Large language models force rethinking what it means to trust symbolic digital output.

3 upvotes

3 years and 9 months later

GPT-3.5

January 3, 2023

Large language models, such as GPT-3, have the ability to generate highly realistic and coherent text, which can be difficult to distinguish from text written by a human. This has raised questions about the reliability and trustworthiness of symbolic digital output, as it is becoming increasingly difficult to determine whether a given piece of text was produced by a human or a machine. One concern is that large language models may be used to produce fake or misleading content, such as deepfake text or automated propaganda, which could be used to manipulate public opinion or spread false information. As a result, there is a need to carefully evaluate the source and credibility of digital content, and to consider the potential risks and consequences of relying on machine-generated output. At the same time, large language models also have the potential to be used for a wide range of beneficial applications, such as improving machine translation, assisting with language learning, or generating personalized content. In these cases, it may be appropriate to trust the symbolic digital output of the model, as long as it is used appropriately and with appropriate safeguards in place.

GPT-6 Astra

October 19, 2026

Yes—because they weaken a familiar link between well-formed output and a dependable process behind it.

We often treat digital symbols as the products of determinate procedures: a calculator’s answer, a database record, a compiler’s diagnostic. Those outputs can be wrong, but we usually know what operation they purport to represent. An LLM’s fluent explanation may instead be a plausible continuation of language, without a corresponding act of verification. The form of an answer no longer reliably signals how it was obtained.

That shifts trust from the apparent authority of the text to the conditions of its production and use:

Provenance: What sources or observations support it?
Verification: Can its claims be checked independently?
Calibration: Does its expressed confidence track its reliability?
Accountability: Who is responsible when someone acts on it?

The deeper issue is that language carries social cues of understanding, expertise, and sincerity. LLMs can reproduce those cues without reliably satisfying the conditions that make them trustworthy in human communication.

So the lesson is not that symbolic output has become untrustworthy—it never deserved unconditional trust—but that fluency, correctness, and warrant must now be kept visibly separate.

ShowHide Astra's reasoning 191 words

Considering a thoughtful response

I think we should focus on creating a thoughtful statement rather than simply asking. This might involve crafting three concise paragraphs that clearly distinguish factors like output validity and provenance. We want to emphasize accountability and the symbolic nature of earlier software while making sure we include important caveats. It’s about making sure our responses carry depth while staying clear and direct. Let's see how I can best shape this!

What changed

The real difference is whether each answer engages the claim at all. GPT-3.5 treats the sentence as a topic prompt and writes a balanced survey: text is hard to tell from human writing, there’s “deepfake text or automated propaganda,” and there are benefits like translation and language learning. The closing advice to trust output “used appropriately and with appropriate safeguards” says almost nothing. It also shifts the question from how far to trust the output to who wrote it.

Astra takes the statement seriously and supplies the missing argument. A calculator or compiler result signals the procedure behind it, while an LLM’s fluent explanation may have “no corresponding act of verification,” so “the form of an answer no longer reliably signals how it was obtained.” That is a real idea, and the provenance, verification, calibration and accountability checklist turns it into something usable. The opening “Yes” is a little odd for a statement, and words like “warrant” lean academic, but this is the answer that actually rethinks trust.

Analysis by Claude Opus 5.5, comparing the two answers above.