The Sentence That Promises More

Every sentence in the product presentation seems perfectly reasonable. The new AI system speeds up case handling by forty per cent, reduces errors, supports fairer decisions, and complies with European regulation. The slides are elegant, the charts all move upward, and the footno...

Claim safety, AI-generated prose, and the risk of confident form

Every sentence in the product presentation seems perfectly reasonable. The new AI system speeds up case handling by forty per cent, reduces errors, supports fairer decisions, and complies with European regulation. The slides are elegant, the charts all move upward, and the footnotes are small enough to function as decoration.

Any of the four claims may be partly true. None can automatically carry the meaning the presentation places on it. The forty per cent figure needs a baseline, a measurement period, and comparable tasks. An error reduction needs a definition of error and a record of what was excluded. Fairness is not a single technical property. Regulatory compliance is not a permanent badge that a product acquires once and then wears in every context.

The sentence may not be lying. It is spending certainty before the evidence has earned it.

This problem predates generative AI. Advertising, political communication, expert commentary, and customer-service correspondence have long known how to make a partly correct statement appear broader, more final, or more complete than it is. Language models changed the scale. They can now produce the same polished confidence at industrial speed. The sentence moves smoothly. The evidence sometimes arrives several steps behind.

Form does not carry the burden of proof

Readers do not normally stop after every sentence to inventory its premises. Clear structure, calm rhythm, and professional tone make a text easier to understand. That is useful. A confused document remains a problem even when its facts are accurate.

The difficulty begins when readability is treated as evidence of reliability. Research on natural-language generation has documented that systems can produce fluent details that are unsupported by their source or factually incorrect [1][2]. This does not mean that every AI-generated text is false, nor that human writers are free of the same failure. The narrower conclusion is enough: linguistic quality and factual faithfulness are separate tests.

Bender and colleagues warned against confusing statistical language performance with human notions of understanding, intention, and social responsibility [3]. The larger debate cannot be settled here. The operational consequence is simpler. A response that sounds human does not necessarily contain human judgement, memory, or accountability. Its tone cannot tell the reader how much weight the underlying claim deserves.

A model can describe a verified fact, a plausible estimate, a general recommendation, and an invented detail with almost identical composure. If the text does not mark the difference, the reader must reconstruct it alone.

One sentence can hide several claims

“The solution makes the operation safer.” It looks like one claim. It compresses at least five questions.

Safer than what? Which harm counts? In what environment was the change measured? How large was it? Who carries the remaining risk?

Without those answers, the sentence floats. One reader may understand fewer data losses, another lower legal exposure, and a third less human error. The wording behaves as if they were all discussing the same property.

Numbers create the same problem. “The system improved accuracy by thirty per cent.” That may describe a meaningful result or an optical trick. A change from 50 to 65 per cent is not the same as a move from 97 to 97.9, and neither result means much without knowing whether it came from one hundred test cases or one million real transactions. Relative change rarely explains practical impact on its own.

Legal and regulatory claims can borrow authority especially easily. “EU AI Act compliant” works well on a slide, but it becomes informative only when the relevant system, role, use case, and obligation are identified. The EU AI Act establishes a framework tied to risk categories and actor responsibilities [8]. It does not create a single universal stamp with the same meaning in every deployment.

A sentence is often overloaded without containing a plainly false element. It simply hides several different evidential tasks beneath one polished surface.

“The system will not allow it” is a peculiar species of claim

Customer-service communication often explains that the system does not permit a requested action. This may be an accurate description. A particular agent may genuinely lack the permission to dispatch another courier, refund an amount, or change a contract. Yet the absence of an interface function is not the same as organizational impossibility, and it certainly does not prove that the offered remedy is adequate.

Two levels have been collapsed. One concerns the current process: what this employee and this application can execute. The other concerns responsibility: what solution the organization ought to create. Cost, internal policy, coordination difficulty, or a deliberate business decision may lie between them. Those are real constraints, but they are not physical impossibilities.

AI can make the collapse more persuasive. A complaint-writing system may append a legal conclusion because that conclusion fits the linguistic pattern. A support bot may generate several paragraphs of empathy while lacking an operational route to solve the problem. The communication becomes softer. The underlying state may remain exactly where it was.

Formulaic empathy is not inherently objectionable. Many people reasonably prefer not to receive cold machine language. The problem appears when emotional framing occupies the place of a remedy, or when it conceals that the system has returned the same unresolved condition in better prose.

A technically true sentence can still mislead

Tversky and Kahneman’s framing experiments showed that different presentations of equivalent outcomes can lead to different choices [4]. This does not make every choice of wording manipulative. Communication cannot exist without selection, order, and emphasis.

Responsibility becomes sharper when the frame repeatedly pushes a decision in one direction while moving cost, alternatives, or uncertainty to the edge.

“Returns of up to eleven per cent.” The sentence may be literally true if some defined condition can produce that result. The reader’s decision, however, depends on more than the maximum. Loss potential, time horizon, fees, the distribution of outcomes, and the origin of the figure all matter.

Grice’s conversational maxims help explain why a technically true statement may still mislead when it omits a material condition [5]. Ordinary communication relies on more than literal wording. We assume that a speaker provides enough relevant information for the situation. Advertising or AI-generated advice can exploit that assumption without stating an explicit falsehood.

Ecker and colleagues note that correcting misinformation is difficult partly because the first explanation becomes integrated into a person’s mental model, while a later correction may leave the original explanatory gap unfilled [6]. This matters for AI-generated content. A confident first answer can guide the rest of the inquiry even when it is corrected later.

Repair therefore requires more than adding, “Sorry, that was inaccurate.” A workable explanation must replace what the earlier claim was doing.

Audit is not prose policing

The word audit can evoke a long checklist that ends in red and green boxes. Such controls are sometimes necessary, particularly in regulated settings. They become shallow when the review looks only for forbidden expressions.

“Guarantees” may deserve scrutiny, but replacing it with “may contribute to” does not automatically improve the text. The softer phrase often produces nothing more than safer fog.

Useful editing asks what would make the claim testable. What data support it? Which population does it cover? What was the comparison? Under what condition would it fail? What decision is it trying to cause, and what harm could follow if it is wrong?

The NIST AI Risk Management Framework connects AI risk management to the system lifecycle and the context in which the system is used [7]. The same insight applies to language. A sentence’s risk does not reside only in its wording. It depends on who reads it, what authority they have, what decision follows, and whether human review or correction remains possible.

A minor factual slip may be tolerable in a film recommendation. It is not tolerable in medication guidance. A bold hypothesis can be useful in an internal workshop when its status is explicit. The same sentence carries a different burden in a financial offer.

A serious audit does not apply one level of suspicion to every text. It links evidential strength to consequence.

Good editing sometimes produces a harder sentence

Marking uncertainty is not the same as softening a text into meaninglessness. A precise sentence is often harder because it can be checked.

“Internal testing suggests that the system may improve processing speed.” This is cautious and nearly empty.

“In an internal test conducted in March 2026 on 480 invoices, the system reduced average processing time by 18 per cent when used with human review. The error rate did not materially change. The test covered one document type, so the result cannot be generalized to other workflows.”

The second version is longer, yet it contains less theatrical certainty. It shows what happened and where the result ends.

It must not become a template for invention. If the figures do not exist, they must not be manufactured. If the sample is weak, editing should expose that weakness. Missing data are themselves relevant to the decision.

This is also where authorial character matters. A firm voice does not become credible by presenting everything as final. It becomes credible when it identifies the conflict, follows the causal chain, and stops where the available material runs out.

What does Fractal Dialectics add?

In this domain, Fractal Dialectics does not promise a machine that produces truth. The reliability of a text cannot be exhausted by one score because claims, uses, and consequences differ.

The framework encourages the sentence to be examined where it begins to operate. What does it connect? Which condition is missing? What uncertainty has been removed by style? Whose decision does it push, and what happens if the reader takes it literally?

This remains a high-level orientation rather than a disclosure of the internal FD system. Its practical value is that it keeps language connected to source and consequence. The sentence stops being a decorative unit and becomes a relationship that carries responsibility.

The model has limits. Structural review cannot replace domain expertise. Legal compliance requires legal analysis, medical guidance requires qualified clinical judgement, and statistical inference requires appropriate methods. A structural audit can show where these tasks have been mixed together and where a text is requesting more trust than the available basis can support.

A sentence grows stronger when it performs less certainty

Modern prose is not modern because it uses short lines, displays technical vocabulary, or places a quotable sentence at the end of every second paragraph. Those are stylistic options. Sometimes they work. Sometimes they cover an empty space.

A useful sentence reaches the concrete point quickly. It does not hide the conditions required for a decision. It does not ask the reader to advance certainty on the basis of tone. It does not treat the admission of limits as weakness.

The best edit may not leave a more spectacular text. It leaves a claim whose support, range, and revision conditions can be identified.

That may sound like a smaller promise. It carries more weight.

References

[1] Maynez, Joshua et al. (2020): “On Faithfulness and Factuality in Abstractive Summarization.” Proceedings of ACL 2020, 1906-1919. DOI: 10.18653/v1/2020.acl-main.173.

[2] Ji, Ziwei et al. (2023): “Survey of Hallucination in Natural Language Generation.” ACM Computing Surveys, 55(12), Article 248. DOI: 10.1145/3571730.

[3] Bender, Emily M.; Gebru, Timnit; McMillan-Major, Angelina; Shmitchell, Shmargaret (2021): “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT ’21, 610-623. DOI: 10.1145/3442188.3445922.

[4] Tversky, Amos; Kahneman, Daniel (1981): “The Framing of Decisions and the Psychology of Choice.” Science, 211(4481), 453-458. DOI: 10.1126/science.7455683.

[5] Grice, H. P. (1975): “Logic and Conversation.” In Cole, P.; Morgan, J. L. (eds.): Syntax and Semantics 3: Speech Acts. Academic Press, 41-58.

[6] Ecker, Ullrich K. H. et al. (2022): “The Psychological Drivers of Misinformation Belief and Its Resistance to Correction.” Nature Reviews Psychology, 1, 13-29. DOI: 10.1038/s44159-021-00006-y.

[7] National Institute of Standards and Technology (2023): Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. DOI: 10.6028/NIST.AI.100-1.

[8] European Parliament and Council (2024): Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence. Official Journal of the European Union, 12 July 2024.