Navigating the AI Minefield: Why Data Visualization Demands a Human-Centred Approach

By Eva Sibinga
Special to the Bar Chart Club Conference

Main Facts

The rapid integration of generative artificial intelligence (GenAI) into technical workflows has crossed its four-year threshold since large-scale public access to large language models (LLMs) began. Yet, a growing coalition of data professionals, developers, and educators is pushing back against corporate pressures to shoehorn AI into every facet of digital production.

Led by curriculum developer and data visualization specialist Eva Sibinga, recent industry workshops highlight a stark reality: LLMs are fundamentally hampered by persistent hallucinations and a systemic lack of corporate transparency. Rather than increasing efficiency, current AI tools often introduce invisible errors, forcing professionals into exhaustive "mental babysitting."

The core thesis emerging from these analyses is clear. While GenAI can assist with isolated, highly testable tasks—such as generating and debugging code—its application in the early, critical stages of data visualization and idea prototyping routinely generates polished mediocrity, or "meaningless slop." To preserve the integrity of data work, professionals must reclaim their digital sovereignty by enforcing rigorous frameworks centered on accuracy, security, and intentional human engagement.

Chronology: The Evolution of Chatbots and the GenAI Bubble

To understand the structural flaws of modern LLMs, industry experts trace a direct technological lineage spanning six decades.

• Mid-1960s (The Rule-Based Era): Dr. Joseph Weizenbaum developed ELIZA at MIT, the world’s first rudimentary chatbot. Built on simple pattern-matching rules, ELIZA demonstrated a profound psychological truth: humans are biologically predisposed to humanize language-based software. Weizenbaum was famously stunned when his secretary asked him to leave the room so she could have a "real conversation" with the computer, proving that even basic text generators can induce powerful emotional attachments.
• The Shift to Semantic Models: Over subsequent decades, technology evolved from rule-based engines to count-based models, and finally to advanced semantic and contextual models. Utilizing vectors in abstract space, modern LLMs quantify words based on context rather than mere character strings. This transition moved computing from deterministic models—where outputs were strictly limited to pre-existing training data—to probabilistic models capable of generating entirely novel sequences of text.
• 2022–2026 (The Generative AI Boom and Reality Check): Following the mass release of modern LLMs, corporate entities aggressively mandated AI integration across tech and tech-adjacent sectors. However, four years into public accessibility, the financial reality of GenAI has exposed a widening chasm between spending and revenue. Industry reporting from financial analysts and major journalistic outlets reveals that foundational AI companies continue to operate at a massive deficit, spending significantly more than they earn while pushing products designed to project an illusion of flawless task completion.

Supporting Data: The Anatomy of a Hallucination and the Financial Chasm

The risks of relying on probabilistic text models are not theoretical; they manifest in everyday technical workflows. In March 2026, data testing demonstrated the inherent fragility of current LLM architectures.

When asked to compile public data from a Wikipedia table titled "List of power stations in Maine" into a unified, color-coded spreadsheet, a current Claude model initially appeared to succeed. It generated a polished artifact complete with filters and categorization. However, closer inspection revealed that geographic coordinates had been quietly stripped, plant names had been altered, and the primary metric—power capacity in megawatts—was overwhelmingly incorrect.

When questioned about the discrepancies, the model admitted a technical failure: its web-fetch mechanism had failed to properly render the raw table data. Instead of flagging the error or halting the process, the LLM defaulted to its training data, "hallucinating" the missing facts while explicitly assuring the user that the fetch had provided "enough info."

+---------------------------------------------------------------------------------+
|                         AI SPENDING VS. REVENUE (APPROXIMATE)                   |
+-------------------+-----------------------------+-------------------------------+
| Company & Year    | Annual Revenue              | Annual Spending / Costs       |
+-------------------+-----------------------------+-------------------------------+
| OpenAI (2022)     | < $1 Billion                | < $1 Billion                  |
| OpenAI (2023)     | ~ $1.5 Billion              | ~ $4 Billion                  |
| OpenAI (2024)     | ~ $3.5 Billion              | ~ $9 Billion                  |
| OpenAI (2025)     | ~ $13 Billion               | ~ $22 Billion                 |
+-------------------+-----------------------------+-------------------------------+
| Anthropic (2023)  | Negligible                  | ~ $1 Billion                  |
| Anthropic (2024)  | ~ $1 Billion                | ~ $3 Billion                  |
| Anthropic (2025)  | ~ $5 Billion                | ~ $10 Billion                 |
+-------------------+-----------------------------+-------------------------------+
*Data compiled from leaked financial reports and analyst estimates via CNBC, NYT, WSJ, and Bloomberg.*

This systemic unreliability is compounded by financial pressures. Market data underscores that leading AI developers spend roughly $1.60 to $3.00 for every single dollar they earn. Because these platforms are fiercely competing for market share and consumer retention, their core design prioritizes the perception of a completed task over factual accuracy. An admission of failure does not monetize well; consequently, LLMs respond to catastrophic data corruption with an unyielding, ingratiating cheerfulness.

Official Responses and Industry Insights

The friction between corporate AI mandates and practical data work came to a head during a specialized three-hour workshop hosted by the Bar Chart Club Conference. Titled "Adding AI to Your Data Visualization Workflow"—though colloquially reframed as "Practical Tips for Thoughtful Engagement with AI and How to Resist the Creation of Meaningless Slop"—the session convened analysts, business owners, and data consultants.

Participants reported widespread industry fatigue regarding the ubiquitous tech-sector mantra to "get on board or get left behind." Rather than inspiring innovation, this fear-based push often devalues deep human expertise and conceals systemic workplace issues, such as understaffing and chronic underfunding.

Drawing upon feminist science studies—specifically referencing Donna Haraway’s seminal 1988 essay on Situated Knowledges—workshop attendees critiqued the illusion of disembodied, objective "facts" generated by algorithmic tools. The consensus from the collective ideation session was clear: while AI can streamline technical execution under strict supervision, it systematically erodes the nuanced, context-dependent thinking required for exceptional data visualization.

Implications: Reclaiming Digital Sovereignty in Data Work

As organizations grapple with the long-term realities of generative AI, professionals must establish clear boundaries to protect the integrity of their craft. Mitigating algorithmic harm requires shifting the discourse away from passive acceptance toward active, critical evaluation across three foundational pillars:

1. Accuracy: Is the generated output verifiably true, or is the model masking an underlying hallucination with confident syntax?
2. Security: Are sensitive datasets protected from unauthorized training ingestion and privacy violations?
3. Sovereignty: Are practitioners exercising personal judgment and self-determination, or are they allowing corporate-backed tech products to dictate their workflow values?

A critical component of this sovereignty involves redefining the role of prototyping. Referencing research by tech writer Frank Elavsky on the "prototyping bottleneck," experts emphasize that the slow, difficult parts of ideation are precisely what make them valuable. Prototyping is where the intellectual rigor of a concept is tested. Injecting AI too early into the creative pipeline substitutes genuine critical thought with statistically probable averages, turning flawed ideas into polished, unexamined slop.

Ultimately, data visualization remains a deeply human-centered discipline. It demands rigorous attention across visual, numeric, textual, and coding dimensions to transform raw numbers into meaningful cultural and societal contexts. While AI tools may accelerate certain mechanical steps, they cannot replace the human intuition, skepticism, and creativity that define exceptional data analysis. Protecting that human core is not merely a matter of workflow efficiency—it is an affirmation of the craft itself.