Preventing and Limiting Hallucinations in LLMs: Confession as a New Safeguard

In recent years, large language models (LLMs)—whether used for document summarization, content generation, or automated analysis—have established themselves as powerful and useful tools. But one problem persists, and it is structural: these models “hallucinate”—that is, they generate made-up information, incorrect facts, or fictitious quotes.

In professional contexts—such as when drafting reports, analytical summaries, or even documents submitted to clients or government agencies—these errors can have serious consequences, not only technical but also legal and reputational.

Recently, the highly publicized case involving Deloitte—which was accused of submitting several AI-generated reports containing fictitious data—served as a reminder of just how real these risks are. In this context, a notable innovation has emerged at OpenAI: “confession”—a method for prompting the model to admit its errors or uncertainties. This approach, still in the experimental stage, could serve as a valuable safeguard for reducing the risks associated with LLM hallucinations.

What is an LLM hallucination, and why does it occur?

LLM hallucinations result from a fundamental misalignment between what the model is optimized to do and factual truth. LLMs are trained to predict coherent sequences of words, not to verify facts or sources. In the absence of reliable sources, or when the context is unclear, the model may invent facts or quotes.

Even with modern techniques such as “reinforcement learning from human feedback (RLHF)” or precise instructions, these models are still prone to inventing passages—especially when they try to satisfy conflicting objectives: being useful, concise, convincing, comprehensive, and so on. This tension can lead the model to take shortcuts: guessing or making assumptions rather than saying “I don’t know.”

Traditional Methods for Limiting LLM Hallucinations

Before addressing the “confession,” several approaches can be combined to minimize risks:

• Retrieval-Augmented Generation (RAG): By combining large language models (LLMs) with databases, corporate documents, archives, or the web, we ground content generation in verifiable information, thereby reducing unfounded speculation.
• Prompt engineering: Clear, structured instructions with explicit constraints (specify sources, highlight uncertainties, and return “I don’t know” when uncertain) can help the model remain rigorous.
• Hyperparameter tuning (temperature, top-p, etc.): Generating at a low temperature reduces random creativity and increases consistency.
• Human validation (human-in-the-loop): Any LLM output intended for a client, government agency, or public use must be reviewed, verified, cross-checked against reliable sources, and validated by a human.
• Cross-validation / Multi-model: Interact with multiple models or repeat the generation process, compare the results, to identify robust assertions.

However, these methods are not sufficient to completely eliminate the risk of error, which is why complementary approaches are useful.

Why Confession Can Really Make a Difference

Until now, there has been no method that could reliably determine, after generation, whether the model had taken a shortcut or invented an element. This is precisely what the so-called “confession” method seeks to address.

The principle is simple: after producing a response, the model generates a second piece of content designed to evaluate its own behavior. This supplementary report does not seek to correct the initial response, but rather to analyze whether it complies with the guidelines: accuracy, lack of fabrication, transparency regarding sources, etc. This serves to make explicit something that, at present, remains implicit: the degree of uncertainty inherent in AI.

This approach is based on a specific training method: the model is not rewarded for being correct, but for telling the truth about its behavior, even when it is wrong.

In practical terms, confession allows us to:

• Transparency in reasoning: The model explicitly indicates whether certain elements have been inferred, assumed, or extrapolated.
• Risk alerts: The model can recognize that it did not have enough reliable data.
• Auditability of results: A user can review the submission to verify whether the answer actually complies with the established rules.
• Automatic interruption or review in case of uncertainty: For example, the workflow can block a response that is not serious enough.

This mechanism does not make the answer correct; it makes the error detectable. Instead of a well-presented but potentially false text, we obtain content accompanied by a usable reliability indicator. In other words, the “confession” does not eliminate LLM hallucinations; rather, it provides a form of traceability for them and thus serves as a building block for governance and quality assurance.

What This Means for Businesses

At DATASOLUTION, we are convinced that transparency is a solution, but that on its own, it cannot prevent bias. Governance must therefore be put in place:

• Adopt a multi-layered approach: combine proven methods (RAG, prompt engineering, checks, human review) with emerging techniques such as confession, but without considering them sufficient on their own.
• Establish responsible AI pipelines: For all automated production, systematically include a verification, review, and audit phase conducted by domain experts.
• Conduct internal experiments: Test a model designed to elicit confessions in non-critical contexts to assess the reliability of the confessions.
• Train teams: Remind them that AI is not infallible, and that verification is a tool, not a guarantee. Encourage caution and systematic verification of all deliverables.
• Documenting “transparency”: In external deliverables, clearly explain the methodology—“AI-assisted content, verification performed, uncertainties identified, sources consulted, human review.” This helps build a credible and responsible narrative.

What are our recommendations

LLM hallucinations are a structural problem: given the current state of technology, we cannot expect models to be infallible. But that doesn’t mean we should give up on AI—on the contrary: it’s a reason to build robust, responsible processes that combine technology, governance, human review, and transparency.

The “confession” feature proposed by OpenAI represents an encouraging step forward—a way to highlight errors, introduce transparency, and enable audits, reviews, and human validation. However, it should not be viewed as a substitute for rigor, nor as an automatic guarantee of truth.

For businesses, the challenge is clear: to use AI to its fullest potential—to drive productivity, speed, and scalability—while rigorously managing risk and implementing operational safeguards.