Responsible AI Design: Exploring the Ethics of LLMs
Seven questions to explore the hidden costs and ethical dilemmas of building AI products.
Large Language Models (LLMs) have exploded into our workflows and products. They’re becoming faster, more powerful, and are even being given agency to complete tasks. It would be easy to imagine this powerful technology emerging synthetically in a sterile laboratory environment, like a circuit board surfacing in an etching bath. But they didn’t appear out of nowhere. They were trained on the work of artists, writers, developers, and communities. They inherited the biases of the world they learned from. They can remember more than they should, and they can confidently say things that are untrue. Meanwhile, the costs—economic, societal, environmental—are only starting to be felt.
This week, I’m taking stock of some key ethical considerations we, as designers, need to grapple with when building with or for LLMs. From copyright and bias to misinformation and sustainability, the challenges feel vast, and often uncomfortable. I don’t pretend to have the answers—or even to have asked all the right questions—but I do know these are things we can’t afford to ignore.
There’s a question that crops up when I walk through older London neighbourhoods: What do we owe the things we build on?
I see the answer in the Victorian brick foundations of iconic London train stations like Paddington or Kings Cross St Pancras—the careful stonework, the arching iron beams. These structures were shaped by the hands of skilled craftspeople and artisans—folks who laid the bricks, cut the stone, and sanded the wood smooth.
But what’s striking is how these places have been transformed—not discarded, but reimagined. The old brick façade of Kings Cross stands proud and restored, while a sleek, modern lattice of steel and glass unfurls behind it. That contrast—historic foundations supporting strikingly modern development—is a reminder that we can honour what came before while building something bold and new.
In these moments, modern architecture doesn’t erase the original work—it elevates it. The old becomes a feature, something to be preserved and celebrated, while supporting the weight of what comes next. The foundation remains, steady as ever, even as we build upward to meet future needs.
But what happens when the ground we build on becomes invisible? When it’s scraped, automated, and scaled beyond recognition? Large Language Models (LLMs) make their home here, on this strange, silent ground—there’s no surface to see or touch to connect you with their lineage. They’re trained on the accumulated works of artists, writers, thinkers, and everyday people. Bits and pieces of someone else’s labour—words written, stories told, images created, all gathered up in massive datasets.
We call this training, but it’s not always clear where training ends and borrowing begins. And if we don’t know whose shoulders we’re standing on, it becomes all too easy to look away. To let the systems spin up, hum along, and speak for us—without stopping to ask: Is this fair? Is this true? Is this good?
LLMs give us incredible tools—tools for creating, scaling, and imagining. But they also come with a weight. There are creators whose work trained them, users whose data shaped them, biases that lurk within them, and consequences that ripple out from them. These are the things we need to wrestle with if we are going to design products that do good in the world.
But, where do we start? It can feel dizzying, overwhelming, especially to those of us under pressure to deliver, or unpracticed in unravelling ethical issues. So, I propose we start here—by taking stock of the ground beneath our feet. By asking: Who built this? Who benefits? Who’s left out? What are the costs, obvious or obscure?
If we’re lucky, design gives us a chance to make things better. To build in ways that are more generous, thoughtful, and clear-eyed. But only if we pay attention.
This article explores seven ethics considerations that every designer and product builder should consider when working with LLMs:
Authorship and Copyright
Privacy
Fairness and Bias
Misinformation and Hallucination
Societal Impact
Accountability and Transparency
Environmental Impact
I’ll take a stab at exploring why each of these matters, the key questions they raise, and what we—as designers, thinkers, and builders—can do about them. Because design is about decisions. And if we’re not making these choices carefully, then someone—or something—else is making them for us.
1. Authorship and Copyright: Whose Work Built This?
The act of creation carries a lineage. A painting is never just paint on canvas—it’s years of practice, a thousand decisions, the artist’s signature voice, and the culture that surrounds it. A writer’s words are shaped by their experiences, their struggles, their craft. Creative work has an origin. It has weight. It has ownership.
And yet, in the training of large language models, much of this creative lineage gets erased. An LLM doesn’t learn from scratch; it learns from us. From artists, writers, coders, thinkers, photographers, and creators whose work was scraped from the internet, collected without their consent, and fed into these systems. These models are built on creative labor—often unpaid, often uncredited—and yet they now compete with the very people who gave them life.
This isn’t just a legal issue, though it certainly is one. It’s an ethical dilemma about value, attribution, and power.
The Dilemma
At the heart of this issue is a question of what counts as fair use. Creators are asking:
Is my style, my voice, my work something I own—or something anyone can replicate and remix without asking?
Does AI transform my work enough to justify using it without consent, or is it derived from my labour?
Take visual artists, for example. Greg Rutkowski, a fantasy illustrator whose detailed, luminous style has become synonymous with AI image generation, never consented to his art being used to train these models. Today, AI tools produce thousands of images “in his style,” often indistinguishable from his own work.
For writers, it’s much the same. An LLM can generate paragraphs that mirror a particular author’s cadence, themes, and structure. If you’ve spent years honing your voice, this feels less like innovation and more like theft—an algorithm devaluing your work and making it endlessly reproducible.
This isn’t just about individual creators. It’s about what happens to creativity as a whole. When systems learn from existing art and writing but don’t acknowledge or compensate the creators, they risk flattening the creative landscape. Styles become templates. Voices become tools. The uniqueness of human creativity becomes cheapened—processed into something commodified and impersonal.
Why It Matters
When creative labor is devalued, the effects ripple outward:
Creators lose out: LLMs built on uncompensated work generate outputs that compete with the very artists and writers who trained them. This cuts off revenue, recognition, and opportunity. What incentive is there to create if your work can be scraped, replicated, and sold back to you?
Trust erodes: AI products rely on user trust—trust in their fairness, transparency, and value. But when systems quietly exploit creative work, that trust breaks down. Users and creators alike are left asking: Whose work am I really looking at? Whose voice am I hearing?
The power imbalance grows: The benefits of LLMs flow upward—to the tech companies who profit—while the costs are pushed downward, onto individual creators. This deepens the divide between massive platforms and the people whose labor built them.
The larger risk here isn’t just to individuals. It’s to creativity itself. We risk a world where technology becomes an endless recycling machine, repackaging what already exists rather than pushing into something new. True creativity is messy, original, often time-consuming, but always deeply human. It comes with a cost to its creators. If we don’t value it, we lose it.
Key Questions for Designers
How do we honour creative labor in the products we design?
Is there a way to surface the lineage of ideas in AI outputs—like attributing inspiration or training data sources?
What would a model of compensation look like? Could creators be paid royalties, much like musicians are for streaming services?
How can we help users understand what’s behind the curtain—the creative works that enabled these tools?
Design Opportunities
While the problem is far from solved, there are pathways forward—small interventions that could begin to redress the imbalance:
Create transparency around training data
Label sources: Could we include indicators showing the types of works or styles an output is built upon? For instance:
“This output draws from fantasy illustrations and epic poetry themes.”
“Inspired by 19th-century literary texts and modern blog structures.”
Imagine this as a “nutrition label” or “ingredients list” for AI outputs—a way to contextualise where they came from.
Attribute inspiration or training sources
Credit builds awareness and begins to acknowledge the creative lineage AI builds upon. While not perfect, systems could note:
“This image is generated in a style similar to X artist.”
Explore alternative compensation models
Look to other industries for models of fair compensation:
Could creators whose work trains these systems earn micro-royalties for their contributions?
Could licensing opt-ins allow artists and writers to choose whether their work can train AI models?
If creators aren’t given tools to take control, they might choose to poison the well of training data.
Frame how users approach tools with ethical defaults
Copyright guardrails: Products can empower users to generate content without directly imitating someone else’s work—discouraging prompts like “in the style of X.”
Subtle nudges: What if “ethical defaults” nudged users toward outputs that respect creative boundaries?
How to Value Creativity
It’s easy to lose track of the human story when working with systems as vast and abstract as LLMs. But behind every dataset is a mosaic of voices, labor, and creativity—people who trusted their work would hold value, even if they didn’t foresee the digital behemoths it would help build.
Designers can help restore that value. We can remind ourselves, and our users, that creativity isn’t just a resource to be mined—it’s something to be celebrated, protected, and paid forward.
2. Privacy: What’s Being Remembered?
LLMs are not supposed to “remember” in the way humans do. Yet, they can unintentionally memorise and reproduce fragments of personal, sensitive, or proprietary information found in their training data.
But it doesn’t stop there. As LLMs interact with users, they can “learn” in real-time, either through explicit fine-tuning or implicitly within sessions. Without the right safeguards, this creates a minefield of risks. Imagine this:
An AI-powered customer service bot accidentally revealing a phone number or credit card detail that was part of its training data.
A design tool, trained on proprietary company data, leaking snippets of confidential content in outputs.
Users sharing sensitive information—financial records, medical details, login credentials—without understanding where that data goes or how it’s handled.
The Dilemma
LLMs offer immense value because they can learn from vast datasets and synthesise them into tailored, context-aware outputs. But this ability to “remember”—whether through the original training data or user interactions—can inadvertently expose information that was never meant to be surfaced. It’s not just a technical flaw—it’s a breach of trust, consent, and autonomy.
Unintentional leaks: Sensitive information like phone numbers, email addresses, or company secrets could resurface in AI-generated outputs. It could harm vulnerable populations like activists, journalists, or marginalised communities, whose political opinions or intellectual property could be exposed.
User misunderstandings: People may share sensitive data in prompts, unaware of how it’s processed, stored, or repurposed—and that there may be risks involved. If they’re unaware of how their data is used, this would also happen without their consent.
Opacity in “learning”: If models update based on user inputs, who controls what’s remembered or forgotten? The role of shaping these models becomes a form of power that must be held accountable and explainable.
The line between “learning” and “leaking” here is thin.
Why It Matters
Privacy is more than a technical issue, it’s a a fundamental human right. It’s foundational to user trust—hard-won but easily lost. Without clear protections, LLMs can:
Violate privacy laws: Data leakage or misuse can breach GDPR, CCPA, and other global privacy regulations, leading to legal and financial consequences.
Create real-world harm: Exposed personal or corporate data can damage reputations, compromise safety, or cause financial loss.
Erode confidence in AI systems: If users fear their data could be memorised, shared, or misused, they will avoid interacting with these tools altogether.
Key Questions for Designers
How can we ensure user data isn’t stored, leaked, or exposed in AI outputs in unintended ways?
Are users fully aware of what they’re sharing when interacting with AI systems? Are they aware of how their data is used and have they consented to that?
How do we handle privacy as LLMs continue to learn from user prompts, either through intentional re-training or implicitly within sessions?
Design Opportunities
How can we, as designers, build privacy into our systems—not as an afterthought, but as a defining principle?
Be transparent about how user data is processed, stored, or reused
Clear, upfront disclosures: Users shouldn’t have to hunt for privacy information buried in settings or legal jargon. Instead, surface simple, reassuring messages at the point of interaction:
“Your data is not stored.”
“This conversation is private and not used to train the model.”
Real-time visibility: Build small affordances that show what’s happening “under the hood.” For instance:
A real-time privacy indicator: “No personal data detected.”
Confirmation when user inputs are anonymised or excluded from training.
Give users control: Let people opt in or out of contributing to AI training with clear toggles. For example:
“Exclude my prompts from model improvement.”?
But more than that, we have an obligation to explain what this means in clear and understandable terms.
Include warnings or guardrails for sharing sensitive information
Proactive nudges: Tools that gently flag risky behaviour in the moment:
“This looks like sensitive information—are you sure you want to share it?”
Sanitisation of inputs: Implement features that automatically redact personal identifiers (e.g., emails, phone numbers, credit cards) before they hit the AI system.
Contextual education: Small prompts that help users make informed decisions:
“Avoid sharing passwords or private details in this chat.”
Create mechanisms to prevent data leakage and ensure compliance
Filter training data: Can systems include strict protocols for filtering out sensitive or proprietary information during model training?
Moderate outputs: Could we design moderation layers that screen AI responses in real time to catch unintentional data exposure?
System-level privacy controls: How do we ensure GDPR-style opt-ins are prominent, and users have meaningful control over their data?
Session-based privacy controls: Offer users an explicit “private session” mode, similar to incognito browsing.
Respecting Users Attention Budget
We have to remember that users bring mental shortcuts and assumptions into AI interactions. People often trust AI products to “just work” without fully understanding their privacy implications.
The challenge for designers: How do we convey complex concepts—data privacy, transparency, control—in ways that don’t overwhelm users or interrupt the flow?
The opportunity: Simplicity and nudges. Treat privacy as a feature, not a friction point. Transparency shouldn’t feel like a legal requirement; it should feel like empowerment.
Some might argue that too many warnings, prompts, or indicators create friction. That they slow down workflows and add complexity.
But here’s the reality: Good privacy design isn’t friction—it’s safety. A well-placed guardrail doesn’t disrupt the experience; it protects it.
3. Fairness and Bias: Whose Voices Are Being Amplified?
The Dilemma
LLMs are mirrors. They reflect back the world they’ve been trained on, and that world is messy: biased, unequal, and riddled with blind spots. When these systems learn from the vastness of the internet—books, blogs, tweets, videos—they inherit not just language but the cultural and societal biases embedded within it.
The result is amplification, not neutrality.
Harmful stereotypes resurface: a hiring assistant “recommends” fewer women for open roles.
Underrepresented voices are drowned out: minority languages or perspectives appear rarely, if at all, because they’re outweighed by other voices.
Dominant norms solidify: AI-generated outputs reinforce Western, English-language narratives as defaults, leaving cultural diversity in the background.
In amplifying the loudest voices, these systems risk minimising others.
But fairness is a loaded word. It’s not something we can measure with a tidy dashboard. Fairness depends on perspective: Whose voices are centred? Whose are ignored? And who gets to decide what “fair” looks like?
These aren’t theoretical concerns, they’re already quietly influencing peoples lives and livelihoods. There are already a multitude of cases where biased systems have impacted groups—from credit scores to hiring tools.
Why It Matters
When bias in AI goes unchecked, the harm isn’t abstract—it’s real, immediate, and deeply human.
For individuals: Reinforced stereotypes can limit opportunities, misrepresent identities, and deepen inequalities.
For cultures: Over time, systems that prioritise dominant perspectives erode diversity, flattening rich traditions, languages, and ways of seeing the world.
For trust: Products that fail to acknowledge and address their biases alienate users—especially those who have already been marginalised by the systems that shape their lives.
Imagine a creative AI tool consistently prioritising Western art forms while neglecting indigenous styles. Or a chatbot struggling to understand non-standard English, dismissing those voices as invalid. These failures perpetuate inequities—and worse, they normalise them.
Left unaddressed, AI becomes a tool that amplifies power imbalances rather than challenging them. Fairness isn’t a nice-to-have—it’s a moral and practical imperative.
Key Questions for Designers
How might we design tools to identify, evaluate, and reduce bias in AI outputs
Could we build systems that intentionally include underrepresented languages, cultures, and voices?
How do we communicate the limitations and biases of an AI product transparently?
Design Opportunities
Fairness starts with intentionality. As designers, we have a responsibility to ensure the systems we build don’t just work for most people but work fairly for all people.
Create signals that alert to bias
Bias detection indicators: Introduce feedback loops to help users flag biased or problematic outputs in real time:
“This response feels biased—let us know why.”
Create affordances for users to report outputs that reinforce stereotypes, lack diversity, or omit key perspectives.
Visual bias annotations: Highlight outputs where fairness is at risk. For example:
“This result might not include all perspectives.”
“Our system is more confident about Western examples. Consider exploring alternatives.”
Bias auditing tools: Enable users or admins to evaluate AI outputs for bias via dashboards that surface patterns, or tracking inclusivity metrics like gender balance and cultural diversity.
Balance underrepresented groups and perspectives
Expand training data: Advocate for datasets that actively include:
Linguistic diversity: Actively include endangered dialects, Indigenous languages, regional variations, and non-dominant forms of expression.
Cultural perspectives: Prioritise the voices, stories, art, and traditions of underrepresented groups.
Multiple forms of knowledge: Value qualitative data, lived experiences, storytelling, and community narratives as equal to quantitative metrics.
Beyond the binary: Move past reductive categorisations like the gender binary or rigid demographic classifications. Ensure that datasets reflect a spectrum of identities and lived realities.
Localised and indigenous knowledge: Prioritise these forms of expertise to ground AI solutions in community needs and lived realities.
Weight outputs towards inclusion: Build mechanisms that prioritise diversity in results. For example, a system prompt for creative outputs might include: “Explore underrepresented cultural styles or voices.”
Reflect multiple viewpoints in responses: Not all questions have a single answer, build adaptability to present multiple viewpoints and let the user choose.
Acknowledge bias where it exists
Bias disclaimers: Include upfront messaging that sets clear expectations:
“This AI system is trained on datasets that may reflect societal biases. We’re working to improve fairness, but outputs may still be imperfect.”
Confidence indicators: Highlight when outputs are likely skewed or incomplete:
“This response draws from limited cultural data and may not reflect all perspectives.”
Educational prompts: Help users understand the “why” behind bias:
“This tool struggles with some non-standard dialects because of gaps in its training data. Here’s how we’re addressing it.”
Whose Fairness Are We Designing For?
As designers, we must ask ourselves:
Whose perspectives are shaping the system?
Who has the power to decide what fairness looks like?
Who’s at the table—and who’s missing—when we define success?
The answers can be uncomfortable. Many AI tools are built by teams that lack the diversity needed to challenge their own blind spots. Fairness begins with us: our teams, our processes, and our willingness to question whose voices we amplify.
LLMs are not neutral tools. They carry the weight of history, society, and culture in their outputs. The systems we design will either reinforce existing inequities or challenge them—and that choice begins with us.
4. Misinformation and Hallucination: What Happens When Fiction Looks the Same as Truth?
The Dilemma
Large language models have a unique gift: they are masters of fluency. They write prose, answer questions, and weave stories that sound human, confident, and seamless. But beneath that polish lies a fundamental flaw: LLMs are predictors of words, not arbiters of truth.
These systems don’t “know” facts in the way we might assume. They generate outputs based on probabilities—what word or phrase is most likely to follow the last. That means they can fabricate information (“hallucinations”) or misrepresent facts in ways that sound plausible, even authoritative. A confident-sounding falsehood can mislead far more effectively than a clumsy one.
Imagine the implications:
An AI assistant confidently offers incorrect medical advice to a worried patient.
A legal tool generates historical legal cases that never happened but are repeated as fact.
A chatbot misrepresents scientific findings in ways that fuel misinformation or public distrust.
The danger grows when these systems are deployed at scale—where misinformation spreads fast and trust evaporates even faster.
Why It Matters
Misinformation isn’t a new problem, but AI makes it faster, cheaper, and more convincing. Hallucinations—innocent as they might seem—can carry serious consequences when AI outputs are trusted implicitly.
For individuals: Misinformation about health, finances, or legal advice can cause harm, sometimes irreversible.
For businesses: AI-generated inaccuracies undermine credibility and create liability risks, especially in regulated industries.
For society: Widespread misinformation erodes trust in institutions, exacerbates polarisation, and weakens our collective sense of truth.
We cannot afford for AI to be seen as “a little too creative” when it comes to facts. For users to trust these tools, they need to know when the AI is reliable—and when it’s guessing.
Key Questions for Designers
How do we ensure users understand that AI outputs are probabilistic, not definitive?
What safeguards can we build to detect and prevent AI-generated misinformation?
How can we balance creativity with factual accuracy, particularly in high-stakes domains like healthcare, education, or journalism?
Design Opportunities
As designers, our role is not to “solve” misinformation alone but to create products that are honest about uncertainty, transparent about limitations, and built with safeguards.
Ensure systems communicate their confidence (or lack thereof)
Confidence indicators: Display visual cues or confidence scores to signal uncertainty.
Uncertainty language: Build humility into the AI’s tone and messaging:
Instead of a definitive statement, use qualifiers: “It’s possible the answer is X, but I’m not certain.”
Add opportunities for users to validate: “Would you like me to cross-check this with other sources?”
Contextual warnings: In high-stakes use cases (health, finance, education), proactively remind users that outputs may not be definitive:
“AI-generated answers may contain inaccuracies. Always consult a professional for critical decisions.”
Connect AI outputs to trusted data
Source attribution: Provide citations for factual claims:
“This response is based on [Source A, Source B].”
Build outputs that cross-reference credible databases (like scientific journals, verified news outlets) and surface references in context.
Fact-checking mechanisms: Enable real-time verification within the product to flag hallucinations:
Allow users to click a “Verify” button that checks the AI’s claims against a pool of credible sources.
Highlight or filter content that doesn’t align with verified facts: “This statement may not match reliable data.”
Be clear about the role the AI is playing, or give users control
Mode switching: Allow users to choose between “Creative” and “Factual” modes. For example:
In Factual Mode, the AI prioritises reliable, grounded responses.
In Creative Mode, the AI has permission to imagine, speculate, or invent.
Make the distinction clear: “This story is fictional. Need factual information instead?”
Disclaimers for imaginative outputs: Signal when outputs are intentionally non-factual:
“This idea is speculative—let’s refine it together.”
“This story is a creative interpretation, not a historical record.”
Trusting AI Without Over-Trusting It
Language is powerful. A well-placed word can inspire, inform, or persuade. But a word can also mislead, harm, or confuse—especially when it comes from a system that doesn’t understand the difference.
The challenge for us as designers is to recalibrate the relationship:
How do we help users approach AI outputs with curiosity, but not blind trust?
How do we encourage validation, skepticism, and collaboration—without undermining the tool’s value?
This is where thoughtful design comes in. Trust doesn’t mean expecting perfection. It means knowing the boundaries, the risks, and the limitations. A user who understands when to trust and how to verify will build a healthier relationship with AI tools—and see them as partners, not oracles.
5. Societal Impact: Who Wins, Who Loses?
The Dilemma
At first glance, LLM-powered systems promise a dazzling future: productivity unlocked, creative barriers lifted, and problems solved at an unprecedented scale. But beneath the glossy veneer of “innovation,” a harder truth emerges: progress is rarely neutral.
For every task that’s automated, a worker may lose a role. For every efficiency gained, someone bears the cost. And behind the scenes of these "self-sufficient" systems lies a foundation of invisible labor: the human annotators, data labellers, and content moderators who toil to train, refine, and sanitise the AI outputs we now take for granted.
This raises uncomfortable questions:
Are we replacing human roles in the name of efficiency—or augmenting their abilities to thrive?
Who benefits most from the AI tools we design? Businesses? Workers? Or just those with the privilege to access and deploy them?
Who pays the hidden costs of AI progress—the precarious workers, the communities left behind in an ever-widening digital divide?
The result is a paradox. AI offers extraordinary potential to empower—but without care, it can deepen inequalities, entrench power imbalances, and render the vulnerable even more invisible.
Why It Matters
The impact of AI is more than economic—it’s societal. It touches livelihoods, opportunities, and human dignity.
Job displacement: From copywriters to customer service agents, LLMs can automate tasks that once sustained workers. Without re-skilling or alternative opportunities, this risks displacing millions of jobs, particularly among those already vulnerable to economic shifts.
Unequal access: The benefits of AI are often concentrated in wealthier countries and among larger businesses. Smaller organisations, marginalised communities, and those without digital infrastructure risk falling further behind.
The hidden workforce: Behind AI’s promise of “automation” are humans—millions of annotators, content moderators, and data labellers. Many work for low wages, in poor conditions, sifting through toxic content to make AI systems safer for the rest of us.
Power imbalances: AI amplifies the reach of those with access to its capabilities while widening gaps for those without—concentrating power in the hands of a few and reducing opportunities for others.
In short, if AI innovation serves only the privileged few, it ceases to be innovation—it becomes exclusion.
Key Questions for Designers
How can we design AI tools that empower, rather than exploit?
How might we design AI tools that augment human roles rather than displace them?
Could we advocate for fairer, more ethical working conditions for those involved in AI development?
How do we prioritise accessibility and equity, ensuring that AI benefits a broader audience?
Design Opportunities
Work with users to enhance their roles
Augmentation over automation: Focus on creating AI systems that complement human skills rather than replace them. For example:
In creative fields: AI that generates drafts, freeing up humans to focus on strategic, high-level ideation.
In customer service: AI tools that handle FAQs while routing complex problems to human agents with richer context.
Empower users: Design for collaboration, not replacement. Frame AI systems as “partners” that augment human decision-making:
“Here’s what I found—does this align with what you need?”
“Let’s refine this together.”
Emphasise the people building these systems
Transparency about the human cost: Use product messaging to acknowledge the labor behind AI systems: “This system is powered by human expertise and AI collaboration.” with links to more information on ethics policies or principles.
Seek out ethical supply chains: Partner with organisations that prioritise ethical AI development practices. Advocate for AI providers to enforce fair wages, mental health support, and humane working conditions for annotators and moderators.
Reduce exposure to harmful content: Consider design strategies that minimise the need for human moderation of toxic material:
Minimise entry-points for harmful content. Consider when to give users liberty to add user generated content like images or to upload files.
Use automated tools to screen out harmful data before it reaches human reviewers.
Develop feedback systems that allow moderators to flag harmful patterns, improving future processes.
Promote accessibility and equity
Accessible AI tools: Build products that are affordable, intuitive, and accessible to smaller businesses, under-resourced communities, and non-tech-savvy users.
Focus on inclusive design principles: Products that work seamlessly across devices, languages, and contexts, meeting users where they are:
Design for older hardware, slower networks, and less powerful devices.
Offer flexible interaction models (voice, text, visuals) to accommodate different abilities and preferences.
Design for rural or isolated communities: Address barriers in regions with poor infrastructure:
Enable offline functionality or lightweight versions for low-bandwidth environments.
Prioritise data-efficient systems that don’t require constant connectivity or high processing power.
Accommodate a range of digital literacy levels: AI should feel approachable, even for those unfamiliar with complex tools:
Use clear, jargon-free language to explain features and outputs.
Provide guided experiences and tutorials to help users get comfortable.
Design error states or nudges that encourage experimentation: “No worries—let’s try that again.”
Designing for Shared Prosperity
It’s easy to be enamoured by AI’s potential to “do more with less.” But the question we must ask is: “Less of what? And for whom?”
We need to challenge the assumption that productivity at all costs is progress. Is it progress if an AI tool replaces jobs but creates no pathways for re-skilling? Is it progress if AI systems disproportionately serve the wealthy while the marginalised are left further behind?
These are hard questions, but they are necessary ones. They remind us that we are not just technologists—we are stewards of systems that will shape lives, economies, and communities.
6. Accountability and Transparency: Who’s Responsible for the Machine?
The Dilemma
Large Language Models operate as opaque systems, capable of producing fluent, confident, and unpredictable outputs. But when things go wrong—when an AI misleads a user, perpetuates bias, or causes harm—one question becomes uncomfortably clear: who is responsible?
The system’s complexity often obscures accountability:
A flawed medical diagnosis from an AI tool—who takes the blame?
A biased hiring decision—who explains the reasoning behind it?
A fabricated legal answer from a chatbot—who owns the consequences?
Compounding this issue is the “black box” nature of LLMs. Users see the output but not the process. They don’t know how the AI arrived at its answer, what data influenced it, or why it failed. This opacity undermines trust, creates confusion, and leaves users powerless to challenge or understand outcomes that impact their lives.
And when systems are deployed in high-stakes contexts like hiring, healthcare, finance, or law, a lack of transparency and accountability stops being an inconvenience—it becomes an ethical failure.
Why It Matters
Accountability and transparency are not abstract principles—they are ethical obligations that protect people from harm and ensure systems operate fairly. Without them:
Harm occurs without ownership: If no one is responsible for an AI failure, those impacted are left without recourse. Trust breaks down when harm becomes nobody’s fault.
Opacity erodes trust: Users cannot trust AI systems they don’t understand. Opacity creates a power imbalance: systems make decisions, and people are expected to accept them without question.
Human oversight disappears: In the absence of clear boundaries, AI systems might be given undue authority in decisions where human judgment is critical.
Marginalised groups are silenced: Opaque systems amplify bias and inequity while depriving affected groups of the tools to challenge unfair decisions.
Whether the stakes are as personal as a job rejection or as critical as life-or-death medical advice, accountability and transparency are what make AI systems trustworthy, contestable, and ethical.
Key Questions for Designers
How might we design for explainability, helping users understand how AI systems generate outputs?
Could we create clearer boundaries for when AI takes the lead and when human oversight is required?
How do we communicate the limitations of AI products in ways that are understandable and actionable for users?
Design Opportunities
Offer a window into the systems processes
Visible cues for AI reasoning: Show the “why” behind outputs in clear, intuitive ways:
“I recommended this candidate because their experience matches 85% of the job requirements.”
“This answer is based on sources published before 2023.”
Influence mapping: Use visuals to communicate what data influenced the output:
“This response was generated based on similar questions asked by other users.”
Step-by-step reasoning: In high-stakes scenarios, provide pathways to unpack the decision or walk users through the process as its happening with operational transparency.
Blend AI with human oversight
Role definition indicators: Make it clear when AI is assisting and when it is in control:
“The AI has suggested this diagnosis, but a doctor will confirm before proceeding.”
“This tool recommends candidates for review—final decisions remain with your team.”
Human oversight triggers: Implement clear boundaries for when a human must intervene. For example, in medical or financial applications: “The AI cannot confirm this result without human validation.”
Manual overrides: Provide users with the ability to question, contest, or adjust AI decisions:
“Does this answer feel incorrect? Let us know, and we’ll review it.”
Acknowledge boundaries
Clear disclaimers: Set expectations upfront so users understand AI limitations:
“This tool generates responses based on probabilities and may contain inaccuracies. Please verify critical information.”
Uncertainty signalling: Indicate when the AI’s output is unreliable:
Use confidence scores: coloured highlights, or soft language:
“I’m not certain about this answer—want to double-check?”
User prompts for validation: Actively encourage users to seek secondary confirmation:
“This response includes legal advice. Please consult a professional before acting.”
Harm Can’t Be Nobody’s Fault
Transparency and accountability are not optional. They are the foundation of ethical, human-centred AI systems. Without them, we create tools that operate unchecked—decisions without explanation, power without responsibility, and harm without recourse.
AI systems influence decisions that affect people’s lives. With that power comes a responsibility to ensure those decisions are fair, explainable, and contestable. When AI goes wrong, someone must take ownership—not only to address the harm but to learn and prevent future failures.
Without accountability, harm becomes nobody’s fault. Without transparency, systems act without scrutiny. Together, these failures undermine trust, erode fairness, and create a world where powerful tools can operate unchecked.
7. Environmental Impact: At What Cost to the Planet?
The Dilemma
AI may be virtual, but its environmental impact is very real. Behind the seamless generation of text, code, or creative outputs lies an invisible and immense infrastructure—data centres buzzing, GPUs crunching, and power grids straining. Training large language models (LLMs) requires staggering computational resources, translating to vast energy consumption and significant carbon emissions.
As these models scale—more data, more parameters, more power—the cost grows exponentially.
Training a single state-of-the-art LLM can emit as much carbon as five cars over their entire lifetimes.
The energy consumption of AI inference—where models generate real-time outputs—adds ongoing strain, especially as AI integrates into everyday tools at scale.
Environmental responsibility is an ethical issue because it forces us to consider the unseen trade-offs and second order impacts of AI innovation that future generations will inherit.
Why It Matters
The climate crisis is the defining challenge of our time. Every industry, every decision, and every system must now be viewed through the lens of sustainability—including AI. If we continue to ignore the environmental cost of scaling LLMs:
Climate inequality deepens: Data centres require enormous resources, often in areas where electricity grids rely on fossil fuels. Developing countries and communities with fewer resources bear the brunt of climate fallout.
Tech’s promise loses credibility: If AI innovation contributes to environmental harm, the tech industry’s vision of a better, more equitable world begins to ring hollow.
Sustainability becomes an afterthought: If efficiency isn’t prioritised now, we normalise unsustainable AI practices—making it harder to course-correct as models, and their power demands, continue to grow.
The ethical imperative is clear: progress that comes at the cost of the planet is not progress. The challenge is not just to innovate, but to innovate responsibly.
Key Questions for Designers
How might we advocate for energy-efficient model training and deployment practices?
Could we optimise AI tools to prioritise performance while minimising environmental cost?
Are there opportunities to design for transparency around the carbon footprint of AI systems?
Design Opportunities
Designers may not control the energy grid or GPU optimisation, but we are not powerless. By advocating for efficiency, transparency, and awareness, we can help ensure sustainability is woven into the design of AI systems.
Prioritise energy efficiency
Highlight energy costs: Advocate for transparency around the environmental impact of training models. Could we encourage organisations to disclose energy usage and carbon emissions for AI systems?
A visible benchmark: “This model used X kWh of energy, equivalent to Y carbon emissions.”
Promote efficient models: Support teams in exploring alternative training methods that reduce energy costs:
Smaller, optimised models that use less data and computation while maintaining performance.
Techniques like model pruning, and knowledge distillation to reduce computational overhead.
Encourage renewable energy use: Collaborate with stakeholders to prioritise cloud providers and infrastructure powered by renewable energy sources.
Offer alternatives and give control to users
Performance-energy balance: Allow users to choose between performance and sustainability modes:
“Optimise for speed” vs. “Optimise for energy efficiency.”
For less time-sensitive tasks, default to more efficient processing.
Offer more efficient shortcuts: Reduce unnecessary AI computation by designing smarter triggers:
Limit inference processes when idle: “The system is waiting for your input to conserve energy.”
Reduce redundant or repetitive queries by suggesting previous results: “Would you like to use the response from earlier?”
User awareness: Educate users about their impact:
A visual cue: “Running this process will consume X energy. Want to proceed?”
Subtle reminders of environmental savings when efficiency is prioritised: “Thanks for choosing the energy-efficient option—together, we’re reducing carbon impact.”
Visualisation and rewards
Carbon impact metrics: Surface energy usage and emissions for AI systems directly within interfaces:
“This tool has saved X emissions by optimising performance.”
Include a comparison: “Equivalent to planting 10 trees this month.”
Feedback loops: Reward users for choosing lower-impact options:
Gamify efficiency: “You’ve reduced your carbon footprint by X% this week—keep it up!”
Balancing Progress With Responsibility
Innovation often moves faster than reflection. The allure of more powerful AI—bigger models, faster responses, more impressive capabilities—can obscure its hidden costs. The dilemma is not just “At what cost does AI innovation come?” but “Who bears that cost?”
If we continue to build without regard for sustainability, the planet—and those most vulnerable to climate change—will pay the price. As designers, we can advocate for solutions that harmonise technological progress with environmental responsibility.
Wrapping Up: Don't Let Perfect be the Enemy of Good
I don’t pretend to have a neat answer for any of this. The ethical challenges LLMs introduce are real, messy, and unresolved. And maybe that’s the hardest thing about all of it: these systems feel so big, so abstract, that even knowing where to begin feels like a victory.
But the enormity of the challenge doesn’t give us permission to look away. As designers, we’re at the interface—where these models stop being abstract technology and start becoming experiences that live in people’s hands, homes, and heads. It’s our job to bring intention to that moment.
Yes, it’s imperfect. The solutions we come up with now—on bias, misinformation, privacy, sustainability—won’t fix everything. They won’t be enough. But they don’t have to be perfect to matter. In fact, they can’t be.
There’s a pragmatism in design that I’ve always loved. You solve the problem in front of you, knowing that it won’t be the last. You try to leave things better than you found them, knowing the work will never be done. That’s how we should approach this.
Instead of paralysis, choose momentum.
Instead of perfection, choose progress.
So what does that look like? It looks like asking the hard questions—even when we don’t have all the answers.
Who built this? Who pays the cost? Who gets left out?
Are these tools fair, transparent, and respectful?
How do we design for humans—messy, diverse, dignified humans?
It looks like putting care into the cracks where harm could slip through—by making bias visible, truth actionable, and sustainability tangible. It means seeing people as more than users—as creators, owners, stakeholders—and honouring their voices at every step.
Progress won’t look like a grand solution. It’ll look like the sum of the small, intentional choices we make along the way:
Nudging users to question outputs that feel too confident.
Giving credit where it’s due—even when it’s messy and incomplete.
Building AI systems that support, augment, and amplify, rather than erase, human creativity.
Choosing transparency, always, even when it feels harder to explain.
Because here’s the quiet truth: design has always been about moving forward without knowing what the end looks like. We chart a path, we take a step, and we pay attention. If something doesn’t feel right, we fix it. And then we keep going.
This moment we’re in—navigating LLMs, AI ethics, and all the grey spaces in between—feels big. It is big. But that doesn’t mean we shrink from it. It means we lean in with humility, clarity, and care.
We design with intention. We advocate for fairness. We build things that empower rather than exclude.
It won’t be perfect. It doesn’t need to be. Let’s just make sure it’s good—and let that be enough to get us started.
Further Reading
Some examples of much deeper thought that I came across in my admittedly limited reading on these ethical issues. Let me know if you have any recommendations!
Discriminating Systems: Gender, Race, and Power in AI - AI Now Institute
When Good Algorithms Go Sexist: Why and How to Advance AI Gender Equity - Stanford Social Innovation Review
Data Feminism - By Catherine D'Ignazio and Lauren F. Klein, MIT Press
Signal Path
AI is reshaping the way we design—our products, our tools, our jobs. Signal Path is a weekly exploration of the challenges and opportunities in making AI products intuitive, trustworthy, and a little more human. Written by Andrew Sims, a design leader working in Fintech, it’s for designers, product thinkers, and technologists grappling with AI’s impact on their work and the products they make. I’m not claiming to have all the answers—this is a way to think through ideas, articulate challenges, and learn as I go. Thank you for joining me as I navigate this path.