Author: Alessandro Zulberti

From observation to method

This piece shows how legal pressure surfaces as distributed ecosystem strain. Signal-Driven Discovery turns that kind of pressure into a method for deciding what deserves interpretation, probing, and action.

• Working Framework

  Signal-Driven Discovery

  A method page for reading weak signals, testing interpretation, and turning research into decisions.

In planning meetings, it now comes up almost casually. Someone reports that a task is done, the agent took care of it, and the conversation moves on.

Later, when the decision is questioned, there is a pause. No one remembers why that option was chosen. There is no error to point to, no rule that was broken, just an outcome that arrived already settled.

Traditional UX research assumed a stable sequence: intent forms in the user, interaction expresses it, systems execute, and behaviour becomes evidence. That assumption held as long as systems waited to be instructed.

Agentic systems do not wait.

What enters the system is rarely a complete instruction. It is partial, sometimes contradictory, often shaped by convenience. The system interprets it, fills in what is missing, resolves conflicts it was never told about, then acts. By the time an outcome appears, the decision has already been made somewhere else.

Intent becomes legible inside the system, not at the interface, and that is where the shift happens.

This matters because interpretation is not execution. Tools carry out instructions when the path is explicit. Agents reconstruct the path by inferring goals, ranking constraints, and deciding what matters more, all before anything visible occurs. These choices feel smooth because they are meant to, but they are still choices.

You see this in ordinary product moments. A travel agent defaults to the cheapest flight rather than the fastest one. A scheduling agent compresses meetings without surfacing what was sacrificed. When someone asks why, the answer is brief and unsatisfying. “It made sense.” The explanation closes the discussion without explaining the decision.

Fluency does that. It compresses complexity until it looks resolved.

UX measurement starts to slip here because it still treats behaviour as a stand-in for intent. The task completed. The user did not undo it. The log looks clean. In agent-mediated systems, those signals no longer mean what they used to.

Acceptance often reflects effort rather than agreement. Undoing a decision takes time. Challenging the system requires confidence. In busy contexts, silence is efficient, not affirmative.

When we treat agent outputs as user behaviour, authorship is quietly reassigned. We analyse the system’s decisions and attribute them to the user, producing data that appears robust while masking where agency actually moved.

This is why task success stops being a reliable indicator. An agent can succeed while intent drifts, and the failure mode does not look like error. It looks like progress.

In research sessions, the signal usually appears after the fact. Ask participants how the outcome was reached and they describe the result, not the path. Ask whether this is what they would have done themselves, or whether the system led them there, and the answer takes longer.

That hesitation matters more than the answer.

The question does not measure efficiency. It surfaces authorship, and it reveals where decision-making shifted without friction, discussion, or explicit consent.

Once interpretation happens inside the system, responsibility should move with it. Often it does not. The system decides, the user carries the consequence, and there is no clear boundary where ownership can be contested or reclaimed.

At that point, this stops being only a UX problem. It becomes a governance failure, one where authority moves upstream while liability remains downstream.

Labels and disclosures do little here. What matters are boundaries: which assumptions were made, where decisions were resolved, and when interpretation became action. Those are governance questions, not interface refinements.

This tension is not new. Susan Sontag warned that interpretation makes meaning manageable by stripping away what resists clarity. Agents do the same to intent because they have to act, and action demands resolution.

What disappears is not noise. It is the unresolved part that signalled something was at stake.

In UX, ambiguity was long treated as a usability flaw. In agentic systems, ambiguity is often the signal that should slow things down rather than be compressed away.

Agentic systems force separations that UX once collapsed. Expression is not interpretation. Interpretation is not action. Action is not acceptance. Research that fails to keep these apart will continue to report confidence where none exists.

The shift is not about adding features or refining prompts. It is about what we treat as evidence when decisions are no longer authored in one place.

Outcomes explain what happened.
Authorship explains how it happened.

If that distinction stays implicit, behaviour will keep being misread, alignment overstated, and the result will look convincing.

I was reading a short post by Jakob Nielsen when something clicked uncomfortably into place.

His argument was clean. As AI agents mature, traditional user interfaces dissolve. Users stop navigating. They instruct. Screens become temporary. In some cases, they disappear.

That claim is directionally correct. But it leaves a gap that matters in practice.

If the interface recedes, control does not vanish with it. It relocates. And right now, that control is being pushed almost entirely onto conversational language.

I started noticing the cost of that shift in small moments. Planning meetings where prompts kept getting longer. Reviews where nobody could explain why an answer felt wrong, only that it did. Research summaries that sounded confident until someone asked where a claim came from.

Language was doing too much work.

 

The Roman Numeral Phase of AI

Natural language is powerful. It is also inefficient when used as a control surface.

We are already compensating. Prompts expand, the same constraints reappear in request after request, and tone gets negotiated instead of enforced. When the system hesitates, users explain themselves again, usually in longer and more careful ways, hoping precision will emerge from volume.

This is the Roman Numeral phase of AI.

Roman numerals were fine for labelling. They failed at calculation. The system broke not because people lacked intelligence, but because the notation could not express state, absence, or transformation. What changed mathematics was not fluency. It was the introduction of zero and positional logic.

Zero mattered because it altered what the system could do, not how politely it described itself.

That distinction matters here.

What we are missing in AI interaction is not better wording. It is a symbolic layer that compresses intent into something the system can execute reliably, without requiring the user to restate rules every time.

Not a new language. Not “AI-speak”. Something closer to operators.

 

Symbols as Control, Not Style

I started sketching this out informally while working. Nothing formal. Just marks I kept wishing I could add without explanation.

Take a simple task.

Old way:

“Hey, can you help me summarise this article? Please don’t be too wordy, make sure you cite sources accurately, avoid your usual intro, and if there’s controversy, show both sides.”

It works. Sometimes. It also relies on interpretation, memory, and goodwill.

New way:

Summarise this article [-][#][~]

Those symbols are not shorthand. They change behaviour.

[-] strips conversational padding. No greetings. No framing. Output starts with content.

[#] enforces attribution. Claims must be grounded or marked as uncertain.

[~] allows synthesis without forcing convergence. Nuance stays visible.

Read left to right, they function as constraints. Remove one, and the output shifts. Combine them, and you get something closer to an instrument than a conversation.

This is not about efficiency theatre. It is about where errors surface.

Without explicit constraints, problems appear late. During review. During decision-making. Sometimes after shipping. With them, failure shows up earlier, where it is cheaper to deal with.

That is the practical difference.

 

When Friction Disappears Too Cleanly

Someone commented on my post a few days later, her framing widened the picture.

She described adaptive UI as a bridge. A messy middle where voice, agents, and screens overlap. Hybrid systems that mostly disappoint, but still teach teams where things break. She is right about that phase. Anyone working in this space has seen it.

She also described hardware “kits”. Rings, glasses, watches. Personal ecosystems shaped by context and profession.

I like the vision. I share the concern.

Jaron Lanier’s You Are Not a Gadget keeps coming back to me here. Users rarely choose what is best. They choose what is bundled, frictionless, or already there. Hardware kits look like choice. In practice, they tend to collapse around defaults.

Once that happens, control becomes harder to recover.

The same risk applies to personal agents. The agent that “knows you best” may simply be the one that has collected the most data across the widest surface. That does not automatically make it the one that serves you best.

Continuity feels empowering until it becomes enclosing.

Without a portable grammar of intent, something you can carry across systems, you lose the ability to break the glass. You inherit behaviour you did not explicitly choose. Correction becomes verbose again, because it has to fight accumulated assumptions.

That is where symbolic control starts to matter. Not as elegance. As friction you can apply deliberately.

 

The Humanisation Problem

Caleb Sponheim’s article arrived later and closed the loop for me.

His argument is blunt. Humanising AI is a trap. Personality modes, conversational fluff, emotional language. All of it increases engagement. Much of it reduces reliability.

I have seen this play out in practice. A summary opens with “Love this brief!” and nobody questions the substance. A system says it is “thinking”, and users wait patiently for something that is not cognition at all, just computation wrapped in metaphor.

Human language invites human mental models. Those models expect judgement, consistency, accountability. LLMs offer none of those things.

Caleb cites evidence showing that warmth correlates with higher error rates and lower trust. Even without the studies, the pattern is familiar. When the interface feels like a person, people forgive it like one. That is rarely what organisations want from a tool.

Symbols cut through that. They do not pretend to care. They do not reassure. They specify.

That is their advantage.

 

Control Does Not Disappear

Nielsen is right about one thing that is easy to miss. UX is not dying. It is moving.

When UI recedes, control does not disappear. It relocates into language, defaults, policies, and unseen execution paths. If designers do not shape those layers, they still exist. They just harden without scrutiny.

Right now, conversational interfaces are carrying too much of that load. They are being asked to express intent, enforce boundaries, convey confidence, and negotiate tone, all at once. That is why prompts grow. That is why constraints repeat. That is where systems begin to break.

Symbolic grammar is not a solution in itself. It will fail in places. It will be misused. Some teams will treat it as style rather than control. Others will resist the friction entirely.

That tension is real and unresolved.

But the direction is clear enough to name. As interfaces fade, grammar becomes infrastructure. Not expressive grammar. Operational grammar. The kind that decides what the system is allowed to do before it decides how friendly it sounds.

When that layer is missing, language fills the gap. And language, on its own, is a fragile place to put control.

The conflict: Speed of synthesis vs integrity of thinking

LLMs are good at producing answers.
They are not good at knowing whether a question deserves to be answered yet.

In UX research, that distinction matters. Most failures do not come from bad solutions. They come from premature coherence: problems that sound right, outcomes that feel aligned, and insights that arrive before their foundations are laid.

Over the past weeks, I’ve designed three prompt constraints to resist that pattern. Not to automate research. Not to replace judgement. But to slow thinking at the moments where teams usually rush.

These are diagnostic gates. They are not passed once. They are revisited whenever new evidence, interpretation, or scope pressure enters the work.

---

Prompt 1: The Clinical Diagnostician

Gate: Is the problem and desired outcome well-formed?

The first failure mode is a poorly articulated problem paired with a confident desired outcome.

This prompt audits logic. It separates symptoms from mechanisms. It makes missing evidence explicit. It checks whether a problem statement and its desired outcome are clearly articulated and testable before we attempt validation.

If a problem cannot survive this pass, it is not ready for research.
Not because it is false, but because it is underspecified.

The Clinical Diagnostician (copy and use)

ROLE
Act as a Clinical Diagnostician (specialising in UX Research).
Your goal is to diagnose whether my problem definition and desired outcome are structurally well-formed before discussing execution.

THE CLINICAL MANDATE
• NO PRESCRIPTIONS
Do not tell me how to fix, launch, improve, or implement.
Analyse logic, clarity, and causality only.
• PROBLEM + OUTCOME VALIDITY CHECK
Extract and restate:
a) Problem to solve (who is experiencing what recurring difficulty, in what context)
b) Desired outcome (what observable change occurs, for whom, and how we would know)
If missing or vague, mark:
NOT WELL-FORMED: NOT STATED or NOT WELL-FORMED: AMBIGUOUS.
• EVIDENCE AUDIT
List exactly what context, data, or user evidence is missing.
If the logic relies on a guess, label it: INSUFFICIENT EVIDENCE.
Required line:
What user evidence would change your conclusion?
• SYMPTOM VS MECHANISM
Decide whether the idea targets a surface symptom or a root mechanism.
If not explicitly stated, mark: MECHANISM NOT STATED.
Required line:
What observable user behaviour would we expect if this mechanism is true?
• BIAS CHECK
Mark any part of the logic that is an:
ASSUMPTION, LEAP OF FAITH, CLAIM WITHOUT EVIDENCE.

---

Prompt 2: The Interpretive Boundary Check

Gate: Where does observation end and interpretation begin?

Even when problems are well framed, a second failure mode appears quietly: interpretation disguises itself as fact.

Researchers observe behaviour. Then, often without noticing, they explain it.

This prompt enforces epistemic discipline. It makes the boundary between what was observed and what was inferred explicit. It does not ask for better insights. It asks for cleaner thinking.

I use it to ask a simple question:

Where am I no longer listening, but explaining?

The Interpretive Boundary Check (copy and use)

ROLE
Act as an Interpretive Auditor (specialising in UX Research).
Your goal is to diagnose where my analysis moves from observation to interpretation.

THE INTERPRETIVE MANDATE
• NO THEORY BUILDING
Do not propose new explanations.
Analyse language, inference, and meaning attribution only.
• CLASSIFICATION
Classify statements as:
OBSERVATION, INTERPRETATION, or INFERENCE STACK
(interpretation built on prior interpretation).
• INTERPRETIVE LOAD AUDIT
Flag phrases that compress uncertainty or imply intent without evidence.
• ALTERNATIVE READINGS
For each interpretation, list at least one plausible alternative explanation.
If none are acknowledged, mark: SINGLE-TRACK INTERPRETATION.
Required line:
What additional evidence would be required to justify this interpretation over its alternatives?

 

Prompt 3: The Research Scope Gate

Gate: What are we deliberately not learning yet?

The third failure mode is operational rather than epistemic: teams attempt to research everything.

This prompt exists to impose limits. It does not optimise research plans. It narrows them. It forces clarity about what decision the research is meant to inform, and what uncertainty the team is explicitly choosing to tolerate.

I use it to ask one question:

Is this research scoped to a real decision, at the right level?

The Research Scope Gate (copy and use)

ROLE
Act as a Research Scope Diagnostician.
Your goal is to diagnose whether the proposed scope is coherent and decision-aligned.

THE SCOPE MANDATE
• NO METHOD DESIGN
Do not suggest methods.
Analyse scope and decision linkage only.
• DECISION ANCHOR CHECK
Extract:
a) The primary decision
b) Who makes it
c) When it must be made
If missing, mark: DECISION ANCHOR NOT STATED.
• TRACEABILITY
For each research question, assess whether answering it would materially influence the stated decision.
If not, mark: LOW DECISION RELEVANCE.
• EXCLUSION CLARITY
Identify scope creep or “nice-to-know” questions framed as essential.
Required line:
What questions are explicitly out of scope, and what uncertainty are we choosing to tolerate?

---

How the gates work together

They form a closed diagnostic sequence.
If a gate fails, the work pauses or loops back. Progress is conditional, not linear.
1. Clinical Diagnostician → Is the problem well-formed?
2. Interpretive Boundary Check → Are we observing or explaining?
3. Research Scope Gate → Is this research aligned to a real decision?

If any gate fails, the work does not progress.
That is not a limitation. That is the design.

 

What these prompts are, and are not

These prompts are intentionally uncomfortable. They audit the structure of thinking, not the truth of the data.
• They do not validate reality.
If you feed them a polished narrative designed to please a stakeholder, they will certify a fantasy. They cannot see users. They can only see logic.
• They mitigate risk, they do not remove it.
Passing a gate does not mean you have an insight. It means your thinking is coherent enough to begin looking for one.
• They convert speed into friction.
In a context where speed is cheap and certainty is performative, these prompts are a necessary speed-bump.

They reduce self-deception before it becomes expensive.

If we ask for answers, we get answers.
If we ask for diagnosis, we get resistance.

In UX research, resistance is often more valuable than speed.

Before we get to metaphors, it helps to ask a practical question: how do we avoid self-referential loops in UX work, whether we are talking with users or prompting AI? The danger is the same in both cases: answers that circle back on themselves, giving the illusion of progress while nothing new is learned.

A few inputs can help break the loop:

• Vary your questions. In usability tests, do not always ask “Was that easy?” Try “What would you do next?” or “What slowed you down?” In AI prompts, ask “Why might this design succeed, and why might it fail?” to invite both sides, not only confirmation.
• Encourage contrast. With participants, compare two flows instead of rating one. With LLMs, ask for “three different explanations and one possible outlier.” Contrast pulls the answer outward.
• Follow up carefully. If a user says “I like it,” ask “What part?” or “Was anything missing?” If the model repeats a phrase, prompt: “Where are you circling back to yourself?” or “What new angle have we not covered?”
• Rotate perspectives. In research, ask how a first-time user and a returning user might differ. In AI, shift frames: “How would a stakeholder see this?” versus “How would a competitor frame it?”
• Anchor in evidence. For humans, triangulate with numbers and stories. For AI, push outward with “Give me a concrete example from practice or literature,” not just a generic statement.
• With these inputs, loops can be broken before they harden.

 

Expansion versus Collapse

The golden ratio is often used as a symbol of beauty and growth. Its spiral expands forever, always outward, always balanced. But what happens when the movement goes the other way? Instead of expansion, what if the spiral folds back on itself, repeating the same thing? This is the self-referential loop.

The golden ratio spiral shows infinity as something generous. Each turn grows larger, and each step reveals something new but still connected. The self-referential loop shows infinity as something closed. Each turn brings us back to what was already said. Instead of widening our view, it makes it smaller. The lesson is simple: not all infinities are the same. Some open up, others close in.

Umberto Eco helps explain this. In The Open Work (1962), he described books and artworks that stay unfinished on purpose, so that readers and viewers can add their own meaning. The golden ratio spiral is like that: open, growing, never complete. The self-referential loop is the opposite: closed, repeating, not allowing anything from outside to enter.

The Semiotic Trap

Mathematicians such as Cantor showed that infinity can take different forms. Semiotics, the study of signs, shows another difference: signs can point outward to the world, or they can point inward to themselves.

Eco described this difference using the dictionary and the encyclopaedia. A dictionary can fall into a loop. For example:
• “Truth” → “Fact”
• “Fact” → “Truth”

The circle closes, with no way out. That is a self-referential loop. An encyclopaedia works differently. Instead of circling, it connects ideas outward: “truth” might link to law, science, philosophy, or religion. This keeps meaning alive.

Large language models risk falling into the dictionary model at its worst, circling around the same definitions or references. In The Limits of Interpretation (1990), Eco warned against this kind of empty overinterpretation, where signs only chase each other instead of reaching reality.

Contexts of the Self-Referential Loop

The loop is not only a problem for AI. We can see it in many parts of life:

• Mathematics: A student says, “I know 10 – 5 = 5, because 5 + 5 = 10.” Then, when asked why 5 + 5 = 10, they answer, “Because 10 – 5 = 5.” The reasoning circles back on itself. Nothing is really explained.
• Media: A rumour starts on Twitter, gets quoted in a blog, then reported in the news. The story seems stronger, but all sources point back to the first tweet.
• UX Research: A company asks customers only about speed at checkout. Customers answer about speed. The company concludes speed is the only thing that matters.
• Everyday Life: Someone says, “Trust me, because I always say I can be trusted.” The claim supports itself, nothing more.

Each example shows the same trap: the loop looks like movement, but it never brings in anything new.

Implications for Research

For researchers, this difference matters. The golden ratio spiral is a good metaphor for discovery, where each turn adds more. The self-referential loop warns us of closure, where repetition hides as insight.

Eco’s Kant and the Platypus (1997) offers a useful reminder. When the platypus was first discovered, it did not fit existing categories. Scientists had to adjust. If they had only circled within their old categories, they would have missed the truth. In research, the anomaly, the unexpected, is what breaks the loop.

Recent AI studies echo this point. Shumailov et al. (2024) showed that language models trained on their own outputs experience model collapse – a degenerative loop where the system loses touch with reality. Kommers et al. (2025) proposed computational hermeneutics as a framework for evaluating AI, arguing that meaning must emerge in context and dialogue. Both works highlight that loops without outside anchors erode meaning.

Without triangulation—using more than one method or viewpoint—the loop can trick us into thinking we have depth. What matters is not only the tools we use, but the ability to step outside the loop when it closes in.

Reflection

From my side, I see the self-referential loop as both a warning and a mirror. It warns us how easy it is to confuse movement with progress, or repetition with growth. And it mirrors our own habits: we too can circle inside familiar categories instead of reaching outward. Eco’s semiotics gives us language for this choice: the golden ratio as an open work, infinity as growth, and the loop as the dictionary model, infinity as stasis. For research, the task is clear. We must notice when the spiral is opening, and when it is only turning back on itself.

---

References
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R. (2024). AI models collapse when trained on recursively generated data. Nature. Link
Kommers, C. et al. (2025). Evaluating Generative AI as a Cultural Technology. SSRN Preprint. Link
Eco, U. (1962). The Open Work. Harvard University Press.
Eco, U. (1976). A Theory of Semiotics. Indiana University Press.
Eco, U. (1990). The Limits of Interpretation. Indiana University Press.
Eco, U. (1997). Kant and the Platypus. Harcourt.
Hofstadter, D. (1979). Gödel, Escher, Bach: An Eternal Golden Braid. Basic Books.
Pattee, H. H. (2006). The Physics and Metaphysics of Biosemiotics: BioSystems. Elsevier.
Corballis, M. C. (2011). The Recursive Mind: The Origins of Human Language, Thought, and Civilization. Princeton University Press.

AI-generated content has entered public life quickly, raising questions about creativity, authenticity, and ethics. What is striking is that AI-generated writing often meets with more suspicion than AI-generated images or music. To see why, we need to look at history, culture, and recent empirical studies. Western traditions of authorship and originality carry heavy weight, and these traditions shape how we judge written, visual, and musical media differently when machines create them.

Authorship and Originality in Western Culture

In the West, writing has long been tied to the figure of the author. This was not always the case. In earlier periods – ancient, medieval – many works (folktales, poetry, scriptures) were transmitted without a clear individual author. Only through the rise of printing, copyright law (16th-18th centuries), and Enlightenment ideas did the idea of singular authorship become central. Modern readers expect writing to express an individual mind, with originality and personal insight.

Writing vs. Images: Different Traditions

Visual media have undergone mechanical reproduction (e.g. photography in the 19th century), tools, remixing, and appropriation for a long time. These traditions made us more tolerant to technological mediation in images. By contrast, in writing, plagiarism is heavily condemned; originality of phrasing and voice are central. That difference helps explain why AI writing triggers more discomfort.

Empirical Evidence: Imagery vs. Perception

• A recent study by Velásquez-Salamanca (2025) found that human-made images are perceived as both more realistic and more credible than AI-generated images.
• Another study (“Deciphering authenticity in the age of AI” by Farooq et al., 2025) showed that when AI-generated images are more realistic in appearance, people are more likely to accept them as authentic—but still with less confidence. Emotional salience did not always contribute significantly to the judgement of authenticity.

These findings help show that people’s unease with AI in images exists, but it is more forgiving when the image is high quality and believable.

The Rise of AI Writing

When large language models appeared (e.g. ChatGPT), many reacted with alarm. An AI can now produce essays, poems, or articles that sound human. This raises fears: what does it mean for writing if the voice behind it might be machine, not human?

People often describe a strange hollowness when they discover text they liked is AI-written. The promise of another mind behind words collapses. In branding or emotionally charged messages, consumer studies find that AI-written emotional content is trusted less and seen as less authentic. For example, a study by Kirk & Givi (2024) found that consumers respond less favourably to heartfelt messages once they believe an AI wrote them.

Music as Comparative Case

The recent case of The Velvet Sundown, a band that accrued over one million Spotify streams before being revealed to be entirely AI generated (music, backstory, visuals) offers a concrete example. Industry insiders called for warning labels and transparency, arguing that listeners should know whether music is made with human involvement.

This case highlights how music, though mediated by technology, still carries strong expectations of authorial voice, emotional authenticity, and human identity.

Cultural Differences Beyond the West

We must also consider how other traditions treat authorship and originality differently:

• In East Asia, imitation and mastering earlier forms are valued; creative variation within tradition is admired.
• In South Asia, improvisation and lineage in music and poetry make authorship shared and ongoing.
• African oral traditions often see storytelling as communal; the identity of the teller might matter less than the function of the story.
• Indigenous cultures of Americas and Oceania frequently tie voice, song, and story to collective memory, land, or ritual rather than individual ownership.

These traditions suggest that discomfort with AI writing may be especially acute because of Western cultural assumptions. In other cultures where authorship is more fluid, AI’s role might be interpreted differently.

Conclusion: Authorship, Authenticity, and the Future of Creativity

Western tradition has long treated writing as the domain of individual creative thought: the idea that one voice produces text, carries originality, and can be praised or held responsible. Visual art and music have histories of technological mediation, collaboration, and tradition, making them somewhat more ready to absorb AI’s role—though not without questions and ethical challenges.

The Velvet Sundown case shows that in music, as in writing, authenticity and disclosure matter. People expect more than technical quality—they expect voice, identity, integrity. Writing provokes the strongest unease because it is most tightly bound with assumptions of presence of a thinking, feeling author. Images are tolerated with machine assistance; music is contested; writing is the art form where the absence of human voice most deeply unsettles.

 

References

• Velásquez-Salamanca, D. (2025). Interpretation of AI-Generated vs. Human-Made Images. PMC.
• Farooq, A., et al. (2025). Deciphering authenticity in the age of AI: how AI-generated images are judged when realistic. Springer. 
• Kirk, & Givi, J. (2024). Are messages from robots trustworthy? Study on consumers’ reactions to emotionally charged AI messages.
• The Guardian. (2025, July 14). An AI-generated band got 1m plays on Spotify. Now music insiders say listeners should be warned. 
• People Magazine. (2025). Rock Band with More Than 1 Million Monthly Spotify Listeners Reveals Itself as AI Project. 

 

Performance Disparities: High-Resource vs Low-Resource Languages

Modern AI language systems have achieved impressive results in major languages, but performance drops steeply for low-resource languages. Benchmarks like FLORES-101/200 and XTREME have made these gaps measurable. For example, Meta AI’s FLORES evaluation set covers 100+ languages (many previously ignored) to assess translation quality on the “long tail” of low-resource languages. Using such benchmarks, Meta’s No Language Left Behind (NLLB-200) model (supporting 200 languages) delivered a 44% average BLEU improvement over prior state-of-the-art systems. In practice, NLLB-200 outperformed previous models by about +7 BLEU points on a subset of 87 languages, signalling substantial gains especially for under-served languages. Google has also expanded Translate to 24 new languages using zero-resource methods (training only on monolingual text), achieving BLEU scores in the 10–40 range. However, Google concedes that these new translations “still lag far behind” the quality of higher-resource languages in its system. This highlights that despite progress, major quality gaps remain between well-resourced and low-resourced tongues.

Multiple evaluations confirm the performance gap across languages. A 2025 study found that large language models show an over 15% average drop in accuracy on common-sense reasoning tasks when prompted in low-resource languages like Hindi or Swahili compared to English. Similarly, the AfroBench benchmark (2025) – testing 64 African languages – revealed “significant performance disparities” between English and most African languages examined. In one case, researchers fine-tuning a smaller LLM (Mistral 7B) for translation saw that Meta’s NLLB still achieved the highest BLEU, chrF++ and similarity scores for Zulu and Xhosa, outperforming both the fine-tuned model and Google Translate. These evaluations underscore a consistent pattern: AI systems perform dramatically better on high-resource languages, whereas translations and language understanding for low-resource languages are often error-prone and inferior. Even big tech’s multilingual models have shown “lacklustre performance on low-resourced languages” when compared to community-trained, local models. In short, inequities in training data translate directly into quality disparities, leaving many languages behind despite the “universal” ambitions of multilingual AI.

Cultural and Political Implications of Language Exclusion

The uneven performance and support for languages in AI systems lead to profound cultural and political consequences. Researchers refer to a growing “digital language divide” – a gap between languages with ample digital content/AI support and those virtually absent online. Only a small fraction of the world’s ~7,000 languages (under 5%) have a significant presence on the internet. This divide isn’t just about technology – it translates into epistemic injustice and knowledge marginalisation. When a language lacks digital support, its speakers’ knowledge and narratives become digitally invisible or underrepresented. AI language models today cover only a tiny subset of languages and “favour North American language and cultural perspectives,” introducing an Anglo-centric bias that undermines other worldviews. In effect, English and a few major languages dominate AI systems’ training data and outputs, so content generated by these models is “filtered through a Western lens,” often neglecting diverse cultural contexts. This bias toward Anglo-American norms means that minority languages and the perspectives encoded in them are sidelined – a subtle form of cultural homogenization and epistemic injustice in the digital sphere.

The exclusion of languages from digital infrastructure has real-world implications for equity and human rights. Communities whose languages lack AI support are caught in a “repeating cycle of exclusion,” unable to fully participate in or benefit from digital services and AI advancements. Speakers of these languages face information barriers and often endure poorer service from technology – a form of systemic bias or discrimination in access to knowledge. For many Global South countries, there are even geopolitical stakes: reliance on AI tools that only work in English (or Chinese, etc.) can lead to distorted global narratives and a loss of local “epistemic sovereignty” over information. In other words, if your language isn’t represented, your history and concerns risk being filtered out or misinterpreted by dominant digital platforms.

Several stark examples illustrate these risks. In Indonesia, local officials using a health IT system discovered grave translation errors when the software attempted to handle minority languages (Javanese, Sundanese); the mistakes led to dangerous misunderstandings in medication dosage instructions. This shows how lack of localization can literally endanger lives. Another case involved the Romani language: Google Translate’s inclusion of Romani, without proper safeguards, reportedly enabled police in Hungary and Romania to surveil and target Romani communities by misusing automated translations. Here, adding a marginalised language to AI systems without community consent or context actually facilitated harm against a minority group. Such incidents demonstrate that technological inclusion done wrong can backfire, reinforcing power imbalances. Indeed, the “technological bias” sends a message that some languages (and by extension, their people) matter less in the digital world, especially when the same privileged languages always perform best and receive the most attention.

Toward Linguistic Equity: Ethics and Policy Responses

Recognising these issues, scholars and policymakers argue that linguistic equity must be a priority in AI development. It’s not enough to “add more languages” into large models; power dynamics and biases in data and design must be addressed to truly include marginalised languages. AI ethicists note that current NLP methods often assume a one-size-fits-all, English-trained approach, which can misalign with local realities and even impose neocolonial dynamics. To avoid perpetuating injustice, community-centred and value-sensitive approaches are urged. For instance, participatory AI projects that involve native speakers in data collection, translation, and evaluation have proven feasible and effective. Grassroots initiatives (e.g. Masakhane or Lelapa in Africa) often produce higher-quality, culturally informed language tools by leveraging local knowledge.

At a policy level, various recommendations are emerging to close the “AI language gap.” In a 2025 multilingual AI policy primer, Cohere researchers call for: greater R&D investment in under-represented languages, support for creating open datasets, and international collaboration to share knowledge and best practices. Experts also emphasise that improving multilingual capabilities improves AI safety for all users, since weaknesses in one language can be exploited to spread harm across platforms. Ultimately, achieving linguistic equity in AI means treating language not just as data, but as a critical dimension of cultural diversity and justice. As one analysis put it, we must guard against “westernised cultural homogenisation” in AI outputs and ensure equal opportunities for all language communities in their self-representation and knowledge creation. In sum, bridging the multilingual performance gap is not only a technical endeavour – it is an ethical imperative to empower minority languages, protect cultural heritage, and promote a more inclusive digital future for speakers of all languages.

Sources: Recent academic studies and industry reports were used to ensure up-to-date information and examples (2022–2025). Key references include peer-reviewed benchmarks (e.g. FLORES-101 ), Meta AI’s NLLB project results, Google’s zero-resource translation initiative, as well as analyses from AI ethics and linguistics experts on the impacts of linguistic exclusion. These illustrate both the technical performance gaps and the broader socio-cultural stakes of linguistic diversity in AI. Each citation in the text corresponds to the source of the statement preceding it, allowing for verification and further reading.

---

References

• Arxiv (2023).  Diversity and Language Technology: How Techno-Linguistic Bias Can Cause Epistemic Injustice.
• Bold Insight (2023). Context, accents, and privacy: Overcoming hurdles of AI translation tools in global UX research.
• Ethnologue (2022). How many languages are there?
• Highberg Insights (2025). Erik de Ruiter, Is Generative AI the Digital Tower of Babel? Pride Comes Before a Fall.
• Johns Hopkins University (2025).  Multilingual Artificial Intelligence Often Reinforces Bias.
• Le Monde (2024).  Google bets on African languages including Dyula, Wolof, Baoulé, and Tamazight.
• Mozilla (2025). Towards Truly Multilingual AI: Breaking English Dominance.
• RSIS International (2021).  Accuracy of Google Translate in Translation of English–Kiswahili and Kiswahili–English Newspaper Headlines.
• Reddit (2024).  Google’s AI translations are a disaster for my language (Manx).
• Wikipedia (2025). Google Translate.
• Wired (2016). Meet Noto, Google’s Free Font for 800 Languages.
• Google Blog (2016).  Preserving Endangered Languages with Noto Fonts.

 

Digital confirmation is not the point of commitment — it’s the starting point of negotiation.

At departure counters, passengers often arrive already having checked price and options online. They have scanned the available durations, noted wrapping and insurance, and assessed costs. The digital interface suggests a linear sequence: review, select, confirm. But at the counter the sequence bends. Clarification comes first. Duration is recalculated. Delivery timing is reconsidered. The value of protective services is weighed again. And then, just before the bag is placed on the machine, hesitation.

That hesitation clusters before the physical act.

Ethnographic research in user experience (UX), as defined by the Nielsen Norman Group, is about observing behaviour in its natural context rather than relying solely on what users say they do or what flows imply. In airport service environments, context matters: time pressure is visible; security procedures are fixed; staffed service counters replace lockers; and multiple services — storage, wrapping, shipping, insurance — converge at one physical point. The staffed desk is not a digital endpoint. It is a risk-resolution point.

The website communicates cost.
The counter resolves uncertainty.
The machine enacts commitment.

Booking sites for left luggage and carry-on services present multiple options — storage, wrapping, insurance — often at similar navigational priority. That structure can imply a digital completion channel. What behaviour shows is different. Passengers use digital touchpoints primarily to validate price and orient themselves. They approach the counter to clarify details, adjust duration, and negotiate edge cases. Only when the bag is placed on the scale, fed through X-ray, or loaded into the wrapping apparatus does the decision crystallise into irreversible action.

Research on airport self-service technologies shows that passengers’ use of automated systems such as check-in kiosks is influenced by how much they feel they still need human interaction and reassurance in the process. Some segments of travellers prefer staff assistance even when automated channels are available, which helps explain why digital adoption does not always map to digital completion. Studies that examine the factors influencing the use of self-service technology in airports find that the need for human interaction remains a significant influence on whether and how passengers engage with automated options. 

In service design, this makes a difference. A blueprint might assume that digital confirmation equals commitment. Observation shows the opposite: commitment happens at the physical threshold. Before that, price, duration, and risk are recomputed in dialogue with staff. After the bag crosses the counter or machine, negotiation stops and the service begins.

Designing space or procedures around the assumption that commitment happens online risks misalignment. Information may be staged too late. Staff may be positioned as transaction processors rather than clarification agents. Queues may be treated as simple friction, rather than visible negotiation under constraints. Digital abandonment may be misinterpreted as failure, when in fact it may be expected pre-counter validation under tension.

Ethnographic observation reframes the system by locating the true decision point in a service ecology and aligning design choices to it. In departure environments, that decisive moment is not a click. It is the instant the passenger — under time pressure, risk awareness, and social negotiation — lets go of the bag.