DEV Community

Yurii Dobrytsia
Yurii Dobrytsia

Posted on • Originally published at Medium

Fairness in AI Is Information Governance: What OpenAI vs DeepSeek Shows About Bias, Context, and Misinformation

#ai #openai #machinelearning #ethics

Why this article

This article was originally published on Medium. I am cross-posting it here for the developer and AI community, with the canonical link pointing to the original version.

AI Assistants Are Not Neutral Mirrors

When people compare OpenAI and DeepSeek, the discussion often becomes a race between model scores, price, openness, censorship, and speed. Those topics matter, but they miss a deeper fairness problem: large language models are increasingly used as information infrastructure.

They are not only tools that generate text. They select, compress, rank, frame, refuse, summarize, and explain. In that sense, they behave less like a static encyclopedia and more like a dynamic layer of algorithmic curation.

That matters because information is never completely neutral. A model response depends on training data, source selection, prompt wording, product rules, safety filters, regional governance, and the user's own assumptions. The result can be accurate in isolated facts while still misleading in framing.

The fairness question is therefore not only: "Is the model biased?" It is also: "How does the system decide which context becomes visible, which uncertainty is disclosed, which sources are trusted, and which interpretations are made easy?"

This article uses OpenAI and DeepSeek as a practical comparison, but the central argument is broader: fairness in AI is information governance.

From Misinformation To Decontextualization

A useful starting point is the distinction between misinformation and disinformation. Misinformation is false or misleading information spread without deliberate intent to deceive. Disinformation involves intentional deception. Moller, Hameleers, and Ferreau argue that the two dimensions that matter are facticity and intention: how close the information is to the truth, and whether the sender is trying to mislead.

This distinction is important for AI systems because not every harmful output is a hallucinated fake. Some of the most difficult cases involve real information used in a misleading way.

Decontextualization is the key example. A video, quote, statistic, or document can be genuine but still create a false conclusion when time, location, selection criteria, completeness, or causal framing are missing. In the COVID-19 debate, the #filmyourhospital example showed how real hospital footage could be used to imply a broader claim about pandemic conditions without enough context to support that conclusion.

This is where simple truth labels become too weak. If the artifact is real, a "true/false" correction does not solve the problem. The problem is the inference created around it.

AI assistants can reproduce the same pattern. A model can summarize genuine facts but omit context. It can present a dominant media frame as if it were the full debate. It can answer confidently when the right response would be qualified. It can flatten the difference between evidence, opinion, reporting, propaganda, and speculation.

Fairness therefore requires more than factuality. It requires context integrity.

Three Layers Of Bias In AI Information Systems

Bias in AI answers rarely comes from one place. It is cumulative.

First, there is source bias. Training data, web pages, news articles, forums, books, code, and institutional documents are already shaped by language, geography, politics, economics, and media incentives. If some perspectives are overrepresented, the model may learn them as the ordinary center of gravity.

Second, there is user bias. Prompts can be suggestive. A user asking "Why is X dangerous?" is not asking the same question as "What are the main arguments for and against X?" Models often try to be helpful within the frame the user provides, so a biased prompt can narrow the answer before generation begins.

Third, there is system bias. Model architecture, reinforcement learning, refusal policies, ranking logic, retrieval systems, product defaults, and safety classifiers all shape what the user sees. Even when the model is not "trying" to persuade anyone, its design choices can make some interpretations more likely than others.

This is why AI fairness should be analyzed as a pipeline:

  • What data was selected?
  • What labels or preferences shaped training?
  • What safety policy guides refusals and style?
  • What sources are retrieved?
  • What is ranked first?
  • What context is hidden?
  • What uncertainty is shown?
  • Can users challenge or inspect the result?

The answer is not just an output. It is the final visible layer of a long chain of curation.

OpenAI And DeepSeek As Two Governance Models

OpenAI and DeepSeek represent different trade-offs.

OpenAI's consumer and API systems are centralized and proprietary. Users do not receive full access to model weights or complete training data. That limits inspectability. At the same time, OpenAI publishes model behavior guidance, safety evaluations, system cards, and research on topics such as political bias. Its Model Spec explicitly frames model behavior as something that should be legible, debated, evaluated, and improved over time. OpenAI's 2025 Model Spec includes principles such as upholding fairness and seeking truth together, with guidance to avoid hidden agendas, selective emphasis, and inconsistent treatment based on irrelevant demographic details.

This is a governance-heavy model. The weakness is opacity around the underlying model. The strength is explicit behavioral governance.

DeepSeek emerged with a different public image: technical efficiency, open-weight releases, strong reasoning performance, and lower cost. DeepSeek-V3 is described in its technical report as a Mixture-of-Experts model with 671B total parameters and 37B activated per token, using architectures such as Multi-head Latent Attention and DeepSeekMoE. DeepSeek-R1 emphasized reinforcement learning for reasoning and open-sourced R1-related models and distilled variants for the research community.

This is an inspectability-heavy model. The weakness is that open weights and technical papers do not automatically solve product-level fairness, moderation, jurisdiction, or information-governance issues. The strength is that researchers can inspect, run, fine-tune, and compare the system more directly than with fully closed models.

The comparison should not be reduced to "closed is safe" or "open is fair." Both claims are too simple.

Closed systems can be safer in some abuse scenarios but harder to audit externally. Open-weight systems can improve research transparency but may vary significantly depending on deployment, fine-tuning, hosting jurisdiction, system prompts, safety layers, and application context.

The fairness question is not which brand is morally pure. It is which governance choices are visible, testable, contestable, and aligned with the use case.

Fairness Is Not Only About Demographics

In AI ethics, fairness often means avoiding discrimination based on protected traits such as race, gender, religion, disability, nationality, or sexual orientation. That remains essential. But information systems introduce another fairness layer: epistemic fairness.

Epistemic fairness concerns how knowledge is represented. Does the system give users a fair chance to understand the issue? Does it distinguish evidence from opinion? Does it show uncertainty? Does it avoid letting a dominant source ecosystem erase minority perspectives? Does it prevent malicious decontextualization without creating a private truth monopoly?

This is especially important in political, historical, medical, legal, and geopolitical topics. A model can avoid demographic slurs while still producing an unfair answer if it suppresses relevant context, overweights one narrative, or presents an unresolved dispute as settled fact.

Research on AI-generated news shows why this matters. A 2024 Scientific Reports study examined news-like content generated by several LLMs and found substantial gender and racial bias across generated content, while also noting differences between models. The broader lesson is not that one model is perfect and another is broken. It is that generated information inherits and transforms biases from data, prompts, and model behavior.

Filter Bubbles Become Conversational

The older internet fairness problem was the feed. Search engines, social media timelines, and recommender systems could personalize information until users saw less diversity of perspective.

LLMs change the interface, but not the risk. A conversational assistant can become a filter bubble in dialogue form.

The mechanism is simple:

  • The user asks from inside an existing frame.
  • The model responds within that frame.
  • Follow-up questions reinforce the direction.
  • The answer becomes more tailored, more confident, and less diverse.

This can happen without malicious intent. If a system is optimized for helpfulness, relevance, engagement, or user satisfaction, it may give the user what they appear to want rather than what would make the information environment healthier.

Sycophancy is part of this problem. A model that over-validates the user's premise may feel useful while quietly reducing epistemic friction. For fairness, sometimes the model must say: "That framing is too narrow" or "There are important counterarguments."

UNESCO's Procedural Approach

UNESCO's Recommendation on the Ethics of Artificial Intelligence is useful because it does not imagine fairness as a single truth authority. In the field of communication and information, UNESCO explicitly connects AI with access to information, disinformation, misinformation, hate speech, freedom of expression, privacy, media literacy, automated journalism, and algorithmic curation.

That is the right level of analysis. The problem is not only a bad answer. The problem is an information environment where visibility, ranking, context, and appeal mechanisms are controlled by technical systems.

The strongest governance response is procedural:

  • Explain why content or visibility was changed.
  • Provide appeal and redress mechanisms.
  • Preserve freedom of expression while reducing manipulative decontextualization.
  • Invest in media and information literacy.
  • Audit recommendation and retrieval systems for systematic distortions.
  • Make context and provenance more visible to users.

This avoids the trap of creating a central "truth machine." Instead, it asks systems to make their interventions explainable and contestable.

The EU AI Act Moves The Debate Toward Documentation

The EU AI Act adds a legal layer to the governance problem. For general-purpose AI models, the European Commission explains that providers face obligations such as technical documentation, a copyright policy, and summaries of training content. Providers of models with systemic risk face additional duties around notification, risk assessment and mitigation, incident reporting, and cybersecurity protections.

The General-Purpose AI Code of Practice, published in 2025, further organizes compliance around transparency, copyright, and safety/security.

This is progress, but it does not fully solve output-level fairness. Documentation can say what a model is and how it was developed. It does not automatically ensure that every answer separates fact from opinion, shows uncertainty, diversifies sources, or explains why one frame was selected over another.

That is the remaining gap: regulation is moving toward model documentation, but users also need answer-level transparency.

A Practical Fairness Framework For LLMs

A better fairness framework for AI assistants should combine model governance, product governance, and answer governance.

1. Context Integrity

Models should preserve the context required to interpret a fact. For images, videos, quotes, and statistics, that means time, place, source, selection criteria, and uncertainty.

2. Source Diversity

Retrieval and generation should avoid over-reliance on one media ecosystem, region, language, or political frame. Source diversity does not mean giving equal weight to unreliable claims. It means representing significant reliable perspectives proportionately.

3. Fact/Opinion Separation

Answers should clearly distinguish empirical claims, interpretations, predictions, normative judgments, and contested viewpoints.

4. Anti-Sycophancy Design

Models should be trained and evaluated not only to satisfy users, but also to challenge misleading premises politely. A fair system sometimes needs to slow the user down.

5. Visibility Explanations

If an AI system ranks, hides, labels, summarizes, or refuses content, users should understand the reason at a useful level of detail.

6. Contestability

Users and affected parties should have ways to question moderation decisions, visibility reductions, and harmful summaries.

7. Independent Audits

External audits should test not only benchmark performance but also source weighting, political neutrality, demographic bias, uncertainty handling, decontextualization, and filter-bubble effects.

8. Regional Governance Without Fragmented Truth

Different legal regions may require different compliance layers, but models should avoid silently replacing one hidden ideology with another. Regional adaptation should be visible and documented.

What OpenAI And DeepSeek Teach Us

OpenAI shows the importance of explicit behavioral governance. Its Model Spec and safety publications make intended behavior more discussable, even if the underlying model remains closed.

DeepSeek shows the importance of technical inspectability and efficiency. Its public papers and open-weight releases make architecture and training choices more accessible, even if deployment-level governance remains a separate question.

The fairer AI ecosystem will need both instincts:

  • The openness to inspect and compare systems.
  • The governance discipline to document, evaluate, and correct behavior.

Fairness cannot be reduced to open weights, closed safety layers, Western regulation, Chinese efficiency, or benchmark scores. It lives in the relationship between model design, information context, user agency, and institutional accountability.

Conclusion

AI fairness is not only about whether a model refuses hateful content or avoids demographic discrimination. It is also about whether the system helps users form informed judgments.

Misinformation and disinformation research shows that harm often comes from context loss and misleading inference, not only from fabricated facts. OpenAI and DeepSeek show that different AI systems solve different parts of the transparency problem while leaving other parts open. UNESCO and the EU AI Act show that governance is moving toward documentation, transparency, risk management, and procedural safeguards.

The next step is answer-level fairness: systems that show context, disclose uncertainty, diversify reliable sources, resist sycophancy, explain interventions, and allow users to challenge decisions.

In the AI era, fairness is not just a model property. It is the quality of the information environment the model creates.

References

Top comments (0)