the data it was trained on, and uncovering potential hidden biases requires diligent inquiry.

Confidence scores, often presented alongside AI predictions, also warrant careful scrutiny. Over-reliance on a high-confidence output risks repeating the Air Canada scenario, where a probable answer is treated as an infallible truth. Conversely, dismissing a low-confidence signal outright can lead teams to overlook genuine insights buried within noisy data. A 90% confidence prediction is not infallible, and a 40% signal is not necessarily useless. Designers must exercise human judgment, weighing possibilities, considering the specific context, and applying their expertise to what the AI suggests.

Transparency is the cornerstone of making this critical evaluation possible. As AI systems increasingly influence decisions across all sectors, users need clear visibility into how outputs are generated, the underlying sources, the reasoning process, and the summaries that inform a recommendation. Opaque, "black-box" systems inevitably breed distrust. Conversely, systems that openly reveal their reasoning empower users to evaluate outputs for themselves, fostering accountability and trust. This commitment to transparency is not just good design; it is an ethical imperative that respects the trust people place in these powerful tools.

Core Principles of Probabilistic Design in Practice

Practicing probabilistic design means recognizing that every design decision is a bet, not a guarantee. Even the most rigorous research and data-backed decisions are based on samples and assumptions about user behavior at scale. A meticulously researched idea can still falter in the real world. This mindset also inherently fosters adaptability: user needs evolve, strategies shift, and ideas occasionally fail. Teams that consistently lean on data signals, continuous experimentation, and learning loops are better positioned to pivot and converge on the most effective solutions.

1. Optimizing for Likelihood, Not Absolute Certainty: The Air Canada chatbot incident serves as a crucial design lesson. The bot generated plausible text, a probabilistic outcome. The interface, however, communicated this prediction with absolute confidence—no caveats, no "here’s what our policy usually says," no clear path to human support. The user interpreted this confidence as a firm commitment, and legally, so did the tribunal. This illustrates the danger of wrapping probabilistic systems in deterministic interfaces. Designers must avoid binary thinking; a brilliant idea doesn’t guarantee success, and a familiar idea isn’t guaranteed to fail. Instead, explore variations, confidence levels, and edge cases. AI can act as a "portfolio-thinking engine," surfacing diverse interpretations, highlighting risks, and generating structured recommendations. The ultimate goal is value-driven optimization, not absolute certainty.
2. Data as a Compass, Not a Rigid Map: Even a precise probability is not a final answer. An AI model might predict an 80% likelihood that users prefer a minimal checkout experience. This doesn’t automatically mean "build a minimal checkout." Data should serve as a compass, guiding direction, not a rigid map dictating every step. Designers must ask: Why do users prefer this? What are the underlying motivations? Are there cultural or contextual factors at play? How does this preference interact with other product goals? These questions facilitate the validation of AI predictions through usability testing and qualitative research. AI excels at identifying patterns, but understanding the human motivations behind those patterns remains a human-centered research task.
3. Experimentation as Continuous Learning: Traditional A/B testing, while valuable, can be expensive in terms of engineering time, traffic allocation, and user exposure, especially when a poorly performing variant is exposed to a significant audience. Probabilistic thinking reframes experimentation not merely as a validation tool, but as a system for reducing uncertainty. AI simulations can efficiently filter weaker ideas before they consume significant production resources, acting as a hypothesis filter and directing engineering efforts towards the most promising directions. User needs are fluid, and effective teams must iterate rapidly. This approach also supports personalization, as different user segments may respond better to varied experiences. Multiple experiences co-existing are not a flaw but a deliberate strategy. Experimentation becomes a continuous feedback loop: Predict → Test → Learn → Adjust → Repeat!
4. Transparent Communication of Uncertainty: One of the most challenging tasks for designers is to make uncertainty both understandable and actionable. When uncertainty is concealed, users tend to treat AI outputs as factual. When communicated clearly, however, trust actually increases. Using ranges, estimates, and confidence indicators is crucial. A delivery window of "Friday to Monday" honestly conveys variability, whereas a precise timestamp that repeatedly slips erodes user trust. A face recognition feature that asks, "This looks like Pratik, is that right?" sets more realistic expectations than one that simply labels a photo with a name. Communicating uncertainty doesn’t weaken trust; it strengthens it by setting honest expectations. Designers must also consider how different user types react to uncertainty: over-trusting users need more prominent displays of uncertainty, distrustful users benefit from historical accuracy or confidence levels, and skeptical/balanced users need AI assistance reinforced while retaining ultimate decision-making.

The Indispensable Human Element: Keeping Judgment in the Loop

AI should always augment human judgment, not replace it. The most trustworthy systems are designed with explicit points where humans can review, challenge, correct, or override machine suggestions. Human-in-the-Loop (HITL) is not merely a safety net; it is a refinement engine. Every human override, correction, or rejection becomes invaluable, high-quality feedback that continuously improves the underlying AI model.

Control is a fundamental prerequisite for user adoption. Users are more willing to rely on AI when they understand how a suggestion was generated, can evaluate its implications, and can easily intervene. Well-designed products make this explicit: identifying who is acting, what happens if a suggestion is incorrect, and where the user can step in. These interactions are also critical for system improvement; every accept, reject, or edit provides a strong signal, producing far more meaningful training data than passive analytics alone. This closes the vital feedback loop between real-world usage and model performance.

In practice, HITL manifests in various ways. GitHub Copilot, for instance, offers inline code suggestions that developers can accept with a tab, edit, or ignore entirely. The system never commits code on its own behalf; authorship remains with the human developer. Similarly, Gmail’s Smart Compose presents predicted text as an optional suggestion, keeping the user’s tone and intent firmly in their hands. In higher-stakes contexts, HITL becomes more explicit. Risk and fraud detection systems often use probability scores to route decisions: low-risk transactions proceed automatically; medium-risk triggers additional verification; and high-risk escalates to a human reviewer. This approach expertly balances speed with human judgment without sacrificing oversight. In safety-critical domains like healthcare, human oversight is non-negotiable. AI may flag anomalies or suggest a diagnosis, but the clinician always retains final authority. Tools that explain the reasoning behind AI recommendations help practitioners understand why a suggestion was made, reinforcing confidence without diminishing accountability.

From a UX perspective, HITL is about matching the interaction pattern to the level of risk. Simple accept/reject affordances suffice for low-risk suggestions that primarily enhance speed. As the stakes rise—impacting data, finances, or human lives—preview and explicit approval steps become essential. Explanations enable users to calibrate their trust rather than blindly accepting outputs. Behind the scenes, the system must capture user decisions with context, feed them into learning workflows, and log overrides for auditability. Over time, teams can track metrics like override rate, confidence accuracy, time-to-approval, and perceived trust. A high override rate is not a user failure; it signals that the design or the model itself requires attention.

However, poorly implemented HITL systems can fail in subtle ways. Human review can devolve into a mere rubber stamp, or workflows can become so cumbersome that users bypass safeguards. Feedback loops can also become skewed towards a narrow subset of users. These risks are real but represent design challenges, not justifications to eliminate HITL. The goal is not to maximize human involvement but to strategically focus it where uncertainty, impact, or ethics demand it. Keeping HITL is ultimately about clarity: clarity regarding who decides, when uncertainty matters, and how responsibility is shared between people and machines.

Building Resilient Systems for a Dynamic Future

Good design adapts as the landscape shifts. Product design, particularly for AI-powered systems, can no longer afford to optimize solely for short-term conversion metrics. User intent is fluid, environments change rapidly, and probabilistic systems continuously evolve. What functions effectively today can quietly break tomorrow. Designing for resilience means building products that remain reliable, trustworthy, and useful even as core assumptions, underlying data, and user behaviors change.

Resilient design fundamentally shifts the core question from: How do we maximize this metric right now?! to: How does this system behave over time, under stress, and amidst uncertainty? A resilient system is characterized by its ability to:

• Maintain functionality despite unexpected inputs or shifts in context.
• Gracefully degrade rather than catastrophically fail when AI confidence is low or data is insufficient.
• Continuously learn and adapt from new data and user feedback.
• Transparently communicate its limitations and uncertainties.
• Provide clear pathways for human intervention and correction.

This means looking beyond last quarter’s numbers and proactively anticipating future shifts to make necessary adjustments.

Building Systems That Adapt as Probabilities Change: Likelihoods are constantly in flux, AI models inevitably drift, contexts evolve, and user needs mature. Designing as if conditions are stable creates inherent fragility in probabilistic environments. A resilient approach assumes volatility as the default state. Consider how recommendation systems evolve: an early version might optimize solely for engagement, and for a period, engagement rises. However, users might eventually perceive the feed as narrow, repetitive, or even exhausting. Resilient systems proactively rebalance, introducing novelty, diversifying signals, and integrating long-term satisfaction metrics alongside short-term clicks. Designers must create interfaces that anticipate change, incorporating dynamic re-ranking, contextual explanations, and "escape hatches" from stale personalization loops, ensuring systems remain useful as probabilities shift.

Optimizing for Long-term Outcomes, Not Just Short-term Wins: Short-term conversion gains often mask significant long-term costs. Accelerating onboarding might reduce comprehension. Maximizing notification click-through rates can erode user trust over time. Optimizing solely for engagement can lead to unhealthy usage patterns. Fragile systems prioritize immediate numbers while neglecting second-order effects—the downstream consequences that manifest weeks or months later. Duolingo’s "hearts" system is an example of designing for long-term resilience. It introduces friction: too many mistakes deplete hearts, requiring users to wait or practice older material to earn more. On paper, this might seem like a conversion killer, reducing lessons per session. In practice, as the Duolingo team has publicly discussed, it supports long-term motivation and retention—the metrics that truly matter for a learning application. Short-term engagement might dip, but long-term outcomes improve. Similarly, Meta (Facebook) has, albeit reluctantly, acknowledged that optimizing purely for "time spent" produced unintended emotional and societal effects, leading to a stated pivot toward "meaningful social interactions." This recognition, whether fully realized or not, underscores the critical point: optimizing for the wrong thing at scale carries substantial downstream costs. Designers must routinely ask: What are the second-order effects of this decision? How might this impact user trust or long-term engagement? What are the potential ethical implications of this optimization?

Planning for Uncertainty the Way You Plan For Scale: Teams routinely plan for traffic spikes and system scalability, but rarely for "uncertainty spikes." Yet, AI systems can degrade, adversarial behaviors evolve, and external shocks can reshape user behavior overnight. Resilient design anticipates variability and prepares for it. This means designing for degrading confidence. What does the interface do when the AI isn’t sure? Does it silently fail, or does it gracefully hand off to a human? Does the user experience remain coherent if AI assistance completely disappears? A robust fallback strategy is as crucial as designing for the "happy path." Practical actions include: implementing dynamic confidence indicators, providing clear human escalation paths, developing graceful degradation modes, and establishing continuous monitoring for model drift and performance anomalies.

Conclusion: Embracing Nuance in the Age of AI

If there is one overarching takeaway from this exploration into probabilistic design, it is this: Stop asking "Will this work?" and start asking "How likely is this to work, and what happens when it doesn’t?" This singular reframe fundamentally alters how hypotheses are formulated, how AI outputs are interpreted, how experiments are scoped, and crucially, how products are designed for the inevitable moments when the system is wrong. Moving forward, teams should commit to explicitly naming the assumptions behind every AI recommendation they accept. They should actively identify instances where probabilistic outputs are presented as certainties within their products, rectify the framing, and design robust fallback mechanisms before perfecting the "happy path."

The shift from deterministic to probabilistic design is less about adopting new tools and more about cultivating a new posture—a fundamental change in mindset. AI has not introduced uncertainty into our world; it has simply made the inherent uncertainty that always existed impossible to ignore. AI can estimate, simulate, and recommend with unprecedented speed and scale, but it cannot definitively decide what truly matters, which user groups are being overlooked, or which unconventional idea is worth defending against a model trained on yesterday’s data. These remain uniquely human responsibilities. In an era where prediction is becoming increasingly commoditized, and sound judgment is a rare and invaluable asset, the most profound contribution a designer can make is to continually ask, "What else might be true?" The future of design lies in thinking in ranges, not points; in testing assumptions, not just features; and in building for adaptation, not an elusive perfection.