Trust Is the Product: What AI Needs to Get Right in the Newsroom

A recent Deloitte article highlights a growing challenge for organizations scaling generative AI: How to ensure AI trust as adoption increases, so are concerns around data quality, model accuracy and trust.

In most industries, that’s a serious operational issue. For organizations delivering news, financial intelligence, research and data-driven insights, trust isn’t a feature. It’s the product. Audiences rely on that trust to form opinions, interpret events and make decisions. It’s built over time through editorial rigor, domain expertise and consistency in how information is presented and understood.

Generative AI is now being introduced into that environment at speed. It can help teams process vast amounts of information, surface insights and expand coverage. But it also introduces a new challenge: how to scale analysis without compromising the editorial standards that make information trustworthy in the first place. At the center of this challenge is an organization’s most strategic asset: unstructured data, the articles, transcripts and archives that make most of an organization’s information and serve as the raw material for any AI system.

Just as importantly, the industry is moving past experimentation. What began as pilots and isolated use cases is now becoming embedded in core workflows. The challenge is no longer what AI can do, but how to make it reliable, repeatable and sustainable as part of everyday editorial operations.

The Limits of Fixing Outputs

Generative AI is no longer experimental in media and publishing. It is already embedded in how information is gathered, analyzed and delivered.

Industry surveys show that most organizations are using it in some capacity, with many journalists incorporating it into regular workflows. In practice, it shows up in a wide range of use cases:

• Analyzing large document sets during investigations
• Translating and indexing content from multiple sources
• Generating summaries and drafting content
• Making archival material accessible through conversational interfaces
• Transforming reporting into structured data for downstream analysis

Journalists are using AI to process thousands of field documents, translating, indexing and surfacing connections that would have been difficult to identify manually. Publishers are enabling users to query decades of archived reporting through AI-powered interfaces, shifting from search-based access to synthesized answers grounded in trusted content. Information providers are leveraging their archives for insight to support decision making. Some organizations are even structuring their journalism as machine-readable data, allowing it to feed directly into analytics tools and decision making systems.

These examples show how AI can extend the reach and efficiency of editorial teams. But they also make the limitations of current approaches more visible, especially as systems are applied to ever-growing volumes of unstructured data that vary in quality, format and context. Because generative AI depends on this data, inconsistencies in how it’s managed and interpreted directly affect the reliability of what it produces.

When generative AI is used to summarize or interpret content, small inaccuracies can alter meaning. Misattributed quotes, incorrect figures or subtle shifts in framing are not uncommon. In fast-moving environments, these errors can pass through initial checks and require correction after publication, undermining confidence in the output.

At the same time, reliance on prompts and retrieval pipelines introduces variability. Two similar queries can produce different answers. Context can be incomplete or inconsistently applied, especially when systems draw from large, unstructured datasets without clear grounding in editorial standards.

Human oversight helps mitigate these risks, but it comes at a cost. If every summary, insight, or generated draft requires verification, what was intended to accelerate workflows can instead introduce new layers of review and coordination. As AI is applied to critical tasks like analysis, synthesis and interpretation, the cost of getting it wrong increases. And, the more that organizations rely on post-generation fixes to maintain quality, the harder it becomes to scale these systems effectively or demonstrate consistent business value.

For digital information providers, this creates a clear constraint. AI cannot simply produce plausible outputs. It has to produce results that are accurate, contextualized and consistent with the standards that define the organization’s credibility. This requires more than refining outputs after the fact. It requires systems that are designed to support trust from the start.

Why AI Trust in the Newsroom Breaks Down

The challenges outlined by Deloitte—fragmented data, integration complexity and hallucinations—take on a different weight in media and publishing environments.

These organizations operate on interconnected, often unstructured information: transcripts, filings, research data, multimedia content and decades of archived material. The task is not just to retrieve information, but to interpret it accurately and consistently. This is where generative AI, on its own, begins to fall short.

Language is inherently ambiguous. The same term can have multiple meanings depending on context, and the same name may refer to different entities. This points to a core limitation in how large language models process language. Approaches based on “similarity,” including semantic search and even retrieval augmented generation (RAG) rely on vector proximity rather than true entity resolution. This means that entities that share a name or context can still be conflated.

The same challenge comes up when journalists need to connect and understand events that are referenced across multiple sources. Because generative models operate within fixed training horizons, they may lack awareness of more recent developments, requiring additional work to stay current and complete. Humans resolve these language challenges through experience and domain knowledge, and AI systems must be designed to do the same. When they are not:

• Entities are misinterpreted or conflated
• Context is incomplete or inconsistently applied
• Outputs cannot be easily traced or verified, due to the black-box nature of many generative models and the lack of clear provenance
• Insights drift from the standards that define the organization’s voice

In media and information services, these are not minor errors. They directly impact credibility.

More fundamentally, many of these issues stem from relying on generalized models that lack domain-specific context. Without systems that reflect how information is structured, interpreted and validated within a given organization, even advanced models struggle to produce consistent, trustworthy results.

Trust Must Be Designed Into the System

To establish AI trust in the newsroom, organizations need to move beyond standalone tools and think in terms of systems that support reliability from end to end.

That means aligning AI with the same principles that guide editorial work and embedding those capabilities directly into the workflows and platforms where content is created, analyzed, and delivered. In practice, this comes down to four foundational pillars.

The four pillars of AI integrity in the newsroom

1. Preserving institutional knowledge
Media organizations have something most AI systems lack: decades of accumulated expertise. Editorial archives, proprietary datasets, domain taxonomies and curated knowledge bases (all largely unstructured data) reflect how an organization understands and explains the world.

Generic AI models can’t replicate this. If systems rely only on external or generalized data, they lose the context, nuance and consistency that define trusted content. Grounding AI in proprietary knowledge ensures that outputs reflect the organization’s voice and standards, not just patterns learned from the web.

In practice, this means structuring and enriching archival content so it can be understood and used by AI systems. Approaches such as semantic enrichment and customizable knowledge graphs allow organizations to map entities, topics and relationships in a way that reflects their domain expertise. Instead of treating archives as static repositories, they become active knowledge layers that inform how AI interprets and generates content.

This is what enables AI that is based on a hybrid approach—with more than one AI technology at work—to outperform one-size-fits-all models by grounding outputs in the context, terminology and perspective that define each organization. This allows AI to build on existing expertise rather than bypass it.

2. Deterministic data disambiguation
Ambiguity is one of the most persistent challenges when working with language and information.

A single term can carry multiple meanings. A company name may refer to different entities across regions. Events often involve overlapping actors and similar ways to identify them. In editorial workflows, resolving these ambiguities is essential.

AI systems need to do this deterministically. By using structured knowledge models to define entities and relationships, systems can distinguish meaning based on context. Techniques such as named entity recognition, relationship extraction and domain-specific knowledge graphs help ensure that references are correctly interpreted across large and complex datasets.

This level of precision allows systems to connect information accurately across sources, reducing the risk of misinterpretation and improving the consistency of outputs. Without it, errors compound quickly and undermine trust.

3. Explainability and auditability
In media environments, it’s not enough for an answer to be plausible. It must be explainable.

Editors and analysts need to understand:

• Where information comes from
• How it was interpreted
• Why a conclusion was reached

Black-box systems make this difficult. When outputs cannot be traced back to their sources or logic, they are difficult to validate and even harder to trust.

Combining generative AI with structured knowledge and reasoning-based approaches makes it possible to provide that visibility. Outputs can be linked to underlying data, enriched with metadata and delivered with clear provenance that shows how insights were generated.

This traceability allows teams to audit results, verify sources and maintain confidence in the system, especially when working with sensitive or high-stakes information. Such transparency is essential for maintaining editorial standards.

4. Expert-in-the-loop governance
AI can accelerate analysis, but its greatest value comes from amplifying human expertise, not replacing it.

Editors and analysts bring judgment, context and accountability that no model can replicate. This is especially critical when dealing with sensitive topics, complex narratives or high-stakes information. The goal is to extend that expertise, allowing teams to work faster and at greater scale while maintaining control over quality. Expert-in-the-loop governance ensures that editors and analysts remain involved where it matters most:

• Validating insights
• Refining interpretations
• Guiding how systems evolve over time

At the same time, AI can help reduce the manual burden by prioritizing high-value insights, organizing knowledge and supporting content reuse across formats and channels. This improves operational efficiency while helping teams focus on the work that drives the most value. This balance of automation with control is what makes AI usable at scale in editorial environments.

From AI Tools to Editorial Systems

Trust cannot be layered on after content is generated. It must be built into the system from the start. This requires a shift in how AI is implemented, from isolated tools and experiments to integrated infrastructure embedded within editorial workflows.

Instead of treating generative AI as a standalone capability, it needs to be part of a broader architecture that combines different approaches: language generation, structured knowledge, reasoning and machine learning.

This is the working behind Expert.ai’s Composite AI approach (also known as Hybrid AI) . By combining neurosymbolic AI to structure knowledge with generative models and knowledge-driven and reasoning-based techniques, organizations can move from loosely guided outputs to more controlled, context-aware results—aligned with their data, their expertise and their editorial standards. The result: accuracy, data independence, scalability and explainability. (Learn more about how we help publishers and digital information providers.)

Scaling Insight without Diluting Trust

The pressure on media and information organizations is not going away. The volume of information will continue to grow and the demand for speed and insight will continue to increase. AI can play a critical role in meeting that demand.

But in environments where trust defines the product, speed alone is not enough. Establishing AI trust in the newsroom requires systems that can interpret information accurately, apply context consistently and produce outputs that can be trusted without constant correction.

Scaling generative AI is not just a technical exercise, it’s about aligning technology with the principles that make information valuable in the first place. Without a strong foundation for managing and leveraging unstructured data, even the strongest AI initiatives will struggle to move beyond experimentation into sustained, production ready use.

The goal is not simply to produce more content faster, but to scale insight without diluting trust. Strengthening AI trust in the newsroom is what ultimately determines whether these systems deliver lasting value.