L’intervention de l’IA dans la production « news » et de textes en général pose des questions semblables à celles des automatismes récents dans la photographie et la priouction vidéo : où en sommes nous par rapport à l’authenticité et à la vérité des contenus censés être les notres ? C’est essentiel pour le droit d’auteur européen (différent du copyright anglo-saxon ou chinois) et pour notre public. Est-ce qu’on peut encore nous faire confiance lors ce que nous employons des aides IA ?
26.2.2026 by Jon Accarrino - How to build a “Zero Trust” architecture that prevents AI errors from destroying your newsroom’s credibility.
As media executives, we often look for the “magic bullet” software that solves our efficiency problems. But when it comes to generative AI, the most critical feature isn’t speed, it’s accuracy.
ThisTVNewsCheck guide breaks down the mechanics of AI hallucinations and offers a “Zero Trust” framework for integrating these tools without sacrificing your newsroom’s credibility.
Everything Gen AI Creates Is A Hallucination
When people ask me how to stop generative AI from hallucinating, I often think back to a lecture from Tommi Jaakkola at MIT. He explained something that fundamentally shifted my perspective: Everything a model like ChatGPT outputs is a hallucination.
Generative AI technology isn’t actually the “intelligent” tool we often think it is. It’s just an advanced prediction engine. Every word gen AI creates is a guess based on training patterns, not necessarily verified fact. We just happen to accept the “good” hallucinations, the ones that align with reality, and panic over the “bad” ones.
For newsrooms, where credibility is the only currency that matters, this distinction is terrifying. A single fabricated quote or invented court case can turn decades of earned trust into a dumpster fire.
But we cannot hide from artificial intelligence technology. New AI features are already embedded in the tools your teams use daily, from email clients to CMS platforms and web browsers.
“The simple availability of attention-getting content does not guarantee that people will trust that content over time,” says Brian Southwell, distinguished fellow at RTI International. “Trust often involves the belief that both parties share values or interests and are accountable for their actions. If you find out that you’ve received false information, can’t easily trace where it came from, and can’t turn to a human author to get an explanation of why it was wrong, you probably won’t feel comfortable going back to that source over time.”
The goal isn’t to eliminate hallucinations. That is currently impossible with foundational models. The goal is to build a “Zero Trust” architecture that catches them before they reach your audience.
What Causes AI To Generate Bad Hallucinations?
What Causes AI To Generate Bad Hallucinations?
Bad AI Hallucinations (Image via ORDO AI)
AI hallucinations happen when models prioritize fluency over accuracy. Because they predict the next likely word rather than retrieving verified facts, they can confidently present fabrications as truth.
The Mechanics:
Data Gaps: Lacking specific info, the model fills the void with plausible-sounding filler.
Overgeneralization: Applying a correct pattern (like legal citation formats) to invent non-existent cases.
Sycophancy: Tuning models to be “helpful” often trains them to invent answers rather than admit ignorance.
Real-World Consequences:
The “Food Bank” Tourist Trap: Microsoft’s AI guide recommended the Ottawa Food Bank as a “tourist hotspot” for hungry travelers.
The Invented Legal Precedent: A lawyer faced sanctions after ChatGPT wrote a brief citing entirely made-up court cases.
The “Pizza Glue” Warning: Google’s AI suggested putting glue on pizza to keep cheese from sliding off, a “correct” prediction based on a sarcastic Reddit comment it ingested. This is the danger of unrestricted web access: You invite the chaos of the internet into your news product.
Google’s “Pizza Glue” AI Overview Fail
Google’s “Pizza Glue” AI Overview Fail (Image via ORDO AI)
What Is The Atomic Unit Of Journalism?
Before we discuss defenses against bad AI hallucinations, we must establish the ground rules of AI use in trusted newsrooms. AI should never write the core news story. As TVNewsCheck Editor, Michael Depp explored during a panel at its NewsTechForum conference last December, the atomic unit of journalism, the original reporting, the interviews, the synthesis of facts, must remain human.
AI’s limited role is performing productivity tasks in the orbit of that atomic core: reversioning content for social media, summarizing archives for research, drafting newsletter teases, generating broadcast lower-thirds, etc.
As Nick Toso, founder & CEO of the journalist discovery platform Rolli, notes: “AI can help increase productivity, but it can’t take responsibility for what gets published. In a newsroom, every claim has to lead back to a real person who knows what they’re talking about and is willing to stand behind it.”
But do not be fooled: Even in these support roles, the risk of hallucination is real. An AI tool summarizing a city council transcript can still invent a quote. A tool generating a lower-third can still misattribute a statement. And as our newsrooms begin to adopt agentic tools that “talk” to each other (e.g. a research bot passing data to a social bot) we risk a digital game of “Telephone” where subtle nuances are lost and hallucinations are amplified.
AI Hallucinations: Strategic Defenses For Media Leaders
AI Hallucinations: Strategic Defenses For Media Leaders
AI Hallucinations: Strategic Defenses For Media Leaders (Image generated using ORDO AI)
We cannot simply trust AI to “do the right thing.” We must engineer environments where it is difficult for AI to do the wrong thing.
“AI is a powerful accelerant for our newsrooms, but we don’t allow that velocity to go unchecked,” says Claire Ferguson, VP & senior technology counsel at Gray Media. “Because Gray’s longstanding position has been our journalism is to be created by humans, every single output flows through our journalists’ rigorous verification systems and their subjective discernment before publication. That’s our firewall against putting unverified AI output in front of our audiences.“
Here are some of the proven safeguards that Ordo Digital recommends for news organizations.
1. Implement Retrieval-Augmented Generation (RAG)
The Concept: Stop letting the AI guess. Instead of relying on a model’s internal training data (a black box of the internet), RAG connects the AI to a trusted, closed ecosystem, your archives, wire services or specific primary documents.
Why It Works: It grounds the output in verified reality. You instruct the AI: “Answer this question using only the provided text.” This forces the model to reference your proprietary knowledge base rather than hallucinating from its training data.
2. The “Human-in-the-Loop” Protocol
The Concept: AI is the intern; you are the editor. Treat every AI output with skepticism. It can draft summaries or structure data, but it should never be the final publisher.
The Workflow:
Atomic Verification: Break down AI-generated text into individual claims.
Source Tracing: Require human editors to trace every quote and stat back to a primary source.
Final Sign-Off: No content goes live without a human editor’s name attached to the log.
“The biggest barrier to adoption isn’t technical, it’s cultural,” admits Kurt Christopher, VP of television digital strategy at Hubbard Broadcasting. “We see real hesitation from journalists who are protective of their work. And rightly so. They need to trust that it won’t distort the nuance of their reporting when it reformats a story for a different platform.”
3. Deploy Automated “Red Teaming” Fact Check Tools
The Concept: Fight AI with AI. Human editors are expensive; AI-generated content is cheap. Before any AI-drafted content reaches a human editor, it should pass through a secondary AI specifically designed to be a critic.
The Workflow: Use a “Self-Reflection” fact check loop or a separate “Judge” model. The first AI writes the content. The second AI (the Judge) scans that content solely to find unsupported claims, logical fallacies, or deviations from your style guide. If the Judge flags an issue, the content is rejected before it ever wastes a human editor’s time.
For example, my team frequently adds a ‘fact checker’ protocol extension to AI tools used by newsrooms.
Example Fact Check Protocol
Example of an Ordo Digital fact checking protocol prompt (Via ELVEX)
4. Limit the Scope of Use Cases
The Concept: Don’t use a bazooka to kill a mosquito. Establish clear “Safe” vs. “Unsafe” zones for AI.
Safe Uses: Summarizing transcripts, SEO metadata, translation, reformatting text.
Unsafe Uses: Writing original reporting from scratch, generating quotes or pulling live data on breaking news where facts are fluid.
Why It Works: Restricting AI to low-variance tasks naturally reduces the opportunity for high-risk hallucinations.
5. Lower the “Temperature”
The Concept: Boring is better. When developers use AI models through an API or other advanced platforms, they can tweak something called “temperature” which ranges from 0.0 to 1.0. This setting helps control how creative the AI can get.
Action Item: Set the temperature low to 0.0–0.2 for any sensitive newsroom task. This forces the model to be deterministic and factual, whereas higher temperatures (0.7+) encourage creative fabrication. This can be controlled either at the tool or enterprise level or just in a prompt. Or try adding “temperature = low” to your standalone prompts and experiment with the output differences.
6. Implement “Chain-of-Thought” Prompting
The Concept: Show your work. LLMs are prone to rushing to an answer. “Chain-of-Thought” forces the model to break down its reasoning step-by-step before generating a final response.
The Strategy: Instead of asking “Summarize this,” ask: “First, list the key entities. Second, identify the main action. Third, write a summary based only on those two steps.” This slows the model down, drastically reducing logical leaps.
7. Strict Prompt Engineering & Role-Based Constraints
The Concept: Garbage in, garbage out. Vague prompts invite improvisation. Strict prompts demand precision.
The Strategy:
Ice Breakers: Hard-code system prompts that set the rules before the user even types.
Roleplay: “You are a cynical, fact-obsessed news editor. You do not speculate.”
Demand Attribution: “Cite the specific paragraph used for every claim.”
8. Enable Uncertainty Scoring
The Concept: A “Confidence Meter” for your AI. Technically sophisticated teams can implement “Uncertainty Scoring” (or log-probability checks).
How It Works: The model assigns a probability score to its own tokens. If the confidence score for a specific claim drops below a certain threshold (e.g., 90%), the system automatically flags that sentence for human review or refuses to generate it entirely.
9. Domain-Specific Fine-Tuning
The Concept: Specialized training, not general guessing. Generic models are trained on the entire internet, including conspiracy blogs and Reddit posts.
The Strategy: “Train” a smaller version of the model on thousands of examples of your CMS or your newsroom’s best work. This teaches the AI your specific voice, formatting rules and ethical boundaries, reducing the “drift” that often leads to hallucinations.
10. The “Sycophancy” Check (Team Education)
The Concept: AI is a “Yes Man.” AI models are reinforcement-learned to please the user. If a reporter asks a leading question (“Find the admission of guilt in this transcript”), the AI will try to find it even if it doesn’t exist.
The Strategy: Train staff to ask neutral, objective questions to avoid bullying the AI into a hallucination. Give the AI explicit permission to say “I don’t know” or “That information is not present.”
“We see it all the time—a reporter asks an AI to ‘find the controversy in this transcript,’ and the model will often manufacture one just to be helpful,” says Daniel Anstandig, CEO of Futuri. To help prevent hallucinations, Futuri’s TopicPulse was grounded using 250,000 sources. “But even with the right tools, you have to train your team to ask neutral questions and explicitly give the AI permission to say ‘there is nothing here.’ If you don’t, you are effectively bullying the model into a hallucination.”
Establishing A ‘Zero Trust’ AI Architecture in Your Newsroom
Establishing A ‘Zero Trust’ AI Architecture in Your Newsroom
‘Zero Trust’ AI Architecture (Image via ORDO AI)
We are moving toward a world of “Agentic AI,” where systems don’t just create content but take actions. In that world, a hallucination isn’t just a typo, it’s a liability. Media organizations must be vigilant, focusing not just on the tools they buy, but on the rigorous architectures they build around them.
AI technology will continue to hallucinate. It is your job to ensure those hallucinations never become headlines.
Disclosure: Gray Media and Hubbard Broadcasting are clients of the AI strategy and consulting firm, Ordo Digital.
